html_url
stringlengths 48
51
| title
stringlengths 1
290
| comments
sequencelengths 0
30
| body
stringlengths 0
228k
⌀ | number
int64 2
7.08k
|
|---|---|---|---|---|
https://github.com/huggingface/datasets/issues/6014 | Request to Share/Update Dataset Viewer Code | [
"Hi ! The huggingface/dataset-viewer code was not maintained anymore because we switched to a new dataset viewer that is deployed available for each dataset the Hugging Face website.\r\n\r\nWhat are you using this old repository for ?",
"I think these parts are outdated:\r\n\r\n* https://github.com/huggingface/datasets-viewer/blob/8efad8eae313a891f713469983bf4c744786f26e/run.py#L126-L131\r\n* https://github.com/huggingface/datasets-viewer/blob/8efad8eae313a891f713469983bf4c744786f26e/run.py#L145-L150\r\n\r\nTo make the viewer work, the first one should be replaced with the following:\r\n```python\r\ndataset_module = datasets.load.dataset_module_factory(path)\r\nbuilder_cls = datasets.load.import_main_class(dataset_module.module_path)\r\nconfs = builder_cls.BUILDER_CONFIGS\r\n```\r\nAnd the second one:\r\n```python\r\ndataset_module = datasets.load.dataset_module_factory(path)\r\nbuilder_cls = datasets.load.import_main_class(dataset_module.module_path)\r\nif conf:\r\n builder_instance = builder_cls(name=conf, cache_dir=path if path_to_datasets is not None else None)\r\nelse:\r\n builder_instance = builder_cls(cache_dir=path if path_to_datasets is not None else None)\r\n```\r\n\r\nBut as @lhoestq suggested, it's better to use the `datasets-server` API nowadays to [fetch the rows](https://huggingface.co/docs/datasets-server/rows).",
"> The dataset viewer on the Hugging Face website is incredibly useful\r\n\r\n@mariosasko i think @lilyorlilypad wants to run the new dataset-viewer, not the old one",
"> wants to run the new dataset-viewer, not the old one\r\n\r\nThanks for the clarification for me. I do want to run the new dataset-viewer. ",
"It should be possible to run it locally using the HF datasets-server API (docs [here](https://huggingface.co/docs/datasets-server)) but the front end part is not open source (yet ?)\r\n\r\nThe back-end is open source though if you're interested: https://github.com/huggingface/datasets-server\r\nIt automatically converts datasets on HF to Parquet, which is the format we use to power the viewer.",
"the new frontend would probably be hard to open source, as is, as it's quite intertwined with the Hub's code.\r\n\r\nHowever, at some point it would be amazing to have a community-driven open source implementation of a frontend to datasets-server! ",
"For the frontend viewer, see https://github.com/huggingface/datasets/issues/6139.\r\n\r\nAlso mentioned in https://github.com/huggingface/datasets-server/issues/213 and https://github.com/huggingface/datasets-server/issues/441\r\n\r\nClosing as a duplicate of https://github.com/huggingface/datasets/issues/6139",
"Hi team,\r\n\r\nI'm currently researching the Dataset Viewer project and would like to understand more about the frontend technologies used. Specifically, I'm interested in knowing:\r\n\r\nWhich frontend framework is being utilized (e.g., React, Vue, etc.)?\r\nAre there any specific libraries or components being used for UI (e.g., Material-UI, Ant Design)?\r\nAny other notable frontend tools or technologies that are part of this project?\r\nYour assistance in providing these details would be greatly appreciated. Thank you for your time and effort!\r\n\r\nBest regards",
"@jacob-rodgers-max we use https://svelte.dev/",
"> @jacob-rodgers-max we use https://svelte.dev/\r\n\r\nThank you very much for your prompt and detailed response!"
] |
Overview:
The repository (huggingface/datasets-viewer) was recently archived and when I tried to run the code, there was the error message "AttributeError: module 'datasets.load' has no attribute 'prepare_module'". I could not resolve the issue myself due to lack of documentation of that attribute.
Request:
I kindly request the sharing of the code responsible for the dataset preview functionality or help with resolving the error. The dataset viewer on the Hugging Face website is incredibly useful since it is compatible with different types of inputs. It allows users to find datasets that meet their needs more efficiently. If needed, I am willing to contribute to the project by testing, documenting, and providing feedback on the dataset viewer code.
Thank you for considering this request, and I look forward to your response. | 6,014 |
https://github.com/huggingface/datasets/issues/6013 | [FR] `map` should reuse unchanged columns from the previous dataset to avoid disk usage | [
"You can use the `remove_columns` parameter in `map` to avoid duplicating the columns (and save disk space) and then concatenate the original dataset with the map result:\r\n```python\r\nfrom datasets import concatenate_datasets\r\n# dummy example\r\nds_new = ds.map(lambda x: {\"new_col\": x[\"col\"] + 2}, remove_columns=ds.column_names)\r\nds_combined = concatenate_datasets([ds, ds_new], axis=1)\r\n```\r\n\r\nDoing this automatically is hard to implement efficiently unless we know ahead of time which existing columns will be modified by a `map` transform. We have this info when `input_columns` are specified, so I think this is the only case we can optimize."
] | ### Feature request
Currently adding a new column with `map` will cause all the data in the dataset to be duplicated and stored/cached on the disk again. It should reuse unchanged columns.
### Motivation
This allows having datasets with different columns but sharing some basic columns. Currently, these datasets would become too expensive to store and one would need some kind of on-the-fly join; which also doesn't seem implemented.
### Your contribution
_ | 6,013 |
https://github.com/huggingface/datasets/issues/6012 | [FR] Transform Chaining, Lazy Mapping | [
"You can use `with_transform` to get a new dataset object.\r\n\r\nSupport for lazy `map` has already been discussed [here](https://github.com/huggingface/datasets/issues/3385) a little bit. Personally, I'm not a fan, as this would make `map` even more complex. ",
"> You can use `with_transform` to get a new dataset object.\r\n> \r\n> Support for lazy `map` has already been discussed [here](https://github.com/huggingface/datasets/issues/3385) a little bit. Personally, I'm not a fan, as this would make `map` even more complex.\r\n\r\nI read about IterableDataset, and it seems to have lazy mapping. But I can't figure out how to convert an IterableDataset into a normal one when needed.\r\n\r\n`with_transform` still does not chain AFAIU.",
"> I read about IterableDataset, and it seems to have lazy mapping. But I can't figure out how to convert an IterableDataset into a normal one when needed.\r\n\r\nYou must cache an `IterableDataset` to disk to load it as a `Dataset`. One way to do this is with `Dataset.from_generator`:\r\n```python\r\nfrom functools import partial\r\nfrom datasets import Dataset\r\n\r\ndef gen_from_iterable_dataset(iterable_ds)\r\n yield from iterable_ds\r\n\r\nds = Dataset.from_generator(partial(gen_from_iterable_dataset, iterable_ds), features=iterable_ds.features})\r\n```\r\n\r\n> with_transform still does not chain AFAIU.\r\n\r\nYes, not supported yet - the solution is to combine the transforms into a single one.",
"I wonder if it would be beneficial to have a dedicated method to do that ? Maybe a `.save_to_disk()` so that the user can reload the resulting dataset later ?",
"> ```python\r\n> from functools import partial\r\n> from datasets import Dataset\r\n> \r\n> def gen_from_iterable_dataset(iterable_ds)\r\n> yield from iterable_ds\r\n> \r\n> ds = Dataset.from_generator(partial(gen_from_iterable_dataset, iterable_ds), features=iterable_ds.features})\r\n> ```\r\n\r\n@mariosasko With these complex mapping functions, what hash will be used to cache this dataset?\r\n",
"The params passed to `Dataset.from_generator` will be used to compute the hash (`partial` encapsulates the `iterable_ds` value, so changing it will also change the hash)",
"Hi, I think this feature would be very useful. I want to concatenate large datasets with heterogeneous columns. I dislike `map` since I don't want multiple copy of that datasets locally. I tried to use \"set_transform\" on each dataset to convert it to a standard features format, but `datasets.concatenate_datasets` ignores the updated format of the datasets. A work around is to use `torch.utils.data.ConcatDataset`. Is there a neat way to do it using HF datasets?"
] | ### Feature request
Currently using a `map` call processes and duplicates the whole dataset, which takes both time and disk space.
The solution is to allow lazy mapping, which is essentially a saved chain of transforms that are applied on the fly whenever a slice of the dataset is requested.
The API should look like `map`, as `set_transform` changes the current dataset while `map` returns another dataset.
### Motivation
Lazy processing allows lower disk usage and faster experimentation.
### Your contribution
_ | 6,012 |
https://github.com/huggingface/datasets/issues/6011 | Documentation: wiki_dpr Dataset has no metric_type for Faiss Index | [
"Hi! You can do `ds.get_index(\"embeddings\").faiss_index.metric_type` to get the metric type and then match the result with the FAISS metric [enum](https://github.com/facebookresearch/faiss/blob/43d86e30736ede853c384b24667fc3ab897d6ba9/faiss/MetricType.h#L22-L36) (should be L2).",
"Ah! Thank you for pointing this out. FYI: the enum indicates it's using the inner product. Using `torch.inner` or `torch.dot` still produces a discrepancy compared to the built-in score. I think this is because of the compression/quantization that occurs with the FAISS index."
] | ### Describe the bug
After loading `wiki_dpr` using:
```py
ds = load_dataset(path='wiki_dpr', name='psgs_w100.multiset.compressed', split='train')
print(ds.get_index("embeddings").metric_type) # prints nothing because the value is None
```
the index does not have a defined `metric_type`. This is an issue because I do not know how the `scores` are being computed for `get_nearest_examples()`.
### Steps to reproduce the bug
System: Python 3.9.16, Transformers 4.30.2, WSL
After loading `wiki_dpr` using:
```py
ds = load_dataset(path='wiki_dpr', name='psgs_w100.multiset.compressed', split='train')
print(ds.get_index("embeddings").metric_type) # prints nothing because the value is None
```
the index does not have a defined `metric_type`. This is an issue because I do not know how the `scores` are being computed for `get_nearest_examples()`.
```py
from transformers import DPRQuestionEncoder, DPRContextEncoder, DPRQuestionEncoderTokenizer, DPRContextEncoderTokenizer
tokenizer = DPRQuestionEncoderTokenizer.from_pretrained("facebook/dpr-question_encoder-multiset-base")
encoder = DPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-multiset-base")
def encode_question(query, tokenizer=tokenizer, encoder=encoder):
inputs = tokenizer(query, return_tensors='pt')
question_embedding = encoder(**inputs)[0].detach().numpy()
return question_embedding
def get_knn(query, k=5, tokenizer=tokenizer, encoder=encoder, verbose=False):
enc_question = encode_question(query, tokenizer, encoder)
topk_results = ds.get_nearest_examples(index_name='embeddings',
query=enc_question,
k=k)
a = torch.tensor(enc_question[0]).reshape(768)
b = torch.tensor(topk_results.examples['embeddings'][0])
print(a.shape, b.shape)
print(torch.dot(a, b))
print((a-b).pow(2).sum())
return topk_results
```
The [FAISS documentation](https://github.com/facebookresearch/faiss/wiki/MetricType-and-distances) suggests the metric is usually L2 distance (without the square root) or the inner product. I compute both for the sample query:
```py
query = """ it catapulted into popular culture along with a line of action figures and other toys by Bandai.[2] By 2001, the media franchise had generated over $6 billion in toy sales.
Despite initial criticism that its action violence targeted child audiences, the franchise has been commercially successful."""
get_knn(query,k=5)
```
Here, I get dot product of 80.6020 and L2 distance of 77.6616 and
```py
NearestExamplesResults(scores=array([76.20431 , 75.312416, 74.945404, 74.866394, 74.68506 ],
dtype=float32), examples={'id': ['3081096', '2004811', '8908258', '9594124', '286575'], 'text': ['actors, resulting in the "Power Rangers" franchise which has continued since then into sequel TV series (with "Power Rangers Beast Morphers" set to premiere in 2019), comic books, video games, and three feature films, with a further cinematic universe planned. Following from the success of "Power Rangers", Saban acquired the rights to more of Toei\'s library, creating "VR Troopers" and "Big Bad Beetleborgs" from several Metal Hero Series shows and "Masked Rider" from Kamen Rider Series footage. DIC Entertainment joined this boom by acquiring the rights to "Gridman the Hyper Agent" and turning it into "Superhuman Samurai Syber-Squad". In 2002,',
```
Doing `k=1` indicates the higher the outputted number, the better the match, so the metric should not be L2 distance. However, my manually computed inner product (80.6) has a discrepancy with the reported (76.2). Perhaps, this has to do with me using the `compressed` embeddings?
### Expected behavior
```py
ds = load_dataset(path='wiki_dpr', name='psgs_w100.multiset.compressed', split='train')
print(ds.get_index("embeddings").metric_type) # METRIC_INNER_PRODUCT
```
### Environment info
- `datasets` version: 2.12.0
- Platform: Linux-4.18.0-477.13.1.el8_8.x86_64-x86_64-with-glibc2.28
- Python version: 3.9.16
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.1 | 6,011 |
https://github.com/huggingface/datasets/issues/6010 | Improve `Dataset`'s string representation | [
"I want to take a shot at this if possible ",
"Yes, feel free to work on this.\r\n\r\nYou can check the PyArrow Table `__repr__` and Polars DataFrame `__repr__`/`_repr_html_` implementations for some pointers/ideas.",
"@mariosasko are there any other similar issues that I could work on? I see this has been already solved. "
] | Currently, `Dataset.__repr__` outputs a dataset's column names and the number of rows. We could improve it by printing its features and the first few rows.
We should also implement `_repr_html_` to have a rich HTML representation in notebooks/Streamlit. | 6,010 |
https://github.com/huggingface/datasets/issues/6008 | Dataset.from_generator consistently freezes at ~1000 rows | [
"By default, we write data to disk (so it can be memory-mapped) every 1000 rows/samples. You can control this with the `writer_batch_size` parameter. Also, when working with fixed-size arrays, the `ArrayXD` feature types yield better performance (e.g., in your case, `features=datasets.Features({\"i\": datasets.Array3D(shape=(512,512,3), dtype=\"float32\")})` should be faster).\r\n\r\nOur support for multi-dim arrays could be better, and we plan to improve it as part of https://github.com/huggingface/datasets/issues/5272.",
"> By default, we write data to disk (so it can be memory-mapped) every 1000 rows/samples. You can control this with the `writer_batch_size` parameter. Also, when working with fixed-size arrays, the `ArrayXD` feature types yield better performance (e.g., in your case, `features=datasets.Features({\"i\": datasets.Array3D(shape=(512,512,3), dtype=\"float32\")})` should be faster).\r\n> \r\n> Our support for multi-dim arrays could be better, and we plan to improve it as part of #5272.\r\n\r\nThanks for the explanation! The Image array was just for demonstration, I use PIL Images in practice. Does that make a difference? What's the best approach for a dataset with PIL Images as rows?",
"It's best to use the `datasets.Image()` feature type for PIL images (to save space) :)"
] | ### Describe the bug
Whenever I try to create a dataset which contains images using `Dataset.from_generator`, it freezes around 996 rows. I suppose it has something to do with memory consumption, but there's more memory available. I
Somehow it worked a few times but mostly this makes the datasets library much more cumbersome to work with because generators are the easiest way to turn an existing dataset into a Hugging Face dataset.
I've let it run in the frozen state for way longer than it can possibly take to load the actual dataset.
Let me know if you have ideas how to resolve it!
### Steps to reproduce the bug
```python
from datasets import Dataset
import numpy as np
def gen():
for row in range(10000):
yield {"i": np.random.rand(512, 512, 3)}
Dataset.from_generator(gen)
# -> 90% of the time gets stuck around 1000 rows
```
### Expected behavior
Should continue and go through all the examples yielded by the generator, or at least throw an error or somehow communicate what's going on.
### Environment info
- `datasets` version: 2.8.0
- Platform: Linux-5.15.0-52-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 12.0.1
- Pandas version: 1.5.1
| 6,008 |
https://github.com/huggingface/datasets/issues/6007 | Get an error "OverflowError: Python int too large to convert to C long" when loading a large dataset | [
"This error means that one of the int32 (`Value(\"int32\")`) columns in the dataset has a value that is out of the valid (int32) range.\r\n\r\nI'll open a PR to print the name of a problematic column to make debugging such errors easier.",
"I am afraid int32 is not the reason for this error.\r\n\r\nI have submitted a commit to use int64 for all ints in the dataset:\r\nhttps://huggingface.co/datasets/liwu/MNBVC/commit/857ac00d9eab96a6708ad6a82bd9001686042a9e\r\n\r\nand I have updated my env to the latest datasets release:\r\nCopy-and-paste the text below in your GitHub issue.\r\n\r\n- `datasets` version: 2.13.1\r\n- Platform: macOS-13.2.1-arm64-arm-64bit\r\n- Python version: 3.11.2\r\n- Huggingface_hub version: 0.13.4\r\n- PyArrow version: 11.0.0\r\n- Pandas version: 1.5.3\r\n\r\nBut the error still exist\r\n\r\n```\r\nDownloading and preparing dataset mnbvc/news_peoples_daily to /Users/silver/.cache/huggingface/datasets/liwu___mnbvc/news_peoples_daily/0.0.1/ee380f6309fe9b8b0d1fb14d77118f132444f22c8c4b28bf5c1645312688e051...\r\nDownloading data files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 9070.40it/s]\r\nExtracting data files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 2697.16it/s]\r\n---------------------------------------------------------------------------\r\nOverflowError Traceback (most recent call last)\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/builder.py:1647, in GeneratorBasedBuilder._prepare_split_single(self, gen_kwargs, fpath, file_format, max_shard_size, split_info, check_duplicate_keys, job_id)\r\n 1646 example = self.info.features.encode_example(record) if self.info.features is not None else record\r\n-> 1647 writer.write(example, key)\r\n 1648 num_examples_progress_update += 1\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:490, in ArrowWriter.write(self, example, key, writer_batch_size)\r\n 488 self.hkey_record = []\r\n--> 490 self.write_examples_on_file()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:448, in ArrowWriter.write_examples_on_file(self)\r\n 444 batch_examples[col] = [\r\n 445 row[0][col].to_pylist()[0] if isinstance(row[0][col], (pa.Array, pa.ChunkedArray)) else row[0][col]\r\n 446 for row in self.current_examples\r\n 447 ]\r\n--> 448 self.write_batch(batch_examples=batch_examples)\r\n 449 self.current_examples = []\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:553, in ArrowWriter.write_batch(self, batch_examples, writer_batch_size)\r\n 552 typed_sequence = OptimizedTypedSequence(col_values, type=col_type, try_type=col_try_type, col=col)\r\n--> 553 arrays.append(pa.array(typed_sequence))\r\n 554 inferred_features[col] = typed_sequence.get_inferred_type()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:236, in pyarrow.lib.array()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:110, in pyarrow.lib._handle_arrow_array_protocol()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:189, in TypedSequence.__arrow_array__(self, type)\r\n 188 trying_cast_to_python_objects = True\r\n--> 189 out = pa.array(cast_to_python_objects(data, only_1d_for_numpy=True))\r\n 190 # use smaller integer precisions if possible\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:320, in pyarrow.lib.array()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:39, in pyarrow.lib._sequence_to_array()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/error.pxi:144, in pyarrow.lib.pyarrow_internal_check_status()\r\n\r\nOverflowError: Python int too large to convert to C long\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nOverflowError Traceback (most recent call last)\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/builder.py:1656, in GeneratorBasedBuilder._prepare_split_single(self, gen_kwargs, fpath, file_format, max_shard_size, split_info, check_duplicate_keys, job_id)\r\n 1655 num_shards = shard_id + 1\r\n-> 1656 num_examples, num_bytes = writer.finalize()\r\n 1657 writer.close()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:584, in ArrowWriter.finalize(self, close_stream)\r\n 583 self.hkey_record = []\r\n--> 584 self.write_examples_on_file()\r\n 585 # If schema is known, infer features even if no examples were written\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:448, in ArrowWriter.write_examples_on_file(self)\r\n 444 batch_examples[col] = [\r\n 445 row[0][col].to_pylist()[0] if isinstance(row[0][col], (pa.Array, pa.ChunkedArray)) else row[0][col]\r\n 446 for row in self.current_examples\r\n 447 ]\r\n--> 448 self.write_batch(batch_examples=batch_examples)\r\n 449 self.current_examples = []\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:553, in ArrowWriter.write_batch(self, batch_examples, writer_batch_size)\r\n 552 typed_sequence = OptimizedTypedSequence(col_values, type=col_type, try_type=col_try_type, col=col)\r\n--> 553 arrays.append(pa.array(typed_sequence))\r\n 554 inferred_features[col] = typed_sequence.get_inferred_type()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:236, in pyarrow.lib.array()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:110, in pyarrow.lib._handle_arrow_array_protocol()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/arrow_writer.py:189, in TypedSequence.__arrow_array__(self, type)\r\n 188 trying_cast_to_python_objects = True\r\n--> 189 out = pa.array(cast_to_python_objects(data, only_1d_for_numpy=True))\r\n 190 # use smaller integer precisions if possible\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:320, in pyarrow.lib.array()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/array.pxi:39, in pyarrow.lib._sequence_to_array()\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/pyarrow/error.pxi:144, in pyarrow.lib.pyarrow_internal_check_status()\r\n\r\nOverflowError: Python int too large to convert to C long\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nDatasetGenerationError Traceback (most recent call last)\r\nCell In[2], line 1\r\n----> 1 dataset = load_dataset(\"liwu/MNBVC\", 'news_peoples_daily', split='train')\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/load.py:1809, in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, verification_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, num_proc, storage_options, **config_kwargs)\r\n 1806 try_from_hf_gcs = path not in _PACKAGED_DATASETS_MODULES\r\n 1808 # Download and prepare data\r\n-> 1809 builder_instance.download_and_prepare(\r\n 1810 download_config=download_config,\r\n 1811 download_mode=download_mode,\r\n 1812 verification_mode=verification_mode,\r\n 1813 try_from_hf_gcs=try_from_hf_gcs,\r\n 1814 num_proc=num_proc,\r\n 1815 storage_options=storage_options,\r\n 1816 )\r\n 1818 # Build dataset for splits\r\n 1819 keep_in_memory = (\r\n 1820 keep_in_memory if keep_in_memory is not None else is_small_dataset(builder_instance.info.dataset_size)\r\n 1821 )\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/builder.py:909, in DatasetBuilder.download_and_prepare(self, output_dir, download_config, download_mode, verification_mode, ignore_verifications, try_from_hf_gcs, dl_manager, base_path, use_auth_token, file_format, max_shard_size, num_proc, storage_options, **download_and_prepare_kwargs)\r\n 907 if num_proc is not None:\r\n 908 prepare_split_kwargs[\"num_proc\"] = num_proc\r\n--> 909 self._download_and_prepare(\r\n 910 dl_manager=dl_manager,\r\n 911 verification_mode=verification_mode,\r\n 912 **prepare_split_kwargs,\r\n 913 **download_and_prepare_kwargs,\r\n 914 )\r\n 915 # Sync info\r\n 916 self.info.dataset_size = sum(split.num_bytes for split in self.info.splits.values())\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/builder.py:1670, in GeneratorBasedBuilder._download_and_prepare(self, dl_manager, verification_mode, **prepare_splits_kwargs)\r\n 1669 def _download_and_prepare(self, dl_manager, verification_mode, **prepare_splits_kwargs):\r\n-> 1670 super()._download_and_prepare(\r\n 1671 dl_manager,\r\n 1672 verification_mode,\r\n 1673 check_duplicate_keys=verification_mode == VerificationMode.BASIC_CHECKS\r\n 1674 or verification_mode == VerificationMode.ALL_CHECKS,\r\n 1675 **prepare_splits_kwargs,\r\n 1676 )\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/builder.py:1004, in DatasetBuilder._download_and_prepare(self, dl_manager, verification_mode, **prepare_split_kwargs)\r\n 1000 split_dict.add(split_generator.split_info)\r\n 1002 try:\r\n 1003 # Prepare split will record examples associated to the split\r\n-> 1004 self._prepare_split(split_generator, **prepare_split_kwargs)\r\n 1005 except OSError as e:\r\n 1006 raise OSError(\r\n 1007 \"Cannot find data file. \"\r\n 1008 + (self.manual_download_instructions or \"\")\r\n 1009 + \"\\nOriginal error:\\n\"\r\n 1010 + str(e)\r\n 1011 ) from None\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/builder.py:1508, in GeneratorBasedBuilder._prepare_split(self, split_generator, check_duplicate_keys, file_format, num_proc, max_shard_size)\r\n 1506 job_id = 0\r\n 1507 with pbar:\r\n-> 1508 for job_id, done, content in self._prepare_split_single(\r\n 1509 gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args\r\n 1510 ):\r\n 1511 if done:\r\n 1512 result = content\r\n\r\nFile ~/git/venv/lib/python3.11/site-packages/datasets/builder.py:1665, in GeneratorBasedBuilder._prepare_split_single(self, gen_kwargs, fpath, file_format, max_shard_size, split_info, check_duplicate_keys, job_id)\r\n 1663 if isinstance(e, SchemaInferenceError) and e.__context__ is not None:\r\n 1664 e = e.__context__\r\n-> 1665 raise DatasetGenerationError(\"An error occurred while generating the dataset\") from e\r\n 1667 yield job_id, True, (total_num_examples, total_num_bytes, writer._features, num_shards, shard_lengths)\r\n\r\nDatasetGenerationError: An error occurred while generating the dataset\r\n```\r\n\r\nBesides, it works fine when I am using streamed dataset.",
"`simhash` is the problematic column - it has values such as `18329103420363166823` that are out of the int64 range. You can fix this by setting the feature type to `Value(\"string\")` (it's advised to use this type for hash values in general)\r\n\r\n> Besides, it works fine when I am using streamed dataset.\r\n\r\nStreaming yields Python dictionaries from the script without converting them to the Arrow representation, as this conversion step is not that cheap performance-wise.",
"i am using uint64 for simhash\r\n\r\nuint64 ranges up to about 3.69E19.\r\n\r\n18329103420363166823 is less than this value.\r\n\r\nmoreover, our simhash algorithm use 64 bits. it should fit in uint64.\r\n\r\n\r\n\r\n",
"You are right. I overlooked the feature type.\r\n\r\nThis is a reproducer:\r\n```python\r\nimport pyarrow as pa\r\nfrom datasets.arrow_writer import TypedSequence\r\n\r\npa.array(TypedSequence([18329103420363166823], type=Value(\"uint64\")))\r\n```\r\n\r\n`pa.array([18329103420363166823])` also fails with the same error, so it seems PyArrow does not always infer the correct type as NumPy does (`uint64` in this case).\r\n\r\nI'll report this issue in the Arrow repo.\r\n\r\n`pa.array([18329103420363166823], pa.uint64)` works, so maybe we can implement a temporary fix (supporting complex input such as `[{\"image\": pil_image, \"num\": uint64_value}]` would be hard though).\r\n\r\nIn the meantime, you should be able to bypass this error by returning the `simhash` values as NumPy scalars in the script:\r\n```python\r\ndef _generate_examples(self, ...):\r\n ...\r\n yield {..., \"simhash\": np.uint64(simhash), ...}\r\n```",
"Thank you for checking this issue in detail.\r\n\r\nHowever, it seems that using `np.uint64(simhash)` does not work. The same issue still exists.\r\n\r\nhttps://huggingface.co/datasets/liwu/MNBVC/commit/1e44f1e400b7e61052647d44c99cdae3bae9c830\r\n\r\nAnyway, we decide to use string type for these simhash values. Hope pyarrow can fix their bug soon.",
"Arrow issue: https://github.com/apache/arrow/issues/36520",
"May be something read your training data line by line.\r\nThen your training data just only one line. \r\nIt is so large.\r\nI guess.\r\n"
] | ### Describe the bug
When load a large dataset with the following code
```python
from datasets import load_dataset
dataset = load_dataset("liwu/MNBVC", 'news_peoples_daily', split='train')
```
We encountered the error: "OverflowError: Python int too large to convert to C long"
The error look something like:
```
OverflowError: Python int too large to convert to C long
During handling of the above exception, another exception occurred:
OverflowError Traceback (most recent call last)
<ipython-input-7-0ed8700e662d> in <module>
----> 1 dataset = load_dataset("liwu/MNBVC", 'news_peoples_daily', split='train', cache_dir='/sfs/MNBVC/.cache/')
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, **config_kwargs)
1749 ignore_verifications=ignore_verifications,
1750 try_from_hf_gcs=try_from_hf_gcs,
-> 1751 use_auth_token=use_auth_token,
1752 )
1753
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/builder.py in download_and_prepare(self, download_config, download_mode, ignore_verifications, try_from_hf_gcs, dl_manager, base_path, use_auth_token, **download_and_prepare_kwargs)
703 if not downloaded_from_gcs:
704 self._download_and_prepare(
--> 705 dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
706 )
707 # Sync info
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/builder.py in _download_and_prepare(self, dl_manager, verify_infos)
1225
1226 def _download_and_prepare(self, dl_manager, verify_infos):
-> 1227 super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
1228
1229 def _get_examples_iterable_for_split(self, split_generator: SplitGenerator) -> ExamplesIterable:
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/builder.py in _download_and_prepare(self, dl_manager, verify_infos, **prepare_split_kwargs)
791 try:
792 # Prepare split will record examples associated to the split
--> 793 self._prepare_split(split_generator, **prepare_split_kwargs)
794 except OSError as e:
795 raise OSError(
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/builder.py in _prepare_split(self, split_generator, check_duplicate_keys)
1219 writer.write(example, key)
1220 finally:
-> 1221 num_examples, num_bytes = writer.finalize()
1222
1223 split_generator.split_info.num_examples = num_examples
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/arrow_writer.py in finalize(self, close_stream)
536 # Re-intializing to empty list for next batch
537 self.hkey_record = []
--> 538 self.write_examples_on_file()
539 if self.pa_writer is None:
540 if self.schema:
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/arrow_writer.py in write_examples_on_file(self)
407 # Since current_examples contains (example, key) tuples
408 batch_examples[col] = [row[0][col] for row in self.current_examples]
--> 409 self.write_batch(batch_examples=batch_examples)
410 self.current_examples = []
411
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/arrow_writer.py in write_batch(self, batch_examples, writer_batch_size)
506 col_try_type = try_features[col] if try_features is not None and col in try_features else None
507 typed_sequence = OptimizedTypedSequence(batch_examples[col], type=col_type, try_type=col_try_type, col=col)
--> 508 arrays.append(pa.array(typed_sequence))
509 inferred_features[col] = typed_sequence.get_inferred_type()
510 schema = inferred_features.arrow_schema if self.pa_writer is None else self.schema
/sfs/MNBVC/venv/lib64/python3.6/site-packages/pyarrow/array.pxi in pyarrow.lib.array()
/sfs/MNBVC/venv/lib64/python3.6/site-packages/pyarrow/array.pxi in pyarrow.lib._handle_arrow_array_protocol()
/sfs/MNBVC/venv/lib64/python3.6/site-packages/datasets/arrow_writer.py in __arrow_array__(self, type)
180 else:
181 trying_cast_to_python_objects = True
--> 182 out = pa.array(cast_to_python_objects(data, only_1d_for_numpy=True))
183 # use smaller integer precisions if possible
184 if self.trying_int_optimization:
/sfs/MNBVC/venv/lib64/python3.6/site-packages/pyarrow/array.pxi in pyarrow.lib.array()
/sfs/MNBVC/venv/lib64/python3.6/site-packages/pyarrow/array.pxi in pyarrow.lib._sequence_to_array()
/sfs/MNBVC/venv/lib64/python3.6/site-packages/pyarrow/error.pxi in pyarrow.lib.pyarrow_internal_check_status()
OverflowError: Python int too large to convert to C long
```
However, that dataset can be loaded in a streaming manner:
```python
from datasets import load_dataset
dataset = load_dataset("liwu/MNBVC", 'news_peoples_daily', split='train', streaming=True)
for i in dataset:
pass # it work well
```
Another issue is reported in our dataset hub:
https://huggingface.co/datasets/liwu/MNBVC/discussions/2
### Steps to reproduce the bug
from datasets import load_dataset
dataset = load_dataset("liwu/MNBVC", 'news_peoples_daily', split='train')
### Expected behavior
the dataset can be safely loaded
### Environment info
- `datasets` version: 2.4.0
- Platform: Linux-3.10.0-1160.an7.x86_64-x86_64-with-centos-7.9
- Python version: 3.6.8
- PyArrow version: 6.0.1
- Pandas version: 1.1.5 | 6,007 |
https://github.com/huggingface/datasets/issues/6006 | NotADirectoryError when loading gigawords | [
"issue due to corrupted download files. resolved after cleaning download cache. sorry for any inconvinence."
] | ### Describe the bug
got `NotADirectoryError` whtn loading gigawords dataset
### Steps to reproduce the bug
When running
```
import datasets
datasets.load_dataset('gigaword')
```
Got the following exception:
```bash
Traceback (most recent call last): [0/1862]
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/builder.py", line 1629, in _prepare_split_single
for key, record in generator:
File "/home/x/.cache/huggingface/modules/datasets_modules/datasets/gigaword/ea83a8b819190acac5f2dae011fad51dccf269a0604ec5dd24795b
64efb424b6/gigaword.py", line 115, in _generate_examples
with open(src_path, encoding="utf-8") as f_d, open(tgt_path, encoding="utf-8") as f_s:
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/streaming.py", line 71, in wrapper
return function(*args, use_auth_token=use_auth_token, **kwargs)
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/download/streaming_download_manager.py", line 493, in xope
n
return open(main_hop, mode, *args, **kwargs)
NotADirectoryError: [Errno 20] Not a directory: '/home/x/.cache/huggingface/datasets/downloads/6da52431bb5124d90cf51a0187d2dbee9046e
89780c4be7599794a4f559048ec/org_data/train.src.txt'
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "gigaword.py", line 38, in <module>
main()
File "gigaword.py", line 35, in main
train, dev, test = dataset.generate_k_shot_data(k=32, seed=seed, path="../data/")
File "/home/x/MICL/preprocess/fewshot_gym_dataset.py", line 199, in generate_k_shot_data
dataset = self.load_dataset()
File "gigaword.py", line 29, in load_dataset
return datasets.load_dataset('gigaword')
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/load.py", line 1809, in load_dataset
builder_instance.download_and_prepare(
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/builder.py", line 909, in download_and_prepare
self._download_and_prepare(
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/builder.py", line 1670, in _download_and_prepare
super()._download_and_prepare(
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/builder.py", line 1004, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/builder.py", line 1508, in _prepare_split
for job_id, done, content in self._prepare_split_single(
File "/home/x/.conda/envs/dataproc/lib/python3.8/site-packages/datasets/builder.py", line 1665, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.builder.DatasetGenerationError: An error occurred while generating the dataset
```
### Expected behavior
Download and process the dataset successfully
### Environment info
- `datasets` version: 2.13.1
- Platform: Linux-5.0.0-1032-azure-x86_64-with-glibc2.10
- Python version: 3.8.0
- Huggingface_hub version: 0.15.1
- PyArrow version: 12.0.1
- Pandas version: 2.0.3
| 6,006 |
https://github.com/huggingface/datasets/issues/6003 | interleave_datasets & DataCollatorForLanguageModeling having a conflict ? | [] | ### Describe the bug
Hi everyone :)
I have two local & custom datasets (1 "sentence" per line) which I split along the 95/5 lines for pre-training a Bert model. I use a modified version of `run_mlm.py` in order to be able to make use of `interleave_dataset`:
- `tokenize()` runs fine
- `group_text()` runs fine
Everytime, on step 19, I get
```pytb
File "env/lib/python3.9/site-packages/transformers/data/data_collator.py", line 779, in torch_mask_tokens
inputs[indices_random] = random_words[indices_random]
RuntimeError: Index put requires the source and destination dtypes match, got Float for the destination and Long for the source.
```
I tried:
- training without interleave on dataset 1, it runs
- training without interleave on dataset 2, it runs
- training without `.to_iterable_dataset()`, it hangs then crash
- training without group_text() and padding to max_length seemed to fix the issue, but who knows if this was just because it was an issue that would come much later in terms of steps.
I might have coded something wrong, but I don't get what
### Steps to reproduce the bug
I have this function:
```py
def build_dataset(path: str, percent: str):
dataset = load_dataset(
"text",
data_files={"train": [path]},
split=f"train[{percent}]"
)
dataset = dataset.map(
lambda examples: tokenize(examples["text"]),
batched=True,
num_proc=num_proc,
)
dataset = dataset.map(
group_texts,
batched=True,
num_proc=num_proc,
desc=f"Grouping texts in chunks of {tokenizer.max_seq_length}",
remove_columns=["text"]
)
print(len(dataset))
return dataset.to_iterable_dataset()
```
I hardcoded group_text:
```py
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, and if the total_length < max_seq_length we exclude this batch and return an empty dict.
# We could add padding if the model supported it instead of this drop, you can customize this part to your needs.
total_length = (total_length // 512) * 512
# Split by chunks of max_len.
result = {
k: [t[i: i + 512] for i in range(0, total_length, 512)]
for k, t in concatenated_examples.items()
}
# result = {k: [el for el in elements if el] for k, elements in result.items()}
return result
```
And then I build datasets using the following code:
```py
train1 = build_dataset("d1.txt", ":95%")
train2 = build_dataset("d2.txt", ":95%")
dev1 = build_dataset("d1.txt", "95%:")
dev2 = build_dataset("d2.txt", "95%:")
```
and finally I run
```py
train_dataset = interleave_datasets(
[train1, train2],
probabilities=[0.8, 0.2],
seed=42
)
eval_dataset = interleave_datasets(
[dev1, dev2],
probabilities=[0.8, 0.2],
seed=42
)
```
Then I run the training part which remains mostly untouched:
> CUDA_VISIBLE_DEVICES=1 python custom_dataset.py --model_type bert --per_device_train_batch_size 32 --do_train --output_dir /var/mlm/training-bert/model --max_seq_length 512 --save_steps 10000 --save_total_limit 3 --auto_find_batch_size --logging_dir ./logs-bert --learning_rate 0.0001 --do_train --num_train_epochs 25 --warmup_steps 10000 --max_step 45000 --fp16
### Expected behavior
The model should then train normally, but fails every time at the same step (19).
printing the variables at `inputs[indices_random] = random_words[indices_random]` shows a magnificient empty tensor (, 32) [if I remember well]
### Environment info
transformers[torch] 4.30.2
Ubuntu
A100 0 CUDA 12
Driver Version: 525.116.04 | 6,003 |
https://github.com/huggingface/datasets/issues/5999 | Getting a 409 error while loading xglue dataset | [
"Thanks for reporting, @Praful932.\r\n\r\nLet's continue the conversation on the Hub: https://huggingface.co/datasets/xglue/discussions/5"
] | ### Describe the bug
Unable to load xglue dataset
### Steps to reproduce the bug
```python
import datasets
dataset = datasets.load_dataset("xglue", "ntg")
```
> ConnectionError: Couldn't reach https://xglue.blob.core.windows.net/xglue/xglue_full_dataset.tar.gz (error 409)
### Expected behavior
Expected the dataset to load
### Environment info
- `datasets` version: 2.13.1
- Platform: Linux-5.15.107+-x86_64-with-glibc2.31
- Python version: 3.10.12
- Huggingface_hub version: 0.15.1
- PyArrow version: 9.0.0
- Pandas version: 1.5.3 | 5,999 |
https://github.com/huggingface/datasets/issues/5998 | The current implementation has a potential bug in the sort method | [
"Thanks for reporting, @wangyuxinwhy. "
] | ### Describe the bug
In the sort method，here's a piece of code
```python
# column_names: Union[str, Sequence_[str]]
# Check proper format of and for duplicates in column_names
if not isinstance(column_names, list):
column_names = [column_names]
```
I get an error when I pass in a tuple based on the column_names type annotation, it will raise an errror.As in the example below, while the type annotation implies that a tuple can be passed.
```python
from datasets import load_dataset
dataset = load_dataset('glue', 'ax')['test']
dataset.sort(column_names=('premise', 'hypothesis'))
# Raise ValueError: Column '('premise', 'hypothesis')' not found in the dataset.
```
Of course, after I modified the tuple into a list, everything worked fine
Change the code to the following so there will be no problem
```python
# Check proper format of and for duplicates in column_names
if not isinstance(column_names, list):
if isinstance(column_names, str):
column_names = [column_names]
else:
column_names = list(column_names)
```
### Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('glue', 'ax')['test']
dataset.sort(column_names=('premise', 'hypothesis'))
# Raise ValueError: Column '('premise', 'hypothesis')' not found in the dataset.
```
### Expected behavior
Passing tuple into column_names should be equivalent to passing list
### Environment info
- `datasets` version: 2.13.0
- Platform: macOS-13.1-arm64-arm-64bit
- Python version: 3.10.11
- Huggingface_hub version: 0.15.1
- PyArrow version: 12.0.1
- Pandas version: 2.0.2 | 5,998 |
https://github.com/huggingface/datasets/issues/5997 | extend the map function so it can wrap around long text that does not fit in the context window | [
"I just noticed the [docs](https://github.com/huggingface/datasets/blob/main/src/datasets/arrow_dataset.py#L2881C11-L2881C200) say:\r\n\r\n>If batched is `True` and `batch_size` is `n > 1`, then the function takes a batch of `n` examples as input and can return a batch with `n` examples, or with an arbitrary number of examples.\r\n\r\nso maybe this is a bug then.",
"All the values in a batch must be of the same length. So one solution is dropping all the input columns:\r\n```python\r\ndata = data.map(lambda samples: tokenizer(samples[\"text\"], max_length=tokenizer.model_max_length, truncation=True, stride=4, return_overflowing_tokens=True), batched=True, remove_columns=data.column_names)\r\n```\r\n\r\nAnother is padding/transforming the input columns to the tokenizer output's length (447). "
] | ### Feature request
I understand `dataset` provides a [`map`](https://github.com/huggingface/datasets/blob/main/src/datasets/arrow_dataset.py#L2849) function. This function in turn takes in a callable that is used to tokenize the text on which a model is trained. Frequently this text will not fit within a models's context window. In this case it would be useful to wrap around the text into multiple rows with each row fitting the model's context window. I tried to do it using this code as example which in turn I have borrowed from [here](https://stackoverflow.com/a/76343993/147530):
```
data = data.map(lambda samples: tokenizer(samples["text"], max_length=tokenizer.model_max_length, truncation=True, stride=4, return_overflowing_tokens=True), batched=True)
```
but running the code gives me this error:
```
File "/llm/fine-tune.py", line 117, in <module>
data = data.map(lambda samples: tokenizer(samples["text"], max_length=tokenizer.model_max_length, truncation=True, stride=4, return_overflowing_tokens=True), batched=True)
File "/llm/.env/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 580, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/llm/.env/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 545, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/llm/.env/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3087, in map
for rank, done, content in Dataset._map_single(**dataset_kwargs):
File "/llm/.env/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3480, in _map_single
writer.write_batch(batch)
File "/llm/.env/lib/python3.9/site-packages/datasets/arrow_writer.py", line 556, in write_batch
pa_table = pa.Table.from_arrays(arrays, schema=schema)
File "pyarrow/table.pxi", line 3798, in pyarrow.lib.Table.from_arrays
File "pyarrow/table.pxi", line 2962, in pyarrow.lib.Table.validate
File "pyarrow/error.pxi", line 100, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: Column 1 named input_ids expected length 394 but got length 447
```
The lambda function I have provided is correctly chopping up long text so it wraps around (and because of this 394 samples become 447 after wrap around) but the dataset `map` function does not like it.
### Motivation
please see above
### Your contribution
I'm afraid I don't have much knowledge to help | 5,997 |
https://github.com/huggingface/datasets/issues/5993 | ValueError: Table schema does not match schema used to create file | [
"We'll do a new release of `datasets` soon to make the fix available :)\r\n\r\nIn the meantime you can use `datasets` from source (main)",
"Thank you very much @lhoestq ! 🚀 "
] | ### Describe the bug
Saving a dataset as parquet fails with a `ValueError: Table schema does not match schema used to create file` if the dataset was obtained out of a `.select_columns()` call with columns selected out of order.
### Steps to reproduce the bug
```python
import datasets
dataset = datasets.Dataset.from_dict(
{
"x1": [1, 2, 3],
"x2": [10, 11, 12],
}
)
ds = dataset.select_columns(["x2", "x1"])
ds.to_parquet("demo.parquet")
```
```shell
>>>
ValueError: Table schema does not match schema used to create file:
table:
x2: int64
x1: int64
-- schema metadata --
huggingface: '{"info": {"features": {"x2": {"dtype": "int64", "_type": "V' + 53 vs.
file:
x1: int64
x2: int64
-- schema metadata --
huggingface: '{"info": {"features": {"x1": {"dtype": "int64", "_type": "V' + 53
```
---
I think this is because after the `.select_columns()` call with out of order columns, the output dataset features' schema ends up being out of sync with the schema of the arrow table backing it.
```python
ds.features.arrow_schema
>>>
x1: int64
x2: int64
-- schema metadata --
huggingface: '{"info": {"features": {"x1": {"dtype": "int64", "_type": "V' + 53
ds.data.schema
>>>
x2: int64
x1: int64
-- schema metadata --
huggingface: '{"info": {"features": {"x2": {"dtype": "int64", "_type": "V' + 53
```
So when we call `.to_parquet()`, the call behind the scenes to `datasets.io.parquet.ParquetDatasetWriter(...).write()` which initialises the backend `pyarrow.parquet.ParquetWriter` with `schema = self.dataset.features.arrow_schema` triggers `pyarrow` on write when [it checks](https://github.com/apache/arrow/blob/11b140a734a516e436adaddaeb35d23f30dcce44/python/pyarrow/parquet/core.py#L1086-L1090) that the `ParquetWriter` schema matches the schema of the table being written 🙌
https://github.com/huggingface/datasets/blob/6ed837325cb539a5deb99129e5ad181d0269e050/src/datasets/io/parquet.py#L139-L141
### Expected behavior
The dataset gets successfully saved as parquet.
*In the same way as it does if saving it as csv:
```python
import datasets
dataset = datasets.Dataset.from_dict(
{
"x1": [1, 2, 3],
"x2": [10, 11, 12],
}
)
ds = dataset.select_columns(["x2", "x1"])
ds.to_csv("demo.csv")
```
### Environment info
`python==3.11`
`datasets==2.13.1`
| 5,993 |
https://github.com/huggingface/datasets/issues/5991 | `map` with any joblib backend | [] | We recently enabled the (experimental) parallel backend switch for data download and extraction but not for `map` yet.
Right now we're using our `iflatmap_unordered` implementation for multiprocessing that uses a shared Queue to gather progress updates from the subprocesses and show a progress bar in the main process.
If a Queue implementation that would work on any joblib backend by leveraging the filesystem that is shared among workers, we can have `iflatmap_unordered` for joblib and therefore a `map` with any joblib backend with a progress bar !
Note that the Queue doesn't need to be that optimized though since we can choose a small frequency for progress updates (like 1 update per second). | 5,991 |
https://github.com/huggingface/datasets/issues/5989 | Set a rule on the config and split names | [
"in this case we need to decide what to do with the existing datasets with white space characters (there shouldn't be a lot of them I think)",
"I imagine that we should stop supporting them, and help the user fix them?",
"See a report where the datasets server fails: https://huggingface.co/datasets/poloclub/diffusiondb/discussions/2#6374ff55b93cbdf65675f564\r\n\r\nThe config name is `random_10k [2m]`!"
] | > should we actually allow characters like spaces? maybe it's better to add validation for whitespace symbols and directly in datasets and raise
https://github.com/huggingface/datasets-server/issues/853
| 5,989 |
https://github.com/huggingface/datasets/issues/5988 | ConnectionError: Couldn't reach dataset_infos.json | [
"Unfortunately, I can't reproduce the error. What does the following code return for you?\r\n```python\r\nimport requests\r\nfrom huggingface_hub import hf_hub_url\r\nr = requests.get(hf_hub_url(\"codeparrot/codeparrot-clean-train\", \"dataset_infos.json\", repo_type=\"dataset\"))\r\n```\r\n\r\nAlso, can you provide more info about your network (region, proxies, etc.)?"
] | ### Describe the bug
I'm trying to load codeparrot/codeparrot-clean-train, but get the following error:
ConnectionError: Couldn't reach https://huggingface.co/datasets/codeparrot/codeparrot-clean-train/resolve/main/dataset_infos.json (ConnectionError(ProtocolError('Connection aborted.', ConnectionResetError(104, 'Connection reset by peer'))))
### Steps to reproduce the bug
train_data = load_dataset('codeparrot/codeparrot-clean-train', split='train')
### Expected behavior
download the dataset
### Environment info
centos7 | 5,988 |
https://github.com/huggingface/datasets/issues/5987 | Why max_shard_size is not supported in load_dataset and passed to download_and_prepare | [
"Can you explain your use case for `max_shard_size`? \r\n\r\nOn some systems, there is a limit to the size of a memory-mapped file, so we could consider exposing this parameter in `load_dataset`.",
"In my use case, users may choose a proper size to balance the cost and benefit of using large shard size. (On azure blob or hdfs which may automatically download the shard from background)",
"But `load_dataset` doesn't support caching (and reading) Arrow datasets from remote storage. \r\n\r\n`load_datset_builder` + `download_and_prepare` is not equal to `load_dataset`. The latter has one more step, `builder.as_dataset`, that memory-maps Arrow files, which only works for local files.",
"Thanks. So if I want to use `IterableDataset` and control the size of single arrow file, how should I organize the data loader? Maybe `load_dataset_build` + `download_and_prepare` + `builder.as_dataset` + `dataset.to_iterable_dataset`?",
"Yes, this should work.\r\n\r\nI think we can expose `max_shard_size` in `load_dataset`, so feel free to open a PR."
] | ### Describe the bug
https://github.com/huggingface/datasets/blob/a8a797cc92e860c8d0df71e0aa826f4d2690713e/src/datasets/load.py#L1809
What I can to is break the `load_dataset` and use `load_datset_builder` + `download_and_prepare` instead.
### Steps to reproduce the bug
https://github.com/huggingface/datasets/blob/a8a797cc92e860c8d0df71e0aa826f4d2690713e/src/datasets/load.py#L1809
### Expected behavior
Users can define the max shard size.
### Environment info
datasets==2.13.1 | 5,987 |
https://github.com/huggingface/datasets/issues/5985 | Cannot reuse tokenizer object for dataset map | [
"This is a known issue: https://github.com/huggingface/datasets/issues/3847.\r\n\r\nFixing this requires significant work - rewriting the `tokenizers` lib to make them immutable.\r\n\r\nThe current solution is to pass `cache_file_name` to `map` to use that file for caching or calling a tokenizer before `map` (with the same set of parameters as the ones in the map transform)",
"Closing since this is a duplicate"
] | ### Describe the bug
Related to https://github.com/huggingface/transformers/issues/24441. Not sure if this is a tokenizer issue or caching issue, so filing in both.
Passing the tokenizer to the dataset map function causes the tokenizer to be fingerprinted weirdly. After calling the tokenizer with arguments like padding and truncation the tokenizer object changes interanally, even though the hash remains the same.
But dumps is able to detect that internal change which causes the tokenizer object's fingerprint to change.
### Steps to reproduce the bug
```python
from transformers import AutoTokenizer
from datasets.utils.py_utils import dumps # Huggingface datasets
t = AutoTokenizer.from_pretrained('bert-base-uncased')
t.save_pretrained("tok1")
th1 = hash(dumps(t))
text = "This is an example text"
ttext = t(text, max_length=512, padding="max_length", truncation=True)
t.save_pretrained("tok2")
th2 = hash(dumps(t))
assert th1 == th2 # Assertion Error
```
But if you use just the hash of the object without dumps, the hashes don't change
```python
from transformers import AutoTokenizer
from datasets.utils.py_utils import dumps # Huggingface datasets
t = AutoTokenizer.from_pretrained('bert-base-uncased')
th1 = hash(t) # Just hash no dumps
text = "This is an example text"
ttext = t(text, max_length=512, padding="max_length", truncation=True)
th2 = hash(t) # Just hash no dumps
assert th1 == th2 # This is OK
```
This causes situations such as the following
1. Create a text file like this `yes "This is an example text" | head -n 10000 > lines.txt`
```python
from transformers import AutoTokenizer
import datasets
class TokenizeMapper(object):
"""Mapper for tokenizer.
This is needed because the caching mechanism of HuggingFace does not work on
lambdas. Each time a new lambda will be created by a new process which will
lead to a different hash.
This way we can have a universal mapper object in init and reuse it with the same
hash for each process.
"""
def __init__(self, tokenizer):
"""Initialize the tokenizer."""
self.tokenizer = tokenizer
def __call__(self, examples, **kwargs):
"""Run the mapper."""
texts = examples["text"]
tt = self.tokenizer(texts, max_length=256, padding="max_length", truncation=True)
batch_outputs = {
"input_ids": tt.input_ids,
"attention_mask": tt.attention_mask,
}
return batch_outputs
t = AutoTokenizer.from_pretrained('bert-base-uncased')
mapper = TokenizeMapper(t)
ds = datasets.load_dataset("text", data_files="lines.txt")
mds1 = ds.map(
mapper,
batched=False,
remove_columns=["text"],
).with_format("torch")
mds2 = ds.map(
mapper,
batched=False,
remove_columns=["text"],
).with_format("torch")
```
The second call to map should reuse the cached processed dataset from mds1, but it instead it redoes the tokenization because of the behavior of dumps.
### Expected behavior
We should be able to initialize a tokenizer. And reusing it should let us reuse the same map computation for the same dataset.
The second call to map should reuse the cached processed dataset from mds1, but it instead it redoes the tokenization because of the behavior of dumps.
### Environment info
- `datasets` version: 2.13.0
- Platform: Linux-6.1.31_1-x86_64-with-glibc2.36
- Python version: 3.9.16
- Huggingface_hub version: 0.15.1
- PyArrow version: 12.0.1
- Pandas version: 2.0.2 | 5,985 |
https://github.com/huggingface/datasets/issues/5984 | AutoSharding IterableDataset's when num_workers > 1 | [
"For this to be possible, we would have to switch from the \"Streaming\" Arrow format to the \"Random Access\" (IPC/Feather) format, which allows reading arbitrary record batches (explained [here](https://arrow.apache.org/docs/python/ipc.html)). We could then use these batches to construct shards.\r\n\r\n@lhoestq @albertvillanova Do you think this use case is worth the switch? Also, we currently shard files, not inner row groups/chunks. Should we also support sharding row groups (e.g. if the number of input files is 1)?\r\n\r\nPS: I don't expect significant speed-up for local, uncompressed Arrow files.",
"Alternatively we could support multiprocessing map for iterable datasets and let the user do the CPU intensive task there ?\r\n\r\nThis way it would work on arrow data but also on any iterable dataset",
"> For this to be possible, we would have to switch from the \"Streaming\" Arrow format to the \"Random Access\" (IPC/Feather) format, which allows reading arbitrary record batches (explained [here](https://arrow.apache.org/docs/python/ipc.html)). We could then use these batches to construct shards.\r\n> \r\n> @lhoestq @albertvillanova Do you think this use case is worth the switch? Also, we currently shard files, not inner row groups/chunks. Should we also support sharding row groups (e.g. if the number of input files is 1)?\r\n> \r\n> PS: I don't expect significant speed-up for local, uncompressed Arrow files.\r\n\r\nCould you explain why you'd need to change the arrow format?\r\n\r\nWhen we use streaming datasets we simply determine the number of worker shards and then add some modulo logic at the appropriate place. Worst case scenario, you'd skip streaming entries according to the number of shards.\r\n\r\nFor PyTorch, I'd be happy to provide an implementation or a sketch thereof, if you point me toward what the testing requirements would be for such a PR.",
"> Could you explain why you'd need to change the arrow format?\r\n\r\nThis way workers have random access to the location of the file where its dataset subset starts. Currently we're using the Arrow streaming format which doesn't include the metadata of the record batches offsets. This is needed here to efficiently split a dataset made of one single file.",
"> > Could you explain why you'd need to change the arrow format?\r\n> \r\n> This way workers have random access to the location of the file where its dataset subset starts. Currently we're using the Arrow streaming format which doesn't include the metadata of the record batches offsets. This is needed here to efficiently split a dataset made of one single file.\r\n\r\nI guess I don't understand why you'd need to subset the dataset in the first place. \r\nIt seems sufficient to figure out how to offset or skip rows.\r\n\r\nFor instance, using pyArrow, you could use RecordBatchStreamReader to zero-copy iterate over records with read_next_batch and then only initiate the next step for records modulo worker shard.\r\nThat's one way to do it, where of course you'd need to account for gpu sharding as well.\r\n\r\n\r\nOtherwise, how did you implement worker/node/GPU sharding for iterable/streaming data where you do not have index information or prior splits (e.g. files)?",
"> For instance, using pyArrow, you could use RecordBatchStreamReader to zero-copy iterate over records with read_next_batch and then only initiate the next step for records modulo worker shard.\r\n\r\nThat works indeed ! And what we meant is that you can make it even faster to instantiate. Indeed using RecordBatchStreamReader you need to get the list of all the record batches in each worker, whereas you could just get the list of record batches per worker if you use the record batches locations in the Arrow IPC file footer. This would be especially appreciated to have a fast instantiation in case you have tens of thousands of Arrow files for example.",
"Any recent updates on this ? ",
"I would also appreciate this feature"
] | ### Feature request
Minimal Example
```
import torch
from datasets import IterableDataset
d = IterableDataset.from_file(<file_name>)
dl = torch.utils.data.dataloader.DataLoader(d,num_workers=3)
for sample in dl:
print(sample)
```
Warning:
Too many dataloader workers: 2 (max is dataset.n_shards=1). Stopping 1 dataloader workers.
To parallelize data loading, we give each process some shards (or data sources) to process. Therefore it's unnecessary to have a number of workers greater than dataset.n_shards=1. To enable more parallelism, please split the dataset in more files than 1.
Expected Behavior:
Dataset is sharded each cpu uses subset (contiguously - so you can do checkpoint loading/saving)
### Motivation
I have a lot of unused cpu's and would like to be able to shard iterable datasets with pytorch's dataloader when num_workers > 1. This is for a very large single file. I am aware that we can use the `split_dataset_by_node` to ensure that each node (for distributed) gets different shards, but we should extend it so that this also continues for multiple workers.
### Your contribution
If someone points me to what needs to change, I can create a PR. | 5,984 |
https://github.com/huggingface/datasets/issues/5982 | 404 on Datasets Documentation Page | [
"This wasn’t working for me a bit earlier, but it looks to be back up now",
"We had a minor issue updating the docs after the latest release. It should work now :)."
] | ### Describe the bug
Getting a 404 from the Hugging Face Datasets docs page:
https://huggingface.co/docs/datasets/index
### Steps to reproduce the bug
1. Go to URL https://huggingface.co/docs/datasets/index
2. Notice 404 not found
### Expected behavior
URL should either show docs or redirect to new location
### Environment info
hugginface.co | 5,982 |
https://github.com/huggingface/datasets/issues/5981 | Only two cores are getting used in sagemaker with pytorch 3.10 kernel | [
"I think it's more likely that this issue is related to PyTorch than Datasets, as PyTorch (on import) registers functions to execute when forking a process. Maybe this is the culprit: https://github.com/pytorch/pytorch/issues/99625",
"From reading that ticket, it may be down in mkl? Is it worth hotfixing in the meantime, with the express intention of turning it off? I know that's a horribly crufty solution, but it's also deeply frustrating to be limited to 2 cores for operations as simple as filtration.",
"This is too specific and unrelated to `datasets`, so this shouldn't be fixed here.",
"@mariosasko @mmr-crexi I had the exact same problem on my kubernetes cluster. the datasets subprocess only user 1 and 17 core"
] | ### Describe the bug
When using the newer pytorch 3.10 kernel, only 2 cores are being used by huggingface filter and map functions. The Pytorch 3.9 kernel would use as many cores as specified in the num_proc field.
We have solved this in our own code by placing the following snippet in the code that is called inside subprocesses:
```os.sched_setaffinity(0, {i for i in range(1000)})```
The problem, as near as we can tell, us that once upon a time, cpu affinity was set using a bitmask ("0xfffff" and the like), and affinity recently changed to a list of processors rather than to using the mask. As such, only processors 1 and 17 are shown to be working in htop.
When running functions via `map`, the above resetting of affinity works to spread across the cores. When using `filter`, however, only two cores are active.
### Steps to reproduce the bug
Repro steps:
1. Create an aws sagemaker instance
2. use the pytorch 3_10 kernel
3. Load a dataset
4. run a filter operation
5. watch as only 2 cores are used when num_proc > 2
6. run a map operation
7. watch as only 2 cores are used when num_proc > 2
8. run a map operation with processor affinity reset inside the function called via map
9. Watch as all cores run
### Expected behavior
All specified cores are used via the num_proc argument.
### Environment info
AWS sagemaker with the following init script run in the terminal after instance creation:
conda init bash
bash
conda activate pytorch_p310
pip install Wand PyPDF pytesseract datasets seqeval pdfplumber transformers pymupdf sentencepiece timm donut-python accelerate optimum xgboost
python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
sudo yum -y install htop
sudo yum -y update
sudo yum -y install wget libstdc++ autoconf automake libtool autoconf-archive pkg-config gcc gcc-c++ make libjpeg-devel libpng-devel libtiff-devel zlib-devel | 5,981 |
https://github.com/huggingface/datasets/issues/5980 | Viewing dataset card returns “502 Bad Gateway” | [
"Can you try again? Maybe there was a minor outage.",
"Yes, it seems to be working now. In case it's helpful, the outage lasted several days. It was failing as late as yesterday morning. ",
"we fixed something on the server side, glad it's fixed now"
] | The url is: https://huggingface.co/datasets/Confirm-Labs/pile_ngrams_trigrams
I am able to successfully view the “Files and versions” tab: [Confirm-Labs/pile_ngrams_trigrams at main](https://huggingface.co/datasets/Confirm-Labs/pile_ngrams_trigrams/tree/main)
Any help would be appreciated! Thanks! I hope this is the right place to report an issue like this.
| 5,980 |
https://github.com/huggingface/datasets/issues/5975 | Streaming Dataset behind Proxy - FileNotFoundError | [
"Duplicate of #",
"Hi ! can you try to set the upper case environment variables `HTTP_PROXY` and `HTTPS_PROXY` ?\r\n\r\nWe use `aiohttp` for streaming and it uses case sensitive environment variables",
"Hi, thanks for the quick reply.\r\n\r\nI set the uppercase env variables with\r\n\r\n`\r\nos.environ['HTTP_PROXY'] = \"http://example.com:xxxx\" \r\nos.environ['HTTPS_PROXY'] = \"http://example.com:xxxx\" \r\n`\r\n\r\nHowever, I still get the same error.\r\n\r\nOne thing that could be helpfull: When downloading a dataset without streaming i get the following message:\r\n_HF google storage unreachable. Downloading and preparing it from source_.\r\nThe download does however work as expected.\r\n",
"Are you able to use `aiohttp` to get the file at `https://huggingface.co/datasets/facebook/voxpopuli/resolve/main/data/n_files.json` using your proxy ?",
"It only works when passing trust_env=True when creating the ClientSession, as well as setting ssl=False.\r\n\r\nWorking Example:\r\n\r\n```\r\nimport os\r\n\r\nos.environ['HTTP_PROXY'] = \"xyz\"\r\nos.environ['HTTPS_PROXY'] = \"xyz\"\r\n\r\nimport asyncio\r\nimport aiohttp\r\n\r\nasync def download_pep(url):\r\n async with aiohttp.ClientSession(trust_env=True) as session:\r\n print(\"1\")\r\n async with session.get(url, ssl=False) as resp:\r\n print(\"2\")\r\n content = await resp.text()\r\n print(content)\r\n return content\r\n\r\nasyncio.run(download_pep(\"https://huggingface.co/datasets/facebook/voxpopuli/resolve/main/data/n_files.json\"))\r\n```\r\n\r\n\r\n\r\nSSL Verification has been a problem with other packages as well. Usually I circumvent the problem by setting\r\n```\r\nimport ssl\r\nssl._create_default_https_context = ssl._create_unverified_context\r\n```\r\n(probably not the best idea for security), although here aiohttp does not seem to use this default context.",
"We do pass `trust_env` as well. Could you share the full stack trace you get when streaming using `datasets` ? That could help locate where we might have forgotten to pass `trust_env`",
"Is there a way to disable ssl verification when streaming a dataset. I suspect this might be the isssue with my proxy.\r\n\r\n\r\nHere you go:\r\n\r\n```\r\nFileNotFoundError Traceback (most recent call last)\r\nCell In[8], line 3\r\n 1 from datasets import load_dataset\r\n----> 3 ds = load_dataset(\"facebook/voxpopuli\", name=\"de\", streaming=True)\r\n 5 sample = next(iter(ds))\r\n\r\nFile [~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/load.py:1790](https://vscode-remote+ssh-002dremote-002bml-002er-002dsoftware-002eat.vscode-resource.vscode-cdn.net/home/wrsbri/projects/audio_course/~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/load.py:1790), in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, verification_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, num_proc, storage_options, **config_kwargs)\r\n 1788 # Return iterable dataset in case of streaming\r\n 1789 if streaming:\r\n-> 1790 return builder_instance.as_streaming_dataset(split=split)\r\n 1792 # Some datasets are already processed on the HF google storage\r\n 1793 # Don't try downloading from Google storage for the packaged datasets as text, json, csv or pandas\r\n 1794 try_from_hf_gcs = path not in _PACKAGED_DATASETS_MODULES\r\n\r\nFile [~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/builder.py:1281](https://vscode-remote+ssh-002dremote-002bml-002er-002dsoftware-002eat.vscode-resource.vscode-cdn.net/home/wrsbri/projects/audio_course/~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/builder.py:1281), in DatasetBuilder.as_streaming_dataset(self, split, base_path)\r\n 1274 dl_manager = StreamingDownloadManager(\r\n 1275 base_path=base_path or self.base_path,\r\n 1276 download_config=DownloadConfig(use_auth_token=self.use_auth_token, storage_options=self.storage_options),\r\n 1277 dataset_name=self.name,\r\n 1278 data_dir=self.config.data_dir,\r\n 1279 )\r\n 1280 self._check_manual_download(dl_manager)\r\n-> 1281 splits_generators = {sg.name: sg for sg in self._split_generators(dl_manager)}\r\n 1282 # By default, return all splits\r\n 1283 if split is None:\r\n\r\nFile [~/.cache/huggingface/modules/datasets_modules/datasets/facebook--voxpopuli/b5ff837284f0778eefe0f642734e142d8c3f574eba8c9c8a4b13602297f73604/voxpopuli.py:120](https://vscode-remote+ssh-002dremote-002bml-002er-002dsoftware-002eat.vscode-resource.vscode-cdn.net/home/wrsbri/projects/audio_course/~/.cache/huggingface/modules/datasets_modules/datasets/facebook--voxpopuli/b5ff837284f0778eefe0f642734e142d8c3f574eba8c9c8a4b13602297f73604/voxpopuli.py:120), in Voxpopuli._split_generators(self, dl_manager)\r\n 118 def _split_generators(self, dl_manager):\r\n 119 n_shards_path = dl_manager.download_and_extract(_N_SHARDS_FILE)\r\n--> 120 with open(n_shards_path) as f:\r\n 121 n_shards = json.load(f)\r\n 123 if self.config.name == \"en_accented\":\r\n\r\nFile [~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/streaming.py:71](https://vscode-remote+ssh-002dremote-002bml-002er-002dsoftware-002eat.vscode-resource.vscode-cdn.net/home/wrsbri/projects/audio_course/~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/streaming.py:71), in extend_module_for_streaming..wrap_auth..wrapper(*args, **kwargs)\r\n 69 @wraps(function)\r\n 70 def wrapper(*args, **kwargs):\r\n---> 71 return function(*args, use_auth_token=use_auth_token, **kwargs)\r\n\r\nFile [~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/download/streaming_download_manager.py:517](https://vscode-remote+ssh-002dremote-002bml-002er-002dsoftware-002eat.vscode-resource.vscode-cdn.net/home/wrsbri/projects/audio_course/~/.conda/envs/audio_hf/lib/python3.10/site-packages/datasets/download/streaming_download_manager.py:517), in xopen(file, mode, use_auth_token, *args, **kwargs)\r\n 515 except FileNotFoundError:\r\n 516 if file.startswith(config.HF_ENDPOINT):\r\n--> 517 raise FileNotFoundError(\r\n 518 file + \"\\nIf the repo is private or gated, make sure to log in with `huggingface-cli login`.\"\r\n 519 ) from None\r\n 520 else:\r\n 521 raise\r\n\r\nFileNotFoundError: https://huggingface.co/datasets/facebook/voxpopuli/resolve/main/data/n_files.json\r\nIf the repo is private or gated, make sure to log in with `huggingface-cli login`.\r\n```",
"> Is there a way to disable ssl verification when streaming a dataset.\r\n\r\nI don't think so.\r\n\r\nWe use `fsspec` HTTPFileSystem implementation that is based on `aiohttp`. If you register a subclass of HTTPFileSystem that has SSL disabled by default it could work, but I wouldn't recommended it because it can raise security issues.",
"Okay thanks for your help! I guess I have to figure out how to improve the proxy environment / see if I can make it work with ssl connections."
] | ### Describe the bug
When trying to stream a dataset i get the following error after a few minutes of waiting.
```
FileNotFoundError: https://huggingface.co/datasets/facebook/voxpopuli/resolve/main/data/n_files.json
If the repo is private or gated, make sure to log in with `huggingface-cli login`.
```
I have already set the proxy environment variables. Downloading a Dataset without streaming works as expected.
Still i suspect that this is connected to being behind a proxy.
Is there a way to set the proxy for streaming datasets? Possibly a keyword argument that gets passed to ffspec?
### Steps to reproduce the bug
This is the code i use.
```
import os
os.environ['http_proxy'] = "http://example.com:xxxx"
os.environ['https_proxy'] = "http://example.com:xxxx"
from datasets import load_dataset
ds = load_dataset("facebook/voxpopuli", name="de", streaming=True)
```
### Expected behavior
I would expect the streaming functionality to use the set proxy settings.
### Environment info
- `datasets` version: 2.13.0
- Platform: Linux-5.15.0-73-generic-x86_64-with-glibc2.35
- Python version: 3.10.11
- Huggingface_hub version: 0.15.1
- PyArrow version: 11.0.0
- Pandas version: 2.0.2
| 5,975 |
https://github.com/huggingface/datasets/issues/5971 | Docs: make "repository structure" easier to find | [
"Loading a local dataset also works the same way when `data_files` are not specified, so I agree we should make this info easier to discover \r\n\r\ncc @stevhliu ",
"Is this issue open? If so, I will self assign. ",
"@benjaminbrown038 Yes, it is. Maybe @stevhliu can give some pointers on improving this doc page's discoverability.",
"I think we can add a version of the [Main use-case](https://huggingface.co/docs/datasets/repository_structure#main-usecase) section to the [Share a dataset to the Hub](https://huggingface.co/docs/datasets/upload_dataset) tutorial. \r\n\r\nCurrently, it doesn't tell you *how* to structure the repository; it only tells you how to create it. So adding the \"main use-case\" will help bridge the gap and make it easier to find. We should also add a link to the [Structure your repository](https://huggingface.co/docs/datasets/repository_structure) guide for users who want to learn about the other options.",
"#self-assign"
] | The page https://huggingface.co/docs/datasets/repository_structure explains how to create a simple repository structure without a dataset script.
It's the simplest way to create a dataset and should be easier to find, particularly on the docs' first pages. | 5,971 |
https://github.com/huggingface/datasets/issues/5970 | description disappearing from Info when Uploading a Dataset Created with `from_dict` | [
"Here's a minimal way to reproduce the bug, for the sake of convenience.\r\n````\r\nfrom datasets import Dataset, DatasetInfo, load_dataset\r\n\r\n\r\nepisodes_dict = {\"test\":[1,2,3],\"test2\": [1,2,4]}\r\n\r\nhugging_face_dataset = Dataset.from_dict(\r\n episodes_dict, info=DatasetInfo(description=\"test_str\")\r\n)\r\nprint(hugging_face_dataset.info)\r\n\r\nhugging_face_dataset.push_to_hub(\"balisujohn/minari_test\", private=True)\r\n\r\nredownloaded_dataset= load_dataset(\"balisujohn/minari_test\")[\"train\"]\r\n\r\n\r\nprint(redownloaded_dataset.info)\r\n````\r\n",
"Thanks for reporting !\r\n\r\nFor now I would recommend uploading a separate JSON file for your metadata.\r\n\r\nAlternatively you can upload a second configuration of the dataset containing your metadata but this feature is not released yet (though you can already use it from [here](https://github.com/huggingface/datasets/pull/5331), it will be released soon)"
] | ### Describe the bug
When uploading a dataset created locally using `from_dict` with a specified `description` field. It appears before upload, but is missing after upload and re-download.
### Steps to reproduce the bug
I think the most relevant pattern in the code might be the following lines:
```
description_json_str = json.dumps(
{
"dataset_id": dataset.spec.dataset_id,
"env_name": dataset.spec.env_spec.id,
"action_space": serialize_space(dataset.spec.action_space),
"observation_space": serialize_space(dataset.spec.observation_space),
}
)
hugging_face_dataset = Dataset.from_dict(
episodes_dict, info=DatasetInfo(description=description_json_str)
)
```
Which comes from this function https://github.com/balisujohn/minarai/blob/8e023727f0a8488c4451651d9f7a79b981412c40/minari/integrations/hugging_face.py#L39
To replicate,
clone this branch of my Minari fork https://github.com/balisujohn/minarai/tree/dev-huggingface then run
```
python3.8 -m venv env
source env/bin/activate
python3 -m pip install -e .
python3 -m pip install pytest
```
The change the hugging face repo path in the test called `test_hugging_face_push_and_pull_dataset` in `tests/integrations/test_hugging_face.py` to one you have permissions to write to.
Then run:
```
pytest tests/integrations/test_hugging_face.py::test_hugging_face_push_and_pull_dataset
```
### Expected behavior
DATASET INFO BEFORE UPLOADING
DatasetInfo(description='{"dataset_id": "dummy-combo-test-v0", "env_name": "DummyComboEnv-v0", "action_space": "{\\"type\\": \\"Tuple\\", \\"subspaces\\": [{\\"type\\": \\"Box\\", \\"dtype\\": \\"float32\\", \\"shape\\": [1], \\"low\\": [2.0], \\"high\\": [3.0]}, {\\"type\\": \\"Box\\", \\"dtype\\": \\"float32\\", \\"shape\\": [1], \\"low\\": [4.0], \\"high\\": [5.0]}]}", "observation_space": "{\\"type\\": \\"Tuple\\", \\"subspaces\\": [{\\"type\\": \\"Box\\", \\"dtype\\": \\"float32\\", \\"shape\\": [1], \\"low\\": [2.0], \\"high\\": [3.0]}, {\\"type\\": \\"Tuple\\", \\"subspaces\\": [{\\"type\\": \\"Box\\", \\"dtype\\": \\"float32\\", \\"shape\\": [1], \\"low\\": [2.0], \\"high\\": [3.0]}, {\\"type\\": \\"Dict\\", \\"subspaces\\": {\\"component_1\\": {\\"type\\": \\"Box\\", \\"dtype\\": \\"float32\\", \\"shape\\": [1], \\"low\\": [-1.0], \\"high\\": [1.0]}, \\"component_2\\": {\\"type\\": \\"Dict\\", \\"subspaces\\": {\\"subcomponent_1\\": {\\"type\\": \\"Box\\", \\"dtype\\": \\"float32\\", \\"shape\\": [1], \\"low\\": [2.0], \\"high\\": [3.0]}, \\"subcomponent_2\\": {\\"type\\": \\"Tuple\\", \\"subspaces\\": [{\\"type\\": \\"Box\\", \\"dtype\\": \\"float32\\", \\"shape\\": [1], \\"low\\": [4.0], \\"high\\": [5.0]}, {\\"type\\": \\"Discrete\\", \\"dtype\\": \\"int64\\", \\"start\\": 0, \\"n\\": 10}]}}}}}]}]}"}', citation='', homepage='', license='', features={'observations': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': {'component_1': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), 'component_2': {'subcomponent_1': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), 'subcomponent_2': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': Value(dtype='int64', id=None)}}}}}, 'actions': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None)}, 'rewards': Value(dtype='int64', id=None), 'truncations': Value(dtype='bool', id=None), 'terminations': Value(dtype='bool', id=None), 'episode_ids': Value(dtype='int64', id=None)}, post_processed=None, supervised_keys=None, task_templates=None, builder_name=None, config_name=None, version=None, splits=None, download_checksums=None, download_size=None, post_processing_size=None, dataset_size=None, size_in_bytes=None)
...
DATASET INFO AFTER UPLOADING AND DOWNLOADING
DatasetInfo(description='', citation='', homepage='', license='', features={'observations': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': {'component_1': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), 'component_2': {'subcomponent_1': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), 'subcomponent_2': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': Value(dtype='int64', id=None)}}}}}, 'actions': {'_index_0': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None), '_index_1': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None)}, 'rewards': Value(dtype='int64', id=None), 'truncations': Value(dtype='bool', id=None), 'terminations': Value(dtype='bool', id=None), 'episode_ids': Value(dtype='int64', id=None)}, post_processed=None, supervised_keys=None, task_templates=None, builder_name=None, config_name=None, version=None, splits={'train': SplitInfo(name='train', num_bytes=4846, num_examples=60, shard_lengths=None, dataset_name='parquet')}, download_checksums={'https://huggingface.co/datasets/balisujohn/minari_test/resolve/8217b614ff9ba5edc1a30c7df430e92a46f65363/data/train-00000-of-00001-7c5900b93b35745e.parquet': {'num_bytes': 9052, 'checksum': None}}, download_size=9052, post_processing_size=None, dataset_size=4846, size_in_bytes=13898)
...
### Environment info
- `datasets` version: 2.13.0
- Platform: Linux-5.15.0-75-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- Huggingface_hub version: 0.15.1
- PyArrow version: 12.0.1
- Pandas version: 2.0.2
| 5,970 |
https://github.com/huggingface/datasets/issues/5968 | Common Voice datasets still need `use_auth_token=True` | [
"cc @pcuenca as well. \r\n\r\nNot super urgent btw",
"The issue commes from the dataset itself and is not related to the `datasets` lib\r\n\r\nsee https://huggingface.co/datasets/mozilla-foundation/common_voice_6_1/blob/2c475b3b88e0f2e5828f830a4b91618a25ff20b7/common_voice_6_1.py#L148-L152",
"Let's remove these lines in the dataset no? cc @anton-l @Vaibhavs10 ",
"Addressed in:\r\n\r\n* `mozilla-foundation/common_voice_1_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_1_0/discussions/4)\r\n* `mozilla-foundation/common_voice_2_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_2_0/discussions/3)\r\n* `mozilla-foundation/common_voice_3_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_3_0/discussions/3)\r\n* `mozilla-foundation/common_voice_4_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_4_0/discussions/3)\r\n* `mozilla-foundation/common_voice_5_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_5_0/discussions/3)\r\n* `mozilla-foundation/common_voice_5_1` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_5_1/discussions/3)\r\n* `mozilla-foundation/common_voice_6_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_6_0/discussions/3)\r\n* `mozilla-foundation/common_voice_6_1` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_6_1/discussions/3)\r\n* `mozilla-foundation/common_voice_7_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0/discussions/3)\r\n* `mozilla-foundation/common_voice_8_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_8_0/discussions/7)\r\n* `mozilla-foundation/common_voice_9_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_9_0/discussions/8)\r\n* `mozilla-foundation/common_voice_10_0` [PR](https://huggingface.co/datasets/mozilla-foundation/common_voice_10_0/discussions/7)"
] | ### Describe the bug
We don't need to pass `use_auth_token=True` anymore to download gated datasets or models, so the following should work if correctly logged in.
```py
from datasets import load_dataset
load_dataset("mozilla-foundation/common_voice_6_1", "tr", split="train+validation")
```
However it throws an error - probably because something weird is hardcoded into the dataset loading script.
### Steps to reproduce the bug
1.)
```
huggingface-cli login
```
2.) Make sure that you have accepted the license here:
https://huggingface.co/datasets/mozilla-foundation/common_voice_6_1
3.) Run:
```py
from datasets import load_dataset
load_dataset("mozilla-foundation/common_voice_6_1", "tr", split="train+validation")
```
4.) You'll get:
```
File ~/hf/lib/python3.10/site-packages/datasets/builder.py:963, in DatasetBuilder._download_and_prepare(self, dl_manager, verification_mode, **prepare_split_kwargs)
961 split_dict = SplitDict(dataset_name=self.name)
962 split_generators_kwargs = self._make_split_generators_kwargs(prepare_split_kwargs)
--> 963 split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
965 # Checksums verification
966 if verification_mode == VerificationMode.ALL_CHECKS and dl_manager.record_checksums:
File ~/.cache/huggingface/modules/datasets_modules/datasets/mozilla-foundation--common_voice_6_1/f4d7854c466f5bd4908988dbd39044ec4fc634d89e0515ab0c51715c0127ffe3/common_voice_6_1.py:150, in CommonVoice._split_generators(self, dl_manager)
148 hf_auth_token = dl_manager.download_config.use_auth_token
149 if hf_auth_token is None:
--> 150 raise ConnectionError(
151 "Please set use_auth_token=True or use_auth_token='<TOKEN>' to download this dataset"
152 )
154 bundle_url_template = STATS["bundleURLTemplate"]
155 bundle_version = bundle_url_template.split("/")[0]
ConnectionError: Please set use_auth_token=True or use_auth_token='<TOKEN>' to download this dataset
```
### Expected behavior
One should not have to pass `use_auth_token=True`. Also see discussion here: https://github.com/huggingface/blog/pull/1243#discussion_r1235131150
### Environment info
```
- `datasets` version: 2.13.0
- Platform: Linux-6.2.0-76060200-generic-x86_64-with-glibc2.35
- Python version: 3.10.6
- Huggingface_hub version: 0.16.0.dev0
- PyArrow version: 11.0.0
- Pandas version: 1.5.3
``` | 5,968 |
https://github.com/huggingface/datasets/issues/5967 | Config name / split name lost after map with multiproc | [
"This must be due to DatasetInfo.from_merge which drops them and is used in `concatenate_datasets`.\r\n\r\nAnd you're experiencing this issue because multiprocessing does concatenate the resulting datasets from each process.\r\n\r\nMaybe they should be kept if all the subdatasets share the same values for config_name and split",
"That sounds like a clean workaround!"
] | ### Describe the bug
Performing a `.map` method on a dataset loses it's config name / split name only if run with multiproc
### Steps to reproduce the bug
```python
from datasets import Audio, load_dataset
from transformers import AutoFeatureExtractor
import numpy as np
# load dummy dataset
libri = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean")
# make train / test splits
libri = libri["validation"].train_test_split(seed=42, shuffle=True, test_size=0.1)
# example feature extractor
model_id = "ntu-spml/distilhubert"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id, do_normalize=True, return_attention_mask=True)
sampling_rate = feature_extractor.sampling_rate
libri = libri.cast_column("audio", Audio(sampling_rate=sampling_rate))
max_duration = 30.0
def preprocess_function(examples):
audio_arrays = [x["array"] for x in examples["audio"]]
inputs = feature_extractor(
audio_arrays,
sampling_rate=feature_extractor.sampling_rate,
max_length=int(feature_extractor.sampling_rate * max_duration),
truncation=True,
return_attention_mask=True,
)
return inputs
# single proc map
libri_encoded = libri.map(
preprocess_function, remove_columns=["audio", "file"], batched=True, num_proc=1
)
print(10 * "=" ,"Single processing", 10 * "=")
print("Config name before: ", libri["train"].config_name, " Split name before: ", libri["train"].split)
print("Config name after: ", libri_encoded["train"].config_name, " Split name after: ", libri_encoded["train"].split)
# multi proc map
libri_encoded = libri.map(
preprocess_function, remove_columns=["audio", "file"], batched=True, num_proc=2
)
print(10 * "=" ,"Multi processing", 10 * "=")
print("Config name before: ", libri["train"].config_name, " Split name before: ", libri["train"].split)
print("Config name after: ", libri_encoded["train"].config_name, " Split name after: ", libri_encoded["train"].split)
```
**Print Output:**
```
========== Single processing ==========
Config name before: clean Split name before: validation
Config name after: clean Split name after: validation
========== Multi processing ==========
Config name before: clean Split name before: validation
Config name after: None Split name after: None
```
=> we can see that the config/split names are lost in the multiprocessing setting
### Expected behavior
Should retain both config / split names in the multiproc setting
### Environment info
- `datasets` version: 2.13.1.dev0
- Platform: Linux-5.15.0-67-generic-x86_64-with-glibc2.35
- Python version: 3.10.6
- Huggingface_hub version: 0.15.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.2 | 5,967 |
https://github.com/huggingface/datasets/issues/5965 | "Couldn't cast array of type" in complex datasets | [
"Thanks for reporting! \r\n\r\nSpecifying the target features explicitly should avoid this error:\r\n```python\r\ndataset = dataset.map(\r\n batch_process,\r\n batched=True,\r\n batch_size=1,\r\n num_proc=1,\r\n remove_columns=dataset.column_names,\r\n features=datasets.Features({\"texts\": datasets.Sequence(datasets.Value(\"string\"))})\r\n)\r\n```\r\n\r\nThis error stems from our type promotion not handling the nested case. But this promotion/casting allocates memory in most scenarios, which can be problematic for large datasets, so explicitly passing the features is the optimal solution.",
"Hi @mariosasko thanks for the context, this is helpful to know. Would it be worth having some logic to generate this explicit feature specification automatically if a type annotation for a .map returns a dataclass that can be inferred?\r\n\r\nFeels like something that would be easy to implement and could save memory / deal with this case in a standardized way.",
"> . Would it be worth having some logic to generate this explicit feature specification automatically if a type annotation for a .map returns a dataclass that can be inferred?\r\n\r\nInteresting proposal! Yes, we could consider doing this if the (return) type hint is `TypedDict`, and raise an error that type hints are incorrect if the cast using the inferred types fails.",
"@mariosasko Put up an initial PR to implement this proposal. Let me know your thoughts on direction and what else should be in-scope here."
] | ### Describe the bug
When doing a map of a dataset with complex types, sometimes `datasets` is unable to interpret the valid schema of a returned datasets.map() function. This often comes from conflicting types, like when both empty lists and filled lists are competing for the same field value.
This is prone to happen in batch mapping, when the mapper returns a sequence of null/empty values and other batches are non-null. A workaround is to manually cast the new batch to a pyarrow table (like implemented in this [workaround](https://github.com/piercefreeman/lassen/pull/3)) but it feels like this ideally should be solved at the core library level.
Note that the reproduction case only throws this error if the first datapoint has the empty list. If it is processed later, datasets already detects its representation as list-type and therefore allows the empty list to be provided.
### Steps to reproduce the bug
A trivial reproduction case:
```python
from typing import Iterator, Any
import pandas as pd
from datasets import Dataset
def batch_to_examples(batch: dict[str, list[Any]]) -> Iterator[dict[str, Any]]:
for i in range(next(iter(lengths))):
yield {feature: values[i] for feature, values in batch.items()}
def examples_to_batch(examples) -> dict[str, list[Any]]:
batch = {}
for example in examples:
for feature, value in example.items():
if feature not in batch:
batch[feature] = []
batch[feature].append(value)
return batch
def batch_process(examples, explicit_schema: bool):
new_examples = []
for example in batch_to_examples(examples):
new_examples.append(dict(texts=example["raw_text"].split()))
return examples_to_batch(new_examples)
df = pd.DataFrame(
[
{"raw_text": ""},
{"raw_text": "This is a test"},
{"raw_text": "This is another test"},
]
)
dataset = Dataset.from_pandas(df)
# datasets won't be able to typehint a dataset that starts with an empty example.
with pytest.raises(TypeError, match="Couldn't cast array of type"):
dataset = dataset.map(
batch_process,
batched=True,
batch_size=1,
num_proc=1,
remove_columns=dataset.column_names,
)
```
This results in crashes like:
```bash
File "/Users/piercefreeman/Library/Caches/pypoetry/virtualenvs/example-9kBqeSPy-py3.11/lib/python3.11/site-packages/datasets/table.py", line 1819, in wrapper
return func(array, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/piercefreeman/Library/Caches/pypoetry/virtualenvs/example-9kBqeSPy-py3.11/lib/python3.11/site-packages/datasets/table.py", line 2109, in cast_array_to_feature
return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/piercefreeman/Library/Caches/pypoetry/virtualenvs/example-9kBqeSPy-py3.11/lib/python3.11/site-packages/datasets/table.py", line 1819, in wrapper
return func(array, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/piercefreeman/Library/Caches/pypoetry/virtualenvs/example-9kBqeSPy-py3.11/lib/python3.11/site-packages/datasets/table.py", line 1998, in array_cast
raise TypeError(f"Couldn't cast array of type {array.type} to {pa_type}")
TypeError: Couldn't cast array of type string to null
```
### Expected behavior
The code should successfully map and create a new dataset without error.
### Environment info
Mac OSX, Linux | 5,965 |
https://github.com/huggingface/datasets/issues/5963 | Got an error _pickle.PicklingError use Dataset.from_spark. | [
"i got error using method from_spark when using multi-node Spark cluster. seems could only use \"from_spark\" in local?",
"@lhoestq ",
"cc @maddiedawson it looks like there an issue with `_validate_cache_dir` ?\r\n\r\nIt looks like the function passed to mapPartitions has a reference to the Spark dataset builder, and therefore contains the SparkContext itself.\r\n\r\nI think it can be fixed by defining `create_cache_and_write_probe` outside the Spark dataset builder, and pass a `partial(create_cache_and_write_probe, cache_dir=self._cache_dir)` to `mapPartitions`",
"Just saw this; thanks for flagging! Your proposed solution sounds good. I can prepare a PR",
"@maddiedawson can you show me the demo ,so i can test in local .before your PR"
] | python 3.9.2
Got an error _pickle.PicklingError use Dataset.from_spark.
Did the dataset import load data from spark dataframe using multi-node Spark cluster
df = spark.read.parquet(args.input_data).repartition(50)
ds = Dataset.from_spark(df, keep_in_memory=True,
cache_dir="/pnc-data/data/nuplan/t5_spark/cache_data")
ds.save_to_disk(args.output_data)
Error :
_pickle.PicklingError: Could not serialize object: RuntimeError: It appears that you are attempting to reference SparkContext from a broadcast variable, action, or transforma
tion. SparkContext can only be used on the driver, not in code that it run on workers. For more information, see SPARK-5063.
23/06/16 21:17:20 WARN ExecutorPodsWatchSnapshotSource: Kubernetes client has been closed (this is expected if the application is shutting down.)
_Originally posted by @yanzia12138 in https://github.com/huggingface/datasets/issues/5701#issuecomment-1594674306_
W
Traceback (most recent call last):
File "/home/work/main.py", line 100, in <module>
run(args)
File "/home/work/main.py", line 80, in run
ds = Dataset.from_spark(df1, keep_in_memory=True,
File "/home/work/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 1281, in from_spark
return SparkDatasetReader(
File "/home/work/.local/lib/python3.9/site-packages/datasets/io/spark.py", line 53, in read
self.builder.download_and_prepare(
File "/home/work/.local/lib/python3.9/site-packages/datasets/builder.py", line 909, in download_and_prepare
self._download_and_prepare(
File "/home/work/.local/lib/python3.9/site-packages/datasets/builder.py", line 1004, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/work/.local/lib/python3.9/site-packages/datasets/packaged_modules/spark/spark.py", line 254, in _prepare_split
self._validate_cache_dir()
File "/home/work/.local/lib/python3.9/site-packages/datasets/packaged_modules/spark/spark.py", line 122, in _validate_cache_dir
self._spark.sparkContext.parallelize(range(1), 1).mapPartitions(create_cache_and_write_probe).collect()
File "/home/work/.local/lib/python3.9/site-packages/pyspark/rdd.py", line 950, in collect
sock_info = self.ctx._jvm.PythonRDD.collectAndServe(self._jrdd.rdd())
File "/home/work/.local/lib/python3.9/site-packages/pyspark/rdd.py", line 2951, in _jrdd
wrapped_func = _wrap_function(self.ctx, self.func, self._prev_jrdd_deserializer,
File "/home/work/.local/lib/python3.9/site-packages/pyspark/rdd.py", line 2830, in _wrap_function
pickled_command, broadcast_vars, env, includes = _prepare_for_python_RDD(sc, command)
File "/home/work/.local/lib/python3.9/site-packages/pyspark/rdd.py", line 2816, in _prepare_for_python_RDD
pickled_command = ser.dumps(command)
File "/home/work/.local/lib/python3.9/site-packages/pyspark/serializers.py", line 447, in dumps
raise pickle.PicklingError(msg)
_pickle.PicklingError: Could not serialize object: RuntimeError: It appears that you are attempting to reference SparkContext from a broadcast variable, action, or transformation. S
parkContext can only be used on the driver, not in code that it run on workers. For more information, see SPARK-5063.
23/06/19 13:51:21 WARN ExecutorPodsWatchSnapshotSource: Kubernetes client has been closed (this is expected if the application is shutting down.)
| 5,963 |
https://github.com/huggingface/datasets/issues/5962 | Issue with train_test_split maintaining the same underlying PyArrow Table | [] | ### Describe the bug
I've been using the train_test_split method in the datasets module to split my HuggingFace Dataset into separate training, validation, and testing subsets. However, I've noticed an issue where the split datasets appear to maintain the same underlying PyArrow Table.
### Steps to reproduce the bug
1. Load any dataset ```dataset = load_dataset("lhoestq/demo1")```
2. Try the next code:
```python
from datasets import Dataset, DatasetDict
train_size = 0.6
split_train = dataset["train"].train_test_split(
train_size=train_size,
)
separate_dataset_dict = DatasetDict({
"train": split_train["train"],
"test": split_train["test"],
})
```
3. The next code ```print(separate_dataset_dict)``` when printing the dataset it gives the indication that they have 3 and 2 rows respectively.
4. But the next code:
```python
print(len(separate_dataset_dict["train"].data['id']))
print(len(separate_dataset_dict["test"].data['id']))
```
Indicates that both tables still have 5 rows.
### Expected behavior
However, I've noticed that train_test_split["train"].data, test_val_split["train"].data, and test_val_split["test"].data are identical, suggesting that they all point to the same underlying PyArrow Table. This means that the split datasets are not independent, as I expected.
I believe this is a bug in the train_test_split implementation, as I would expect this function to return datasets with separate underlying PyArrow Tables. Could you please help me understand if this is expected behavior, or if there's a workaround to create truly independent split datasets?
I would appreciate any assistance with this issue. Thank you.
### Environment info
I tried in Colab:
- `datasets` version: 2.13.0
- Platform: Windows-10-10.0.22621-SP0
- Python version: 3.10.11
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.1
and my PC:
- `datasets` version: 2.13.0
- Platform: Linux-5.15.107+-x86_64-with-glibc2.31
- Python version: 3.10.12
- Huggingface_hub version: 0.15.1
- PyArrow version: 9.0.0
- Pandas version: 1.5.3 | 5,962 |
https://github.com/huggingface/datasets/issues/5961 | IterableDataset: split by node and map may preprocess samples that will be skipped anyway | [
"Does \"number of shards\" refer to the total number of data?\r\n\r\nmy config:\r\nnproc_per_node=2\r\nds=ds['train'] = load_dataset(streaming=True).take(50000)\r\n\r\nI'm test again: in prepare_data(), data have the same for each GPU\r\n",
"The number of shards is `ds.n_shards`. It corresponds generally to the number of files the dataset is made of, to be able to distribute to several nodes.\r\n\r\n**You don't end up with the same data per GPU**. But all the samples are going through your preprocessing function you pass to map. They are just skipped afterwards to only keep 1 sample out of n(GPUs)",
"For each GPU, although see the same data in prepare_data(), the actual training data will not be the same in the end. \r\nIs my understanding correct?\r\n\r\nWhere can I print the actual training data for each GPU?",
"> For each GPU, although see the same data in prepare_data(), the actual training data will not be the same in the end.\r\nIs my understanding correct?\r\n\r\nYes exactly :)\r\n\r\n> Where can I print the actual training data for each GPU?\r\n\r\nYou should call print in the data_collator",
"I print out n_shards, and under multiple GPUs, this value is always 1.\r\nIs this value correct?",
"Yes it's correct, and it explains why you always have the same data passed to your map function (the data can't be split).\r\n\r\nBut after being passed to `map`, each GPU keeps one example out of n(GPUs) so that you don't end up with duplicate data across GPUs",
"> > For each GPU, although see the same data in prepare_data(), the actual training data will not be the same in the end.\r\n> > Is my understanding correct?\r\n> \r\n> Yes exactly :)\r\n> \r\n> > Where can I print the actual training data for each GPU?\r\n> \r\n> You should call print in the data_collator\r\n\r\nOK, when printing the train data in the data collator, each GPU sees different data.\r\n\r\nThanks for your reply",
"Do we have a solution for this one? Or it's required to get \"number of shards is a factor of number of GPUs: in that case the shards are evenly distributed per GPU\"",
"For now it's required to have a number of shards that is a factor of the number of GPUs to not have all the workers process the same data (and then skip the right ones to not end up training on duplicate data).\r\n\r\nIt would be quite complex to implement a strategy that would utilize all the GPUs with an arbitrary number of shards even at the end of training"
] | There are two ways an iterable dataset can be split by node:
1. if the number of shards is a factor of number of GPUs: in that case the shards are evenly distributed per GPU
2. otherwise, each GPU iterate on the data and at the end keeps 1 sample out of n(GPUs) - skipping the others.
In case 2. it's therefore possible to have the same examples passed to `prepare_dataset` for each GPU.
This doesn't sound optimized though, because it runs the preprocessing on samples that won't be used in the end.
Could you open a new issue so that we can discuss about this and find a solution ?
_Originally posted by @lhoestq in https://github.com/huggingface/datasets/issues/5360#issuecomment-1592729051_
| 5,961 |
https://github.com/huggingface/datasets/issues/5959 | read metric glue.py from local file | [
"Sorry, I solve this by call `evaluate.load('glue_metric.py','sst-2')`\r\n"
] | ### Describe the bug
Currently, The server is off-line. I am using the glue metric from the local file downloaded from the hub.
I download / cached datasets using `load_dataset('glue','sst2', cache_dir='/xxx')` to cache them and then in the off-line mode, I use `load_dataset('xxx/glue.py','sst2', cache_dir='/xxx')`. I can successfully reuse cached datasets.
My problem is about the load_metric.
When I run `load_dataset('xxx/glue_metric.py','sst2',cache_dir='/xxx')` , it returns
` File "xx/lib64/python3.9/site-packages/datasets/utils/deprecation_utils.py", line 46, in wrapper
return deprecated_function(*args, **kwargs)
File "xx//lib64/python3.9/site-packages/datasets/load.py", line 1392, in load_metric
metric = metric_cls(
TypeError: 'NoneType' object is not callable`
Thanks in advance for help!
### Steps to reproduce the bug
N/A
### Expected behavior
N/A
### Environment info
`datasets == 2.12.0` | 5,959 |
https://github.com/huggingface/datasets/issues/5955 | Strange bug in loading local JSON files, using load_dataset | [
"This is the actual error:\r\n```\r\nFailed to read file '/home/lakala/hjc/code/pycode/glm/temp.json' with error <class 'pyarrow.lib.ArrowInvalid'>: cannot mix list and non-list, non-null values\r\n```\r\nWhich means some samples are incorrectly formatted.\r\n\r\nPyArrow, a storage backend that we use under the hood, requires that all the list elements have the same level of nesting (same number of dimensions) or are `None`.\r\n```python\r\nimport pyarrow as pa\r\npa.array([[1, 2, 3], 2]) # ArrowInvalid: cannot mix list and non-list, non-null values\r\npa.array([[1, 2, 3], [2]]) # works\r\n``` ",
"@mariosasko \r\nI used the same operation to check the original data before and after slicing.\r\nThis is reflected in my code.\r\n160000 is not a specific number.\r\nI can also get output using 150000.\r\nThis doesn't seem to align very well with what you said.\r\nBecause if only some sample formats are incorrect.\r\nSo there should be an error in one of the front and back slices.\r\nthank you for your reply.",
"Our JSON loader does the following in your case:\r\n\r\n```python\r\nimport json\r\nimport pyarrow as pa\r\n\r\nwith open(file, encoding=\"utf-8\") as f:\r\n dataset = json.load(f)\r\nkeys = set().union(*[row.keys() for row in dataset])\r\nmapping = {col: [row.get(col) for row in dataset] for col in keys}\r\npa_table = pa.Table.from_pydict(mapping) # the ArrowInvalid error comes from here\r\n```\r\n\r\nSo if this code throws an error with correctly-formatted JSON, then this is an Arrow bug and should be reported in their repo.\r\n\r\n> I used the same operation to check the original data before and after slicing.\r\nThis is reflected in my code.\r\n160000 is not a specific number.\r\nI can also get output using 150000.\r\nThis doesn't seem to align very well with what you said.\r\nBecause if only some sample formats are incorrect.\r\nSo there should be an error in one of the front and back slices.\r\n\r\nYou should shuffle the data to make sure that's not the case",
"@mariosasko \r\nThank you.\r\nI will try again."
] | ### Describe the bug
I am using 'load_dataset 'loads a JSON file, but I found a strange bug: an error will be reported when the length of the JSON file exceeds 160000 (uncertain exact number). I have checked the data through the following code and there are no issues. So I cannot determine the true reason for this error.
The data is a list containing a dictionary. As follows:
[
{'input': 'someting...', 'target': 'someting...', 'type': 'someting...', 'history': ['someting...', ...]},
...
]
### Steps to reproduce the bug
```
import json
from datasets import load_dataset
path = "target.json"
temp_path = "temp.json"
with open(path, "r") as f:
data = json.load(f)
print(f"\n-------the JSON file length is: {len(data)}-------\n")
with open(temp_path, "w") as f:
json.dump(data[:160000], f)
dataset = load_dataset("json", data_files=temp_path)
print("\n-------This works when the JSON file length is 160000-------\n")
with open(temp_path, "w") as f:
json.dump(data[160000:], f)
dataset = load_dataset("json", data_files=temp_path)
print("\n-------This works and eliminates data issues-------\n")
with open(temp_path, "w") as f:
json.dump(data[:170000], f)
dataset = load_dataset("json", data_files=temp_path)
```
### Expected behavior
```
-------the JSON file length is: 173049-------
Downloading and preparing dataset json/default to /root/.cache/huggingface/datasets/json/default-acf3c7f418c5f4b4/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4...
Downloading data files: 100%|███████████████████| 1/1 [00:00<00:00, 3328.81it/s]
Extracting data files: 100%|█████████████████████| 1/1 [00:00<00:00, 639.47it/s]
Dataset json downloaded and prepared to /root/.cache/huggingface/datasets/json/default-acf3c7f418c5f4b4/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4. Subsequent calls will reuse this data.
100%|████████████████████████████████████████████| 1/1 [00:00<00:00, 265.85it/s]
-------This works when the JSON file length is 160000-------
Downloading and preparing dataset json/default to /root/.cache/huggingface/datasets/json/default-a42f04b263ceea6a/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4...
Downloading data files: 100%|███████████████████| 1/1 [00:00<00:00, 2038.05it/s]
Extracting data files: 100%|█████████████████████| 1/1 [00:00<00:00, 794.83it/s]
Dataset json downloaded and prepared to /root/.cache/huggingface/datasets/json/default-a42f04b263ceea6a/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4. Subsequent calls will reuse this data.
100%|████████████████████████████████████████████| 1/1 [00:00<00:00, 681.00it/s]
-------This works and eliminates data issues-------
Downloading and preparing dataset json/default to /root/.cache/huggingface/datasets/json/default-63f391c89599c7b0/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4...
Downloading data files: 100%|███████████████████| 1/1 [00:00<00:00, 3682.44it/s]
Extracting data files: 100%|█████████████████████| 1/1 [00:00<00:00, 788.70it/s]
Generating train split: 0 examples [00:00, ? examples/s]Failed to read file '/home/lakala/hjc/code/pycode/glm/temp.json' with error <class 'pyarrow.lib.ArrowInvalid'>: cannot mix list and non-list, non-null values
Traceback (most recent call last):
File "/home/lakala/conda/envs/glm/lib/python3.8/site-packages/datasets/builder.py", line 1858, in _prepare_split_single
for _, table in generator:
File "/home/lakala/conda/envs/glm/lib/python3.8/site-packages/datasets/packaged_modules/json/json.py", line 146, in _generate_tables
raise ValueError(f"Not able to read records in the JSON file at {file}.") from None
ValueError: Not able to read records in the JSON file at /home/lakala/hjc/code/pycode/glm/temp.json.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/home/lakala/hjc/code/pycode/glm/test.py", line 22, in <module>
dataset = load_dataset("json", data_files=temp_path)
File "/home/lakala/conda/envs/glm/lib/python3.8/site-packages/datasets/load.py", line 1797, in load_dataset
builder_instance.download_and_prepare(
File "/home/lakala/conda/envs/glm/lib/python3.8/site-packages/datasets/builder.py", line 890, in download_and_prepare
self._download_and_prepare(
File "/home/lakala/conda/envs/glm/lib/python3.8/site-packages/datasets/builder.py", line 985, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/lakala/conda/envs/glm/lib/python3.8/site-packages/datasets/builder.py", line 1746, in _prepare_split
for job_id, done, content in self._prepare_split_single(
File "/home/lakala/conda/envs/glm/lib/python3.8/site-packages/datasets/builder.py", line 1891, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.builder.DatasetGenerationError: An error occurred while generating the dataset
```
### Environment info
```
Ubuntu==22.04
python==3.8
pytorch-transformers==1.2.0
transformers== 4.27.1
datasets==2.12.0
numpy==1.24.3
pandas==1.5.3
``` | 5,955 |
https://github.com/huggingface/datasets/issues/5953 | Bad error message when trying to download gated dataset | [
"cc @sanchit-gandhi @Vaibhavs10 @lhoestq - this is mainly for demos that use Common Voice datasets as done here: https://github.com/facebookresearch/fairseq/tree/main/examples/mms#-transformers\r\n",
"Hi ! the error for me is\r\n\r\n```\r\nFileNotFoundError: Couldn't find a dataset script at /content/mozilla-foundation/common_voice_13_0/common_voice_13_0.py or any data file in the same directory. Couldn't find 'mozilla-foundation/common_voice_13_0' on the Hugging Face Hub either: FileNotFoundError: Dataset 'mozilla-foundation/common_voice_13_0' doesn't exist on the Hub. If the repo is private or gated, make sure to log in with `huggingface-cli login`.\r\n```\r\n\r\nAnd tbh idk how you managed to get your error. \"n_shards.json\" is not even a thing in `datasets`",
"Okay, I am able to reproduce @patrickvonplaten's original error: https://github.com/Vaibhavs10/scratchpad/blob/main/cv13_datasets_test.ipynb\r\n\r\nAlso not sure why it looks for `n_shards.json`",
"Ok I see, this file is downloaded from the CV dataset script - let me investigate",
"Ok I see: when you log out you no longer have access to the repository.\r\n\r\nTherefore the dataset script is loaded from cache:\r\n```\r\nWARNING:datasets.load:Using the latest cached version of the module from /root/.cache/huggingface/modules/datasets_modules/datasets/mozilla-foundation--common_voice_13_0/22809012aac1fc9803eaffc44122e4149043748e93933935d5ea19898587e4d7 (last modified on Wed Jun 14 10:13:17 2023) since it couldn't be found locally at mozilla-foundation/common_voice_13_0., or remotely on the Hugging Face Hub.\r\n```\r\n\r\nand the script tries to download the n_shards.json but fails",
"Is this ok for you https://github.com/huggingface/datasets/pull/5954 ?\r\n\r\nI'll do a release this afternoon",
"Cool! ",
"this is included in the new release 2.13.0"
] | ### Describe the bug
When I attempt to download a model from the Hub that is gated without being logged in, I get a nice error message. E.g.:
E.g.
```sh
Repository Not Found for url: https://huggingface.co/api/models/DeepFloyd/IF-I-XL-v1.0.
Please make sure you specified the correct `repo_id` and `repo_type`.
If you are trying to access a private or gated repo, make sure you are authenticated.
Invalid username or password..
Will try to load from local cache.
```
If I do the same for a gated dataset on the Hub, I'm not gated a nice error message IMO:
```sh
File ~/hf/lib/python3.10/site-packages/fsspec/implementations/http.py:430, in HTTPFileSystem._info(self, url, **kwargs)
427 except Exception as exc:
428 if policy == "get":
429 # If get failed, then raise a FileNotFoundError
--> 430 raise FileNotFoundError(url) from exc
431 logger.debug(str(exc))
433 return {"name": url, "size": None, **info, "type": "file"}
FileNotFoundError: https://huggingface.co/datasets/mozilla-foundation/common_voice_13_0/resolve/main/n_shards.json
```
### Steps to reproduce the bug
```
huggingface-cli logout
```
and then:
```py
from datasets import load_dataset, Audio
# English
stream_data = load_dataset("mozilla-foundation/common_voice_13_0", "en", split="test", streaming=True)
stream_data = stream_data.cast_column("audio", Audio(sampling_rate=16000))
en_sample = next(iter(stream_data))["audio"]["array"]
# Swahili
stream_data = load_dataset("mozilla-foundation/common_voice_13_0", "sw", split="test", streaming=True)
stream_data = stream_data.cast_column("audio", Audio(sampling_rate=16000))
sw_sample = next(iter(stream_data))["audio"]["array"]
```
### Expected behavior
Better error message
### Environment info
Copy-and-paste the text below in your GitHub issue.
- `datasets` version: 2.12.0
- Platform: Linux-6.2.0-76060200-generic-x86_64-with-glibc2.35
- Python version: 3.10.6
- Huggingface_hub version: 0.16.0.dev0
- PyArrow version: 11.0.0
- Pandas version: 1.5.3
| 5,953 |
https://github.com/huggingface/datasets/issues/5951 | What is the Right way to use discofuse dataset?? | [
"Thanks for opening https://huggingface.co/datasets/discofuse/discussions/3, let's continue the discussion over there if you don't mind",
"I have posted there also sir, please check\r\n@lhoestq"
] | [Click here for Dataset link](https://huggingface.co/datasets/discofuse/viewer/discofuse-wikipedia/train?row=6)
**Below is the following way, as per my understanding , Is it correct :question: :question:**
The **columns/features from `DiscoFuse dataset`** that will be the **input to the `encoder` and `decoder`** are:
[Click here for Dataset link](https://huggingface.co/datasets/discofuse/viewer/discofuse-wikipedia/train?row=6)
1. **coherent_first_sentence**
2. **coherent_second_sentence**
3. **incoherent_first_sentence**
4. **incoherent_second_sentence**
[Click here for Dataset link](https://huggingface.co/datasets/discofuse/viewer/discofuse-wikipedia/train?row=6)
The **`encoder` will take these four columns as input and encode them into a sequence of hidden states. The `decoder` will then take these hidden states as input and decode them into a new sentence that fuses the two original sentences together.**
The **discourse type, connective_string, has_coref_type_pronoun, and has_coref_type_nominal columns will not be used as input to the encoder or decoder.** These columns are used to provide additional information about the dataset, but they are not necessary for the task of sentence fusion.
Please correct me if I am wrong; otherwise, if this understanding is right, how shall I implement this task practically? | 5,951 |
https://github.com/huggingface/datasets/issues/5950 | Support for data with instance-wise dictionary as features | [
"Hi ! We use the Arrow columnar format under the hood, which doesn't support such dictionaries: each field must have a fixed type and exist in each sample.\r\n\r\nInstead you can restructure your data like\r\n```\r\n{\r\n \"index\": 0,\r\n \"keys\": [\"2 * x + y >= 3\"],\r\n \"values\": [[\"2 * x + y >= 3\", \"4 * x + 2 * y >= 6\"]],\r\n }\r\n},\r\n...\r\n{\r\n \"index\": 9999,\r\n \"keys\": [\"x >= 6\"],\r\n \"values\": [[\"x >= 6\", \"x >= 0\", \"x >= -1\"]],\r\n},\r\n...\r\n```"
] | ### Feature request
I notice that when loading data instances with feature type of python dictionary, the dictionary keys would be broadcast so that every instance has the same set of keys. Please see an example in the Motivation section.
It is possible to avoid this behavior, i.e., load dictionary features as it is and do not broadcast the keys among instances? Please note that these dictionaries would have to be processed dynamically at each training iteration into strings (and tokenized).
### Motivation
I am trying to load a dataset from a json file. Each instance of the dataset has a feature that is a dictionary but its keys depend on the instance. Every two instances may have different keys. For example, imagine a dataset that contains a set of math expressions from a bunch of mutually redundant expressions:
```
{
"index": 0,
"feature": {
"2 * x + y >= 3": ["2 * x + y >= 3", "4 * x + 2 * y >= 6"],
...
}
},
...
{
"index": 9999,
"feature": {
"x >= 6": ["x >= 6", "x >= 0", "x >= -1"],
...
}
},
...
```
When directly loading the dataset using `data = load_dataset("json", data_files=file_paths, split='train')`, each instance would have all the keys from other instances and None as values. That is, instance of index 0 becomes:
```
{
"index": 0,
"feature": {
"2 * x + y >= 3": ["2 * x + y >= 3", "4 * x + 2 * y >= 6"],
...
"x >= 6": None, # keys from other instances
...
}
},
```
This is not desirable. Moreover, issue would be raised if I attempt to combine two such datasets using `data = concatenate_datasets(multi_datasets)`, perhaps because their dictionary features contain different keys.
A solution I can think of is to store the dictionary features as a long string, and evaluate it later. Please kindly suggest any other solution using existing methods of datasets.
### Your contribution
N/A | 5,950 |
https://github.com/huggingface/datasets/issues/5947 | Return the audio filename when decoding fails due to corrupt files | [
"Hi ! The audio data don't always exist as files on disk - the blobs are often stored in the Arrow files. For now I'd suggest disabling decoding with `.cast_column(\"audio\", Audio(decode=False))` and apply your own decoding that handles corrupted files (maybe to filter them out ?)\r\n\r\ncc @sanchit-gandhi since it's related to our discussion about allowing users to make decoding return `None` and show a warning when there are corrupted files",
"Thanks @lhoestq, I wasn't aware of the decode flag. It makes more sense as you say to show a warning when there are corrupted files together with some metadata of the file that allows to filter them from the dataset.\r\n\r\nMy workaround was to catch the LibsndfileError and generate a dummy audio with an unsual sample rate to filter it later. However returning `None` seems better. \r\n\r\n`try:\r\n array, sampling_rate = sf.read(file)\r\nexcept sf.LibsndfileError:\r\n print(\"bad file\")\r\n array = np.array([0.0])\r\n sampling_rate = 99.000` \r\n\r\n"
] | ### Feature request
Return the audio filename when the audio decoding fails. Although currently there are some checks for mp3 and opus formats with the library version there are still cases when the audio decoding could fail, eg. Corrupt file.
### Motivation
When you try to load an object file dataset and the decoding fails you can't know which file is corrupt
```
raise LibsndfileError(err, prefix="Error opening {0!r}: ".format(self.name))
soundfile.LibsndfileError: Error opening <_io.BytesIO object at 0x7f5ab7e38290>: Format not recognised.
```
### Your contribution
Make a PR to Add exceptions for LIbsndfileError to return the audio filename or path when soundfile decoding fails. | 5,947 |
https://github.com/huggingface/datasets/issues/5946 | IndexError Not Solving -> IndexError: Invalid key: ?? is out of bounds for size 0 or ?? | [
"https://colab.research.google.com/#scrollTo=AQ_HCYruWIHU&fileId=https%3A//huggingface.co/dfurman/falcon-40b-chat-oasst1/blob/main/finetune_falcon40b_oasst1_with_bnb_peft.ipynb\r\n\r\nI ran the same administration exactly the same but got the same error",
"Looks related to /static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Findexerror-invalid-key-16-is-out-of-bounds-for-size-0%2F14298%2F4%3Fu%3Dlhoestq",
"> Looks related to /static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Findexerror-invalid-key-16-is-out-of-bounds-for-size-0%2F14298%2F4%3Fu%3Dlhoestq%5Cn%5CnThe problem has not been solved, I have tried this before, but the problem is the same",
"> \r\n\r\n@syngokhan did u solve it? \r\nI am desperate ",
"data = data[\"train\"].shuffle().map(generate_and_tokenize_prompt, batched = False) # change this line to -\r\n\r\ndata[\"train\"] = data[\"train\"].shuffle().map(generate_and_tokenize_prompt, batched = False)\r\nAfter doing this change you code should run fine.",
"> > \r\n> \r\n> @syngokhan did u solve it? I am desperate\r\n\r\nrefer to my earlier comment. you will find the solution."
] | ### Describe the bug
in <cell line: 1>:1 │
│ │
│ /usr/local/lib/python3.10/dist-packages/transformers/trainer.py:1537 in train │
│ │
│ 1534 │ │ inner_training_loop = find_executable_batch_size( │
│ 1535 │ │ │ self._inner_training_loop, self._train_batch_size, args.auto_find_batch_size │
│ 1536 │ │ ) │
│ ❱ 1537 │ │ return inner_training_loop( │
│ 1538 │ │ │ args=args, │
│ 1539 │ │ │ resume_from_checkpoint=resume_from_checkpoint, │
│ 1540 │ │ │ trial=trial, │
│ │
│ /usr/local/lib/python3.10/dist-packages/transformers/trainer.py:1789 in _inner_training_loop │
│ │
│ 1786 │ │ │ │ rng_to_sync = True │
│ 1787 │ │ │ │
│ 1788 │ │ │ step = -1 │
│ ❱ 1789 │ │ │ for step, inputs in enumerate(epoch_iterator): │
│ 1790 │ │ │ │ total_batched_samples += 1 │
│ 1791 │ │ │ │ if rng_to_sync: │
│ 1792 │ │ │ │ │ self._load_rng_state(resume_from_checkpoint) │
│ │
│ /usr/local/lib/python3.10/dist-packages/accelerate/data_loader.py:377 in __iter__ │
│ │
│ 374 │ │ dataloader_iter = super().__iter__() │
│ 375 │ │ # We iterate one batch ahead to check when we are at the end │
│ 376 │ │ try: │
│ ❱ 377 │ │ │ current_batch = next(dataloader_iter) │
│ 378 │ │ except StopIteration: │
│ 379 │ │ │ yield │
│ 380 │
│ │
│ /usr/local/lib/python3.10/dist-packages/torch/utils/data/dataloader.py:633 in __next__ │
│ │
│ 630 │ │ │ if self._sampler_iter is None: │
│ 631 │ │ │ │ # TODO(https://github.com/pytorch/pytorch/issues/76750) │
│ 632 │ │ │ │ self._reset() # type: ignore[call-arg] │
│ ❱ 633 │ │ │ data = self._next_data() │
│ 634 │ │ │ self._num_yielded += 1 │
│ 635 │ │ │ if self._dataset_kind == _DatasetKind.Iterable and \ │
│ 636 │ │ │ │ │ self._IterableDataset_len_called is not None and \ │
│ │
│ /usr/local/lib/python3.10/dist-packages/torch/utils/data/dataloader.py:677 in _next_data │
│ │
│ 674 │ │
│ 675 │ def _next_data(self): │
│ 676 │ │ index = self._next_index() # may raise StopIteration │
│ ❱ 677 │ │ data = self._dataset_fetcher.fetch(index) # may raise StopIteration │
│ 678 │ │ if self._pin_memory: │
│ 679 │ │ │ data = _utils.pin_memory.pin_memory(data, self._pin_memory_device) │
│ 680 │ │ return data │
│ │
│ /usr/local/lib/python3.10/dist-packages/torch/utils/data/_utils/fetch.py:49 in fetch │
│ │
│ 46 │ def fetch(self, possibly_batched_index): │
│ 47 │ │ if self.auto_collation: │
│ 48 │ │ │ if hasattr(self.dataset, "__getitems__") and self.dataset.__getitems__: │
│ ❱ 49 │ │ │ │ data = self.dataset.__getitems__(possibly_batched_index) │
│ 50 │ │ │ else: │
│ 51 │ │ │ │ data = [self.dataset[idx] for idx in possibly_batched_index] │
│ 52 │ │ else: │
│ │
│ /usr/local/lib/python3.10/dist-packages/datasets/arrow_dataset.py:2782 in __getitems__ │
│ │
│ 2779 │ │
│ 2780 │ def __getitems__(self, keys: List) -> List: │
│ 2781 │ │ """Can be used to get a batch using a list of integers indices.""" │
│ ❱ 2782 │ │ batch = self.__getitem__(keys) │
│ 2783 │ │ n_examples = len(batch[next(iter(batch))]) │
│ 2784 │ │ return [{col: array[i] for col, array in batch.items()} for i in range(n_example │
│ 2785 │
│ │
│ /usr/local/lib/python3.10/dist-packages/datasets/arrow_dataset.py:2778 in __getitem__ │
│ │
│ 2775 │ │
│ 2776 │ def __getitem__(self, key): # noqa: F811 │
│ 2777 │ │ """Can be used to index columns (by string names) or rows (by integer index or i │
│ ❱ 2778 │ │ return self._getitem(key) │
│ 2779 │ │
│ 2780 │ def __getitems__(self, keys: List) -> List: │
│ 2781 │ │ """Can be used to get a batch using a list of integers indices.""" │
│ │
│ /usr/local/lib/python3.10/dist-packages/datasets/arrow_dataset.py:2762 in _getitem │
│ │
│ 2759 │ │ format_kwargs = kwargs["format_kwargs"] if "format_kwargs" in kwargs else self._ │
│ 2760 │ │ format_kwargs = format_kwargs if format_kwargs is not None else {} │
│ 2761 │ │ formatter = get_formatter(format_type, features=self._info.features, **format_kw │
│ ❱ 2762 │ │ pa_subtable = query_table(self._data, key, indices=self._indices if self._indice │
│ 2763 │ │ formatted_output = format_table( │
│ 2764 │ │ │ pa_subtable, key, formatter=formatter, format_columns=format_columns, output │
│ 2765 │ │ ) │
│ │
│ /usr/local/lib/python3.10/dist-packages/datasets/formatting/formatting.py:578 in query_table │
│ │
│ 575 │ │ _check_valid_column_key(key, table.column_names) │
│ 576 │ else: │
│ 577 │ │ size = indices.num_rows if indices is not None else table.num_rows │
│ ❱ 578 │ │ _check_valid_index_key(key, size) │
│ 579 │ # Query the main table │
│ 580 │ if indices is None: │
│ 581 │ │ pa_subtable = _query_table(table, key) │
│ │
│ /usr/local/lib/python3.10/dist-packages/datasets/formatting/formatting.py:531 in │
│ _check_valid_index_key │
│ │
│ 528 │ │ │ _check_valid_index_key(min(key), size=size) │
│ 529 │ elif isinstance(key, Iterable): │
│ 530 │ │ if len(key) > 0: │
│ ❱ 531 │ │ │ _check_valid_index_key(int(max(key)), size=size) │
│ 532 │ │ │ _check_valid_index_key(int(min(key)), size=size) │
│ 533 │ else: │
│ 534 │ │ _raise_bad_key_type(key) │
│ │
│ /usr/local/lib/python3.10/dist-packages/datasets/formatting/formatting.py:521 in │
│ _check_valid_index_key │
│ │
│ 518 def _check_valid_index_key(key: Union[int, slice, range, Iterable], size: int) -> None: │
│ 519 │ if isinstance(key, int): │
│ 520 │ │ if (key < 0 and key + size < 0) or (key >= size): │
│ ❱ 521 │ │ │ raise IndexError(f"Invalid key: {key} is out of bounds for size {size}") │
│ 522 │ │ return │
│ 523 │ elif isinstance(key, slice): │
│ 524 │ │ pass
### Steps to reproduce the bug
``
import json
import os
from pprint import pprint
import bitsandbytes as bnb
import pandas as pd
import torch
import torch.nn as nn
import transformers
from datasets import Dataset,load_dataset
from peft import (
LoraConfig,
PeftConfig,
PeftModel,
get_peft_model,
prepare_model_for_kbit_training
)
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoTokenizer,
BitsAndBytesConfig,
)
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
def print_trainable_parameters(model):
"""
Prints the number of trainable parameters in the model.
"""
trainable_params = 0
all_param = 0
for _, param in model.named_parameters():
all_param += param.numel()
if param.requires_grad:
trainable_params += param.numel()
print(
f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
)
MODEL_NAME = "tiiuae/falcon-7b"
bnb_config = BitsAndBytesConfig(
load_in_4bit = True,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16,
)
model = AutoModelForCausalLM.from_pretrained(
MODEL_NAME,
device_map = "auto",
trust_remote_code = True,
quantization_config = bnb_config
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
tokenizer.pad_token = tokenizer.eos_token
model.gradient_checkpointing_enable()
model = prepare_model_for_kbit_training(model)
config = LoraConfig(
r = 16,
lora_alpha = 32,
target_modules = ["query_key_value"],
lora_dropout = 0.05,
bias = "none",
task_type = "CASUAL_LM"
)
model = get_peft_model(model,config)
print_trainable_parameters(model)
def generate_prompt(data_point):
return f"""
<human>: {data_point["question"]}
<assistant>: {data_point["answer"]}
""".strip()
def generate_and_tokenize_prompt(data_point):
full_prompt = generate_prompt(data_point)
tokenized_full_prompt = tokenizer(full_prompt, padding = True, truncation = True,return_tensors = None)
return dict({
"input_ids" : tokenized_full_prompt["input_ids"],
"attention_mask" : tokenized_full_prompt["attention_mask"]
})
data = data["train"].shuffle().map(generate_and_tokenize_prompt, batched = False)
OUTPUT_DIR = "experiments"
trainings_args = transformers.TrainingArguments(
per_device_train_batch_size = 1,
gradient_accumulation_steps = 4,
num_train_epochs = 1,
learning_rate = 2e-4,
fp16 = True,
save_total_limit = 3,
logging_steps = 1,
output_dir = OUTPUT_DIR,
max_steps = 80,
optim = "paged_adamw_8bit",
lr_scheduler_type = "cosine",
warmup_ratio = 0.05,
#remove_unused_columns=True
)
trainer = transformers.Trainer(
model = model,
train_dataset = data,
args = trainings_args,
data_collator = transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),
)
model.config.use_cache = False
trainer.train()
IndexError: Invalid key: 32 is out of bounds for size 0
DataSet Format is like :
[{"question": "How can I create an account?", "answer": "To create an account, click on the 'Sign Up' button on the top right corner of our website and follow the instructions to complete the registration process."}, .... ]
### Expected behavior
-
### Environment info
!pip install -q pip
!pip install -q bitsandbytes==0.39.0
!pip install -q torch==2.0.1
!pip install -q git+https://github.com/huggingface/transformers.git
!pip install -q git+https://github.com/huggingface/peft.git
!pip install -q git+https://github.com/huggingface/accelerate.git
!pip install -q datasets
!pip install -q loralib==0.1.1
!pip install -q einops==0.6.1
import json
import os
from pprint import pprint
import bitsandbytes as bnb
import pandas as pd
import torch
import torch.nn as nn
import transformers
from datasets import Dataset,load_dataset
from peft import (
LoraConfig,
PeftConfig,
PeftModel,
get_peft_model,
prepare_model_for_kbit_training
)
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoTokenizer,
BitsAndBytesConfig,
)
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
| 5,946 |
https://github.com/huggingface/datasets/issues/5945 | Failing to upload dataset to the hub | [
"Hi ! Feel free to re-run your code later, it will resume automatically where you left",
"Tried many times in the last 2 weeks, problem remains.",
"Alternatively you can save your dataset in parquet files locally and upload them to the hub manually\r\n\r\n```python\r\nfrom tqdm import tqdm\r\nnum_shards = 60\r\nfor index in tqdm(range(num_shards)):\r\n ds.shard(num_shards=num_shards, index=index, contiguous=True).to_parquet(f\"{index:05d}.parquet\")\r\n````"
] | ### Describe the bug
Trying to upload a dataset of hundreds of thousands of audio samples (the total volume is not very large, 60 gb) to the hub with push_to_hub, it doesn't work.
From time to time one piece of the data (parquet) gets pushed and then I get RemoteDisconnected even though my internet is stable.
Please help.
I'm trying to upload the dataset for almost a week.
Thanks
### Steps to reproduce the bug
not relevant
### Expected behavior
Be able to upload thedataset
### Environment info
python: 3.9 | 5,945 |
https://github.com/huggingface/datasets/issues/5941 | Load Data Sets Too Slow In Train Seq2seq Model | [
"Hi ! you can speed it up using multiprocessing by passing `num_proc=` to `load_dataset()`",
"already did，but not useful for step Generating train split，it works in step \"Resolving data files\" & \"Downloading data files\" ",
"@mariosasko some advice ， thanks！",
"I met the same problem, terrible experience",
"@mariosasko ",
"We need more info about the issue to provide help. \r\n\r\nCan you interrupt the process (with `num_proc=None`) after the `load_dataset` call when the slowdown occurs? So we can know what part of the code is causing it.\r\n\r\nThe `audiofolder` \\ `imagefolder` with metadata is not performant for large datasets. Luckily, we can make them much faster if drop the nested metadata files feature (not that useful). I plan to work on this soon.\r\n\r\nIn the meantime, it's better to use `Dataset.from_generator` (requires replacing the `load_dataset` calls in the transformers script with `Dataset.from_generator`) or write a dataset loading script for large datasets.",
"Can you interrupt the process (with num_proc=None) after the load_dataset call when the slowdown occurs? So we can know what part of the code is causing it.\r\n（I'll try this operation）\r\nThe audiofolder \\ imagefolder with metadata is not performant for large datasets. Luckily, we can make them much faster if drop the nested metadata files feature (not that useful). I plan to work on this soon.\r\n(My data is indeed a bit large, exceeding 10000 hours of audio data. Looking forward to your improvement work very much)\r\n\r\nIn the meantime, it's better to use Dataset.from_generator (requires replacing the load_dataset calls in the transformers script with Dataset.from_generator) or write a dataset loading script for large datasets.\r\n（I want to use Dataset.from_generator instead of load_dataset ，where can i found sample code to load audio&label dataset， I was to do asr task）",
"Can you interrupt the process (with num_proc=None) after the load_dataset call when the slowdown occurs? So we can know what part of the code is causing it.\r\n================================================================================\r\nHere is the log：\r\n[load_dataset.log](https://github.com/huggingface/datasets/files/12169362/load_dataset.log)\r\n（The larger my training data, the slower it loads）\r\n\r\n\r\n",
"In the meantime, it's better to use Dataset.from_generator (requires replacing the load_dataset calls in the transformers script with Dataset.from_generator) or write a dataset loading script for large datasets.\r\n================================================================================\r\nI tried ‘Dataset. from_generator’ implements data loading, but the testing results show no improvement",
"I have already solved this problem, referring to #5990 : read audio frist, then use data_generator to change format ."
] | ### Describe the bug
step 'Generating train split' in load_dataset is too slow：
### Steps to reproduce the bug
Data： own data，16K16B Mono wav
Oficial Script:[ run_speech_recognition_seq2seq.py](https://github.com/huggingface/transformers/blob/main/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py)
Add Code：
if data_args.data_path is not None:
print(data_args.data_path)
raw_datasets = load_dataset("audiofolder", data_dir=data_args.data_path, cache_dir=model_args.cache_dir)
raw_datasets = raw_datasets.cast_column("audio", Audio(sampling_rate=16000))
raw_datasets = raw_datasets["train"].train_test_split(test_size=0.005, shuffle=True)
（change cache_dir to other path ，ex:/DATA/cache）
### Expected behavior
load data fast,at least 1000+
`Generating train split: 387875 examples [32:24:45, 1154.83 examples/s]`
### Environment info
- `transformers` version: 4.28.0.dev0
- Platform: Linux-5.4.0-149-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.16
- Huggingface_hub version: 0.13.2
- PyTorch version (GPU?): 1.13.1+cu116 (True)
- Tensorflow version (GPU?): not installed (NA)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Using GPU in script?: <fill in>
- Using distributed or parallel set-up in script?: <fill in> | 5,941 |
https://github.com/huggingface/datasets/issues/5990 | Pushing a large dataset on the hub consistently hangs | [
"Hi @AntreasAntoniou , sorry to know you are facing this issue. To help debugging it, could you tell me:\r\n- What is the total dataset size?\r\n- Is it always failing on the same shard or is the hanging problem happening randomly?\r\n- Were you able to save the dataset as parquet locally? This would help us determine if the problem comes from the upload or the file generation.\r\n\r\nI'm cc-ing @lhoestq who might have some insights from a `datasets` perspective.",
"One trick that can also help is to check the traceback when you kill your python process: it will show where in the code it was hanging",
"Right. So I did the trick @lhoestq suggested. Here is where things seem to hang\r\n\r\n```\r\nError while uploading 'data/train-00120-of-00195-466c2dbab2eb9989.parquet' to the Hub. \r\nPushing split train to the Hub. \r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:03<00:00, 1.15s/ba]\r\nUpload 1 LFS files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:52<00:00, 52.12s/it]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:03<00:00, 1.08s/ba]\r\nUpload 1 LFS files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:45<00:00, 45.54s/it]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:03<00:00, 1.08s/ba]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:03<00:00, 1.03s/ba^Upload 1 LFS files: 0%| | 0/1 [\r\n21:27:35<?, ?it/s] \r\nPushing dataset shards to the dataset hub: 63%|█████████████████████████████████████████████████████████████▎ | 122/195 [23:37:11<14:07:59, 696.98s/it]\r\n^CError in sys.excepthook: \r\nTraceback (most recent call last): \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1699, in print \r\n extend(render(renderable, render_options)) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1335, in render \r\n yield from self.render(render_output, _options) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1331, in render \r\n for render_output in iter_render: \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/constrain.py\", line 29, in __rich_console__ \r\n yield from console.render(self.renderable, child_options) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1331, in render \r\n for render_output in iter_render: \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/panel.py\", line 220, in __rich_console__ \r\n lines = console.render_lines(renderable, child_options, style=style) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1371, in render_lines \r\n lines = list( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/segment.py\", line 292, in split_and_crop_lines \r\n for segment in segments: \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1331, in render \r\n for render_output in iter_render: \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/padding.py\", line 97, in __rich_console__ \r\n lines = console.render_lines( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1371, in render_lines \r\n lines = list( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/segment.py\", line 292, in split_and_crop_lines \r\n for segment in segments: \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1335, in render \r\n yield from self.render(render_output, _options) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/console.py\", line 1331, in render \r\n for render_output in iter_render: \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/syntax.py\", line 611, in __rich_console__ \r\n segments = Segments(self._get_syntax(console, options)) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/segment.py\", line 668, in __init__ \r\n self.segments = list(segments) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/syntax.py\", line 674, in _get_syntax \r\n lines: Union[List[Text], Lines] = text.split(\"\\n\", allow_blank=ends_on_nl) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/text.py\", line 1042, in split \r\n lines = Lines( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/containers.py\", line 70, in __init__ \r\n self._lines: List[\"Text\"] = list(lines) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/text.py\", line 1043, in <genexpr> \r\n line for line in self.divide(flatten_spans()) if line.plain != separator \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/rich/text.py\", line 385, in plain \r\n if len(self._text) != 1: \r\nKeyboardInterrupt \r\n \r\nOriginal exception was: \r\nTraceback (most recent call last): \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/tqdm/contrib/concurrent.py\", line 51, in _executor_map \r\n return list(tqdm_class(ex.map(fn, *iterables, chunksize=chunksize), **kwargs)) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/tqdm/std.py\", line 1178, in __iter__ \r\n for obj in iterable: \r\n File \"/opt/conda/envs/main/lib/python3.10/concurrent/futures/_base.py\", line 621, in result_iterator \r\n yield _result_or_cancel(fs.pop()) \r\n File \"/opt/conda/envs/main/lib/python3.10/concurrent/futures/_base.py\", line 319, in _result_or_cancel \r\n return fut.result(timeout) \r\n File \"/opt/conda/envs/main/lib/python3.10/concurrent/futures/_base.py\", line 453, in result \r\n self._condition.wait(timeout) \r\n File \"/opt/conda/envs/main/lib/python3.10/threading.py\", line 320, in wait \r\n waiter.acquire() \r\nKeyboardInterrupt \r\n \r\nDuring handling of the above exception, another exception occurred: \r\n \r\nTraceback (most recent call last): \r\n File \"/TALI/tali/scripts/validate_dataset.py\", line 127, in <module> \r\n train_dataset.push_to_hub(repo_id=\"Antreas/TALI-base\", max_shard_size=\"5GB\") \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/datasets/dataset_dict.py\", line 1583, in push_to_hub \r\n repo_id, split, uploaded_size, dataset_nbytes, _, _ = self[split]._push_parquet_shards_to_hub( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_dataset.py\", line 5275, in _push_parquet_shards_to_hub \r\n _retry( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/datasets/utils/file_utils.py\", line 282, in _retry \r\n return func(*func_args, **func_kwargs) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py\", line 118, in _inner_fn \r\n return fn(*args, **kwargs) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/hf_api.py\", line 826, in _inner \r\n return fn(self, *args, **kwargs) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/hf_api.py\", line 3205, in upload_file \r\n commit_info = self.create_commit( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py\", line 118, in _inner_fn \r\n return fn(*args, **kwargs) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/hf_api.py\", line 826, in _inner \r\n return fn(self, *args, **kwargs) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/hf_api.py\", line 2680, in create_commit \r\n upload_lfs_files( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py\", line 118, in _inner_fn \r\n return fn(*args, **kwargs) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/_commit_api.py\", line 353, in upload_lfs_files \r\n thread_map( \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/tqdm/contrib/concurrent.py\", line 69, in thread_map \r\n return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs) \r\n File \"/opt/conda/envs/main/lib/python3.10/site-packages/tqdm/contrib/concurrent.py\", line 49, in _executor_map \r\n with PoolExecutor(max_workers=max_workers, initializer=tqdm_class.set_lock, \r\n File \"/opt/conda/envs/main/lib/python3.10/concurrent/futures/_base.py\", line 649, in __exit__ \r\n self.shutdown(wait=True) \r\n File \"/opt/conda/envs/main/lib/python3.10/concurrent/futures/thread.py\", line 235, in shutdown \r\n t.join() \r\n File \"/opt/conda/envs/main/lib/python3.10/threading.py\", line 1096, in join \r\n self._wait_for_tstate_lock() \r\n File \"/opt/conda/envs/main/lib/python3.10/threading.py\", line 1116, in _wait_for_tstate_lock \r\n if lock.acquire(block, timeout): \r\nKeyboardInterrupt \r\n```",
"@Wauplin \r\n\r\n>What is the total dataset size?\r\n\r\nThere are three variants, and the random hanging happens on all three. The sizes are 2TB, 1TB, and 200GB. \r\n\r\n>Is it always failing on the same shard or is the hanging problem happening randomly?\r\n\r\nIt seems to be very much random, as restarting can help move past the previous hang, only to find a new one, or not. \r\n\r\n>Were you able to save the dataset as parquet locally? This would help us determine if the problem comes from the upload or the file generation.\r\n\r\nYes. The dataset seems to be locally stored as parquet. ",
"Hmm it looks like an issue with TQDM lock. Maybe you can try updating TQDM ?",
"I am using the latest version of tqdm\r\n\r\n```\r\n⬢ [Docker] ❯ pip install tqdm --upgrade\r\nRequirement already satisfied: tqdm in /opt/conda/envs/main/lib/python3.10/site-packages (4.65.0)\r\nWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\r\n```",
"I tried trying to catch the hanging issue in action again\r\n\r\n```\r\nPushing dataset shards to the dataset hub: 65%|█████████████████████████████████████████████████████████████████▊ | 127/195 [2:28:02<1:19:15, 69.94s/it] \r\nError while uploading 'data/train-00127-of-00195-3f8d036ade107c27.parquet' to the Hub. \r\nPushing split train to the Hub. \r\nPushing dataset shards to the dataset hub: 64%|████████████████████████████████████████████████████████████████▏ | 124/195 [2:06:10<1:12:14, 61.05s/it]C^[^C^C^C \r\n╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮ \r\n│ /TALI/tali/scripts/validate_dataset.py:127 in <module> │ \r\n│ │ \r\n│ 124 │ │ \r\n│ 125 │ while not succesful_competion: │ \r\n│ 126 │ │ try: │ \r\n│ ❱ 127 │ │ │ train_dataset.push_to_hub(repo_id=\"Antreas/TALI-base\", max_shard_size=\"5GB\") │ \r\n│ 128 │ │ │ succesful_competion = True │ \r\n│ 129 │ │ except Exception as e: │ \r\n│ 130 │ │ │ print(e) │ \r\n│ │ \r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/dataset_dict.py:1583 in push_to_hub │ \r\n│ │ \r\n│ 1580 │ │ for split in self.keys(): │ \r\n│ 1581 │ │ │ logger.warning(f\"Pushing split {split} to the Hub.\") │ \r\n│ 1582 │ │ │ # The split=key needs to be removed before merging │ \r\n│ ❱ 1583 │ │ │ repo_id, split, uploaded_size, dataset_nbytes, _, _ = self[split]._push_parq │ \r\n│ 1584 │ │ │ │ repo_id, │ \r\n│ 1585 │ │ │ │ split=split, │ \r\n│ 1586 │ │ │ │ private=private, │ \r\n│ │ \r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_dataset.py:5263 in │ \r\n│ _push_parquet_shards_to_hub │ \r\n│ │ \r\n│ 5260 │ │ │ \r\n│ 5261 │ │ uploaded_size = 0 │ \r\n│ 5262 │ │ shards_path_in_repo = [] │ \r\n│ ❱ 5263 │ │ for index, shard in logging.tqdm( │ \r\n│ 5264 │ │ │ enumerate(itertools.chain([first_shard], shards_iter)), │ \r\n│ 5265 │ │ │ desc=\"Pushing dataset shards to the dataset hub\", │ \r\n│ 5266 │ │ │ total=num_shards, │ \r\n│ │ \r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/tqdm/std.py:1178 in __iter__ │ \r\n│ │ \r\n│ 1175 │ │ time = self._time │ \r\n│ 1176 │ │ │ \r\n│ 1177 │ │ try: │\r\n│ ❱ 1178 │ │ │ for obj in iterable: │\r\n│ 1179 │ │ │ │ yield obj │\r\n│ 1180 │ │ │ │ # Update and possibly print the progressbar. │\r\n│ 1181 │ │ │ │ # Note: does not call self.update(1) for speed optimisation. │\r\n│ │\r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_dataset.py:5238 in │\r\n│ shards_with_embedded_external_files │\r\n│ │\r\n│ 5235 │ │ │ │ for shard in shards: │\r\n│ 5236 │ │ │ │ │ format = shard.format │\r\n│ 5237 │ │ │ │ │ shard = shard.with_format(\"arrow\") │\r\n│ ❱ 5238 │ │ │ │ │ shard = shard.map( │\r\n│ 5239 │ │ │ │ │ │ embed_table_storage, │\r\n│ 5240 │ │ │ │ │ │ batched=True, │\r\n│ 5241 │ │ │ │ │ │ batch_size=1000, │\r\n│ │\r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_dataset.py:578 in wrapper │\r\n│ │\r\n│ 575 │ │ else: │\r\n│ 576 │ │ │ self: \"Dataset\" = kwargs.pop(\"self\") │\r\n│ 577 │ │ # apply actual function │\r\n│ ❱ 578 │ │ out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs) │ \r\n│ 579 │ │ datasets: List[\"Dataset\"] = list(out.values()) if isinstance(out, dict) else [ou │ \r\n│ 580 │ │ for dataset in datasets: │ \r\n│ 581 │ │ │ # Remove task templates if a column mapping of the template is no longer val │ \r\n│ │ \r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_dataset.py:543 in wrapper │ \r\n│ │ \r\n│ 540 │ │ │ \"output_all_columns\": self._output_all_columns, │ \r\n│ 541 │ │ } │ \r\n│ 542 │ │ # apply actual function │ \r\n│ ❱ 543 │ │ out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs) │ \r\n│ 544 │ │ datasets: List[\"Dataset\"] = list(out.values()) if isinstance(out, dict) else [ou │ \r\n│ 545 │ │ # re-apply format to the output │ \r\n│ 546 │ │ for dataset in datasets: │ \r\n│ │ \r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_dataset.py:3073 in map │ \r\n│ │ \r\n│ 3070 │ │ │ │ │ leave=False, │ \r\n│ 3071 │ │ │ │ │ desc=desc or \"Map\", │ \r\n│ 3072 │ │ │ │ ) as pbar: │ \r\n│ ❱ 3073 │ │ │ │ │ for rank, done, content in Dataset._map_single(**dataset_kwargs): │ \r\n│ 3074 │ │ │ │ │ │ if done: │ \r\n│ 3075 │ │ │ │ │ │ │ shards_done += 1 │ \r\n│ 3076 │ │ │ │ │ │ │ logger.debug(f\"Finished processing shard number {rank} of {n │ \r\n│ │ \r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_dataset.py:3464 in _map_single │ \r\n│ │ \r\n│ 3461 │ │ │ │ │ │ │ │ buf_writer, writer, tmp_file = init_buffer_and_writer() │ \r\n│ 3462 │ │ │ │ │ │ │ │ stack.enter_context(writer) │ \r\n│ 3463 │ │ │ │ │ │ │ if isinstance(batch, pa.Table): │ \r\n│ ❱ 3464 │ │ │ │ │ │ │ │ writer.write_table(batch) │ \r\n│ 3465 │ │ │ │ │ │ │ else: │ \r\n│ 3466 │ │ │ │ │ │ │ │ writer.write_batch(batch) │ \r\n│ 3467 │ │ │ │ │ │ num_examples_progress_update += num_examples_in_batch │ \r\n│ │ \r\n│ /opt/conda/envs/main/lib/python3.10/site-packages/datasets/arrow_writer.py:567 in write_table │ \r\n│ │ \r\n│ 564 │ │ │ writer_batch_size = self.writer_batch_size │ \r\n│ 565 │ │ if self.pa_writer is None: │ \r\n│ 566 │ │ │ self._build_writer(inferred_schema=pa_table.schema) │ \r\n│ ❱ 567 │ │ pa_table = pa_table.combine_chunks() │ \r\n│ 568 │ │ pa_table = table_cast(pa_table, self._schema) │ \r\n│ 569 │ │ if self.embed_local_files: │ \r\n│ 570 │ │ │ pa_table = embed_table_storage(pa_table) │ \r\n╰──────────────────────────────────────────────────────────────────────────────────────────────────╯ \r\nKeyboardInterrupt \r\n```",
"I'm on my phone so can't help that much. What I'd advice to do is to [save_to_disk](https://huggingface.co/docs/datasets/package_reference/main_classes#save_to_disk) if it's not already done and then upload the files/folder to the Hub separately. You can find what you need in the [upload guide](https://huggingface.co/docs/huggingface_hub/guides/upload). It might not help finding the exact issue for now but at least it can unblock you. ",
"In your last stacktrace it interrupted while embedding external content - in case your dataset in made of images or audio files that live on your disk. Is it the case ?",
"Yeah, the dataset has images, audio, video and text. ",
"It's maybe related to https://github.com/apache/arrow/issues/34455: are you using ArrayND features ?\r\n\r\nAlso what's your `pyarrow` version ? Could you try updating to >= 12.0.1 ?",
"I was using pyarrow == 12.0.0\r\n\r\nI am not explicitly using ArrayND features, unless the hub API automatically converts my files to such. ",
"I have now updated to pyarrow == 12.0.1 and retrying",
"You can also try to reduce the `max_shard_size` - Sometimes parquet has a hard time working with data bigger than 2GB",
"So, updating the pyarrow seems to help. It can still throw errors here and there but I can retry when that happens. It's better than hanging. \r\n\r\nHowever, I am a bit confused about something. I have uploaded my datasets, but while earlier I could see all three sets, now I can only see 1. What's going on? \r\nhttps://huggingface.co/datasets/Antreas/TALI-base\r\n\r\nI have seen this happen before as well, so I deleted and reuploaded, but this dataset is way too large for me to do this. ",
"It's a bug on our side, I'll update the dataset viewer ;)\r\n\r\nThanks for reporting !",
"Apparently this happened because of bad modifications in the README.md split metadata.\r\n\r\nI fixed them in this PR: https://huggingface.co/datasets/Antreas/TALI-base/discussions/1",
"@lhoestq It's a bit odd that when uploading a dataset, one set at a time \"train\", \"val\", \"test\", the push_to_hub function overwrites the readme and removes differently named sets from previous commits. i.e., you push \"val\", all is well. Then you push \"test\", and the \"val\" entry disappears from the readme, while the data remain intact. ",
"Also, just found another related issue. One of the many that make things hang or fail when pushing to hub. \r\n\r\nIn the following code:\r\n\r\n```python\r\ntrain_generator = lambda: data_generator(\"train\", percentage=1.0)\r\n val_generator = lambda: data_generator(\"val\")\r\n test_generator = lambda: data_generator(\"test\")\r\n\r\n train_data = datasets.Dataset.from_generator(\r\n train_generator,\r\n num_proc=mp.cpu_count(),\r\n writer_batch_size=5000,\r\n cache_dir=tali_dataset_dir,\r\n )\r\n\r\n val_data = datasets.Dataset.from_generator(\r\n val_generator,\r\n writer_batch_size=5000,\r\n num_proc=mp.cpu_count(),\r\n cache_dir=tali_dataset_dir,\r\n )\r\n\r\n test_data = datasets.Dataset.from_generator(\r\n test_generator,\r\n writer_batch_size=5000,\r\n num_proc=mp.cpu_count(),\r\n cache_dir=tali_dataset_dir,\r\n )\r\n\r\n print(f\"Pushing TALI-large to hub\")\r\n\r\n dataset = datasets.DatasetDict(\r\n {\"train\": train_data, \"val\": val_data, \"test\": test_data}\r\n )\r\n succesful_competion = False\r\n\r\n while not succesful_competion:\r\n try:\r\n dataset.push_to_hub(repo_id=\"Antreas/TALI-large\", max_shard_size=\"2GB\")\r\n succesful_competion = True\r\n except Exception as e:\r\n print(e)\r\n ```\r\n \r\n \r\n Things keep failing in the push_to_repo step, at random places, with the following error:\r\n \r\n ```bash\r\n Pushing dataset shards to the dataset hub: 7%|██████████▋ | 67/950 [42:41<9:22:37, 38.23s/it]\r\nError while uploading 'data/train-00067-of-00950-a4d179ed5a593486.parquet' to the Hub.\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:01<00:00, 1.81ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:11<00:00, 11.20s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.48ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:15<00:00, 15.30s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.39ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:11<00:00, 11.52s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.47ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.39s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.26ba/s]\r\nUpload 1 LFS files: 0%| | 0/1 [16:38<?, ?it/s]\r\nPushing dataset shards to the dataset hub: 7%|███████████▎ | 71/950 [44:37<9:12:28, 37.71s/it]\r\nError while uploading 'data/train-00071-of-00950-72bab6e5cb223aee.parquet' to the Hub.\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.18ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.94s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.36ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.67s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.57ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.16s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.68ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:09<00:00, 9.63s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.36ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.67s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.37ba/s]\r\nUpload 1 LFS files: 0%| | 0/1 [16:39<?, ?it/s]\r\nPushing dataset shards to the dataset hub: 8%|████████████ | 76/950 [46:21<8:53:08, 36.60s/it]\r\nError while uploading 'data/train-00076-of-00950-b90e4e3b433db179.parquet' to the Hub.\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.21ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:25<00:00, 25.40s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:01<00:00, 1.56ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.40s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.49ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:23<00:00, 23.53s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.27ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.25s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.42ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:11<00:00, 11.03s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.39ba/s]\r\nUpload 1 LFS files: 0%| | 0/1 [16:39<?, ?it/s]\r\nPushing dataset shards to the dataset hub: 9%|████████████▊ | 81/950 [48:30<8:40:22, 35.93s/it]\r\nError while uploading 'data/train-00081-of-00950-84b0450a1df093a9.parquet' to the Hub.\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.18ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:11<00:00, 11.65s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:01<00:00, 1.92ba/s]\r\nUpload 1 LFS files: 0%| | 0/1 [16:38<?, ?it/s]\r\nPushing dataset shards to the dataset hub: 9%|█████████████ | 82/950 [48:55<8:37:57, 35.80s/it]\r\nError while uploading 'data/train-00082-of-00950-0a1f52da35653e08.parquet' to the Hub.\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.31ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:26<00:00, 26.29s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.42ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.57s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.64ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:10<00:00, 10.35s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.64ba/s]\r\nUpload 1 LFS files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:11<00:00, 11.74s/it]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.31ba/s]\r\nUpload 1 LFS files: 0%| | 0/1 [16:40<?, ?it/s]\r\nPushing dataset shards to the dataset hub: 9%|█████████████▋ | 86/950 [50:48<8:30:25, 35.45s/it]\r\nError while uploading 'data/train-00086-of-00950-e1cc80dd17191b20.parquet' to the Hub.\r\n```\r\n\r\nI have a while loop that forces retries, but it seems that the progress itself is randomly getting lost as well. Any ideas on how to improve this? It has been blocking me for way too long. \r\n\r\nShould I build the parquet manually and then push manually as well? If I do things manually, how can I ensure my dataset works properly with \"stream=True\"? \r\n\r\nThank you for your help and time. ",
"> @lhoestq It's a bit odd that when uploading a dataset, one set at a time \"train\", \"val\", \"test\", the push_to_hub function overwrites the readme and removes differently named sets from previous commits. i.e., you push \"val\", all is well. Then you push \"test\", and the \"val\" entry disappears from the readme, while the data remain intact.\r\n\r\nHmm this shouldn't happen. What code did you run exactly ? Using which version of `datasets` ?",
"> I have a while loop that forces retries, but it seems that the progress itself is randomly getting lost as well. Any ideas on how to improve this? It has been blocking me for way too long.\r\n\r\nCould you also print the cause of the error (`e.__cause__`) ? Or show the full stack trace when the error happens ?\r\nThis would give more details about why it failed and would help investigate.",
"> Should I build the parquet manually and then push manually as well? If I do things manually, how can I ensure my dataset works properly with \"stream=True\"?\r\n\r\nParquet is supported out of the box ^^\r\n\r\nIf you want to make sure it works as expected you can try locally first:\r\n```python\r\nds = load_dataset(\"path/to/local\", streaming=True)\r\n```",
"@lhoestq @AntreasAntoniou I transferred this issue to the `datasets` repository as the questions and answers are more related to this repo. Hope it can help other users find the bug and fixes more easily (like updating [tqdm](https://github.com/huggingface/datasets/issues/5990#issuecomment-1607120204) and [pyarrow](https://github.com/huggingface/datasets/issues/5990#issuecomment-1607120278) or [setting a lower `max_shard_size`](https://github.com/huggingface/datasets/issues/5990#issuecomment-1607120328)).\r\n\r\n~For the initial \"pushing large dataset consistently hangs\"-issue, I still think it's best to try to `save_to_disk` first and then upload it manually/with a script (see [upload_folder](https://huggingface.co/docs/huggingface_hub/guides/upload#upload-a-folder)). It's not the most satisfying solution but at least it would confirm from where the problem comes from.~\r\n\r\n**EDIT:** removed suggestion about saving to disk first (see https://github.com/huggingface/datasets/issues/5990#issuecomment-1607186914).",
"> @lhoestq @AntreasAntoniou I transferred this issue to the datasets repository as the questions and answers are more related to this repo. Hope it can help other users find the bug and fixes more easily (like updating https://github.com/huggingface/datasets/issues/5990#issuecomment-1607120204 and https://github.com/huggingface/datasets/issues/5990#issuecomment-1607120278 or https://github.com/huggingface/datasets/issues/5990#issuecomment-1607120328).\r\n\r\nthanks :)\r\n\r\n> For the initial \"pushing large dataset consistently hangs\"-issue, I still think it's best to try to save_to_disk first and then upload it manually/with a script (see [upload_folder](https://huggingface.co/docs/huggingface_hub/guides/upload#upload-a-folder)). It's not the most satisfying solution but at least it would confirm from where the problem comes from.\r\n\r\nAs I've already said in other discussions, I would not recommend pushing files saved with `save_to_disk` to the Hub but save to parquet shards and upload them instead. The Hub does not support datasets saved with `save_to_disk`, which is meant for disk only.",
"> As I've already said in other discussions, I would not recommend pushing files saved with save_to_disk to the Hub but save to parquet shards and upload them instead. The Hub does not support datasets saved with save_to_disk, which is meant for disk only.\r\n\r\nWell noted, thanks. That part was not clear to me :)",
"Sorry for not replying in a few days, I was on leave. :) \r\n\r\nSo, here are more information as to the error that causes some of the delay\r\n\r\n```bash\r\nPushing Antreas/TALI-tiny to hub\r\nAttempting to push to hub\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:24<00:00, 4.06s/ba]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:24<00:00, 4.15s/ba]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:26<00:00, 4.45s/ba]\r\n/opt/conda/envs/main/lib/python3.10/site-packages/huggingface_hub/lfs.py:310: UserWarning: hf_transfer is enabled but does not support uploading from bytes or BinaryIO, falling back to regular upload\r\n warnings.warn(\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:25<00:00, 4.26s/ba]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:27<00:00, 4.58s/ba]\r\nCreating parquet from Arrow format: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:24<00:00, 4.10s/ba]\r\nPushing dataset shards to the dataset hub: 22%|████████████████████████▎ | 5/23 [52:23<3:08:37, 628.74s/it]\r\nException: Error while uploading 'data/train-00005-of-00023-e224d901fd65e062.parquet' to the Hub., with stacktrace: <traceback object at 0x7f745458d0c0>, and type: <class 'RuntimeError'>, and \r\ncause: HTTPSConnectionPool(host='s3.us-east-1.amazonaws.com', port=443): Max retries exceeded with url: \r\n/lfs.huggingface.co/repos/7c/d3/7cd385d9324302dc13e3986331d72d9be6fa0174c63dcfe0e08cd474f7f1e8b7/3415166ae28c0beccbbc692f38742b8dea2c197f5c805321104e888d21d7eb90?X-Amz-Algorithm=AWS4-HMAC-SHA256\r\n&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=AKIA4N7VTDGO27GPWFUO%2F20230627%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20230627T003349Z&X-Amz-Expires=86400&X-Amz-Signature=5a12ff96f2\r\n91f644134170992a6628e5f3c4e7b2e7fc3e940b4378fe11ae5390&X-Amz-SignedHeaders=host&partNumber=1&uploadId=JSsK8r63XSF.VlKQx3Vf8OW4DEVp5YIIY7LPnuapNIegsxs5EHgM1p4u0.Nn6_wlPlQnvxm8HKMxZhczKE9KB74t0etB\r\noLcxqBIvsgey3uXBTZMAEGwU6y7CDUADiEIO&x-id=UploadPart (Caused by SSLError(SSLEOFError(8, 'EOF occurred in violation of protocol (_ssl.c:2426)')))\r\nPush failed, retrying\r\nAttempting to push to hub\r\nPushing split train to the Hub.\r\n```\r\n\r\nOne issue is that the uploading does not continue from the chunk it failed off. It often continues from a very old chunk. e.g. if it failed on chunk 192/250, it will continue from say 53/250, and this behaviour appears almost random. ",
"Are you using a proxy of some sort ?",
"I am using a kubernetes cluster built into a university VPN. ",
"So, other than the random connection drops here and there, any idea why the progress does not continue where it left off?\r\n\r\n```bash\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 10.79ba/s]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 13.65ba/s]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 13.39ba/s]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 13.04ba/s]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 13.52ba/s]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 12.28ba/s]\r\nPushing dataset shards to the dataset hub: 20%|██████████████████████ | 75/381 [1:34:39<6:26:11, 75.72s/it]\r\nException: Error while uploading 'data/train-00075-of-00381-1614bc251b778766.parquet' to the Hub., with stacktrace: <traceback object at 0x7fab6d9a4980>, and type: <class 'RuntimeError'>, and \r\ncause: HTTPSConnectionPool(host='s3.us-east-1.amazonaws.com', port=443): Max retries exceeded with url: \r\n/lfs.huggingface.co/repos/3b/31/3b311464573d8d63b137fcd5b40af1e7a5b1306843c88e80372d0117157504e5/ed8dae933fb79ae1ef5fb1f698f5125d3e1c02977ac69438631f152bb3bfdd1e?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-\r\nAmz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=AKIA4N7VTDGO27GPWFUO%2F20230629%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20230629T053004Z&X-Amz-Expires=86400&X-Amz-Signature=da2b26270edfd6d0\r\nd069c015a5a432031107a8664c3f0917717e5e40c688183c&X-Amz-SignedHeaders=host&partNumber=1&uploadId=2erWGHTh3ICqBLU_QvHfnygZ2tkMWbL0rEqpJdYohCKHUHnfwMjvoBIg0TI_KSGn4rSKxUxOyqSIzFUFSRSzixZeLeneaXJOw.Qx8\r\nzLKSV5xV7HRQDj4RBesNve6cSoo&x-id=UploadPart (Caused by SSLError(SSLEOFError(8, 'EOF occurred in violation of protocol (_ssl.c:2426)')))\r\nPush failed, retrying\r\nAttempting to push to hub\r\nPushing split train to the Hub.\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 12.09ba/s]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 11.51ba/s]\r\nCreating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:02<00:00, 10.77ba/s]\r\nPushing dataset shards to the dataset hub: 20%|██████████████████████▋ | 77/381 [1:32:50<6:06:34, 72.35s/it]\r\nException: Error while uploading 'data/train-00077-of-00381-368b2327a9908aab.parquet' to the Hub., with stacktrace: <traceback object at 0x7fab45b27f80>, and type: <class 'RuntimeError'>, and \r\ncause: HTTPSConnectionPool(host='s3.us-east-1.amazonaws.com', port=443): Max retries exceeded with url: \r\n/lfs.huggingface.co/repos/3b/31/3b311464573d8d63b137fcd5b40af1e7a5b1306843c88e80372d0117157504e5/9462ff2c5e61283b53b091984a22de2f41a2f6e37b681171e2eca4a998f979cb?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-\r\nAmz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=AKIA4N7VTDGO27GPWFUO%2F20230629%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20230629T070510Z&X-Amz-Expires=86400&X-Amz-Signature=9ab8487b93d443cd\r\n21f05476405855d46051a0771b4986bbb20f770ded21b1a4&X-Amz-SignedHeaders=host&partNumber=1&uploadId=UiHX1B.DcoAO2QmIHpWpCuNPwhXU_o1dsTkTGPqZt1P51o9k0yz.EsFD9eKpQMwgAST3jOatRG78I_JWRBeLBDYYVNp8r0TpIdeSg\r\neUg8uwPZOCPw9y5mWOw8MWJrnBo&x-id=UploadPart (Caused by SSLError(SSLEOFError(8, 'EOF occurred in violation of protocol (_ssl.c:2426)')))\r\nPush failed, retrying\r\nAttempting to push to hub\r\nPushing split train to the Hub.\r\nPushing dataset shards to the dataset hub: 8%|████████▋ | 29/381 [27:39<5:50:03, 59.67s/it]\r\nMap: 36%|████████████████████████████████████████████████████ | 1000/2764 [00:35<00:34, 51.63 examples/Map: 72%|████████████████████████████████████████████████████████████████████████████████████████████████████████▏ | 2000/2764 [00:40<00:15, 49.06 examples/Map: 72%|████████████████████████████████████████████████████████████████████████████████████████████████████████▏ | 2000/2764 [00:55<00:15, 49.06 examples/Map: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2764/2764 [00:56<00:00, 48.82 examples/Pushing dataset shards to the dataset hub: 8%|████████▉ | 30/381 [28:35<5:43:03, 58.64s/iPushing dataset shards to the dataset hub: 8%|█████████▎ | 31/381 [29:40<5:52:18, 60.40s/iPushing dataset shards to the dataset hub: 8%|█████████▌ | 32/381 [30:46<6:02:20, 62.29s/it] \r\nMap: 36%|███████████████████████████████████████████████████▎ \r\n```\r\n\r\nThis is actually the issue that wastes the most time for me, and I need it fixed. Please advice on how I can go about it.\r\n\r\nNotice how the progress goes from \r\n| 77/381 to 30/381",
"If the any shard is missing on the Hub, it will re-upload it. It looks like the 30th shard was missing on the Hub in your case. \r\n\r\nIt also means that the other files up to the 77th that were successfully uploaded won't be uploaded again.\r\n\r\ncc @mariosasko who might know better"
] | ### Describe the bug
Once I have locally built a large dataset that I want to push to hub, I use the recommended approach of .push_to_hub to get the dataset on the hub, and after pushing a few shards, it consistently hangs. This has happened over 40 times over the past week, and despite my best efforts to try and catch this happening and kill a process and restart, it seems to be extremely time wasting -- so I came to you to report this and to seek help.
I already tried installing hf_transfer, but it doesn't support Byte file uploads so I uninstalled it.
### Reproduction
```python
import multiprocessing as mp
import pathlib
from math import ceil
import datasets
import numpy as np
from tqdm.auto import tqdm
from tali.data.data import select_subtitles_between_timestamps
from tali.utils import load_json
tali_dataset_dir = "/data/"
if __name__ == "__main__":
full_dataset = datasets.load_dataset(
"Antreas/TALI", num_proc=mp.cpu_count(), cache_dir=tali_dataset_dir
)
def data_generator(set_name, percentage: float = 1.0):
dataset = full_dataset[set_name]
for item in tqdm(dataset):
video_list = item["youtube_content_video"]
video_list = np.random.choice(
video_list, int(ceil(len(video_list) * percentage))
)
if len(video_list) == 0:
continue
captions = item["youtube_subtitle_text"]
captions = select_subtitles_between_timestamps(
subtitle_dict=load_json(
captions.replace(
"/data/",
tali_dataset_dir,
)
),
starting_timestamp=0,
ending_timestamp=100000000,
)
for video_path in video_list:
temp_path = video_path.replace("/data/", tali_dataset_dir)
video_path_actual: pathlib.Path = pathlib.Path(temp_path)
if video_path_actual.exists():
item["youtube_content_video"] = open(video_path_actual, "rb").read()
item["youtube_subtitle_text"] = captions
yield item
train_generator = lambda: data_generator("train", percentage=0.1)
val_generator = lambda: data_generator("val")
test_generator = lambda: data_generator("test")
train_data = datasets.Dataset.from_generator(
train_generator,
num_proc=mp.cpu_count(),
writer_batch_size=5000,
cache_dir=tali_dataset_dir,
)
val_data = datasets.Dataset.from_generator(
val_generator,
writer_batch_size=5000,
num_proc=mp.cpu_count(),
cache_dir=tali_dataset_dir,
)
test_data = datasets.Dataset.from_generator(
test_generator,
writer_batch_size=5000,
num_proc=mp.cpu_count(),
cache_dir=tali_dataset_dir,
)
dataset = datasets.DatasetDict(
{
"train": train_data,
"val": val_data,
"test": test_data,
}
)
succesful_competion = False
while not succesful_competion:
try:
dataset.push_to_hub(repo_id="Antreas/TALI-small", max_shard_size="5GB")
succesful_competion = True
except Exception as e:
print(e)
```
### Logs
```shell
Pushing dataset shards to the dataset hub: 33%|██████████████████████████████████████▎ | 7/21 [24:33<49:06, 210.45s/it]
Error while uploading 'data/val-00007-of-00021-6b216a984af1a4c8.parquet' to the Hub.
Pushing split train to the Hub.
Resuming upload of the dataset shards.
Pushing dataset shards to the dataset hub: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 46/46 [42:10<00:00, 55.01s/it]
Pushing split val to the Hub.
Resuming upload of the dataset shards.
Creating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:01<00:00, 1.55ba/s]
Upload 1 LFS files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:23<00:00, 23.51s/it]
Creating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.39ba/s]
Upload 1 LFS files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:30<00:00, 30.19s/it]
Creating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.28ba/s]
Upload 1 LFS files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:24<00:00, 24.08s/it]
Creating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.42ba/s]
Upload 1 LFS files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:23<00:00, 23.97s/it]
Creating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.49ba/s]
Creating parquet from Arrow format: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.54ba/s^
Upload 1 LFS files: 0%| | 0/1 [04:42<?, ?it/s]
Pushing dataset shards to the dataset hub: 52%|████████████████████████████████████████████████████████████▏ | 11/21 [17:23<15:48, 94.82s/it]
That's where it got stuck
```
### System info
```shell
- huggingface_hub version: 0.15.1
- Platform: Linux-5.4.0-147-generic-x86_64-with-glibc2.35
- Python version: 3.10.11
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Token path ?: /root/.cache/huggingface/token
- Has saved token ?: True
- Who am I ?: Antreas
- Configured git credential helpers: store
- FastAI: N/A
- Tensorflow: N/A
- Torch: 2.1.0.dev20230606+cu121
- Jinja2: 3.1.2
- Graphviz: N/A
- Pydot: N/A
- Pillow: 9.5.0
- hf_transfer: N/A
- gradio: N/A
- numpy: 1.24.3
- ENDPOINT: https://huggingface.co
- HUGGINGFACE_HUB_CACHE: /root/.cache/huggingface/hub
- HUGGINGFACE_ASSETS_CACHE: /root/.cache/huggingface/assets
- HF_TOKEN_PATH: /root/.cache/huggingface/token
- HF_HUB_OFFLINE: False
- HF_HUB_DISABLE_TELEMETRY: False
- HF_HUB_DISABLE_PROGRESS_BARS: None
- HF_HUB_DISABLE_SYMLINKS_WARNING: False
- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False
- HF_HUB_DISABLE_IMPLICIT_TOKEN: False
- HF_HUB_ENABLE_HF_TRANSFER: False
```
| 5,990 |
https://github.com/huggingface/datasets/issues/5939 | . | [] | null | 5,939 |
https://github.com/huggingface/datasets/issues/5936 | Sequence of array not supported for most dtype | [
"Related, `float16` is the only dtype not supported by `Array2D` (probably by every `ArrayND`):\r\n\r\n```python\r\nfrom datasets import Array2D, Features, Dataset\r\n\r\nimport numpy as np\r\n\r\nfor dtype in [\r\n \"bool\", # ok\r\n \"int8\", # ok\r\n \"int16\", # ok\r\n \"int32\", # ok\r\n \"int64\", # ok\r\n \"uint8\", # ok\r\n \"uint16\", # ok\r\n \"uint32\", # ok\r\n \"uint64\", # ok\r\n \"float16\", # failed\r\n \"float32\", # ok\r\n \"float64\", # ok\r\n]:\r\n features = Features({\"foo\": Array2D(dtype=dtype, shape=(3, 4))})\r\n array = np.zeros((3, 4), dtype=dtype)\r\n try:\r\n dataset = Dataset.from_dict({\"foo\": [array]}, features=features)\r\n except Exception as e:\r\n print(f\"Failed for dtype={dtype}\")\r\n```",
"Here's something I can't explain:\r\n\r\nWhen an array is encoded in the `from_dict` method, the numpy array is converted to a list (thus losing the original dtype, which is transfromed to the nearest builtin Python type)\r\n\r\nhttps://github.com/huggingface/datasets/blob/6ee61e6e695b1df9f232d47faf3a5e2b30b33737/src/datasets/features/features.py#L524-L525\r\n\r\nHowever, later on, this same data is written to memory, and it seems authorized that the data is an array (or in this case, a list of arrays). \r\n\r\nhttps://github.com/huggingface/datasets/blob/6ee61e6e695b1df9f232d47faf3a5e2b30b33737/src/datasets/arrow_writer.py#L185-L186\r\n\r\nSo the question is: why convert it to a Python list? This seems to be quite expensive both in terms of write time (all data is copied) and memory (e.g., an int8 is converted to an int64).\r\n\r\nFinally, if I try to remove this step, it solves all the previous problems, and it seems to me that it doesn't break anything (the CI passes without problem).",
"Arrow only support 1d numpy arrays, so we convert multidim arrays to lists of 1s arrays (and keep the dtype).\r\n\r\nThough you noticed that it's concerting to lists and lose the dtype. If it's the case then it's a bug.",
"Ok the conversion to list shouldn't be there indeed ! Could you open a PR to remove it ?"
] | ### Describe the bug
Create a dataset composed of sequence of array fails for most dtypes (see code below).
### Steps to reproduce the bug
```python
from datasets import Sequence, Array2D, Features, Dataset
import numpy as np
for dtype in [
"bool", # ok
"int8", # failed
"int16", # failed
"int32", # failed
"int64", # ok
"uint8", # failed
"uint16", # failed
"uint32", # failed
"uint64", # failed
"float16", # failed
"float32", # failed
"float64", # ok
]:
features = Features({"foo": Sequence(Array2D(dtype=dtype, shape=(2, 2)))})
sequence = [
[[1.0, 2.0], [3.0, 4.0]],
[[5.0, 6.0], [7.0, 8.0]],
]
array = np.array(sequence, dtype=dtype)
try:
dataset = Dataset.from_dict({"foo": [array]}, features=features)
except Exception as e:
print(f"Failed for dtype={dtype}")
```
Traceback for `dtype="int8"`:
```
Traceback (most recent call last):
File "/home/qgallouedec/datasets/a.py", line 29, in <module>
raise e
File "/home/qgallouedec/datasets/a.py", line 26, in <module>
dataset = Dataset.from_dict({"foo": [array]}, features=features)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/arrow_dataset.py", line 899, in from_dict
pa_table = InMemoryTable.from_pydict(mapping=mapping)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/table.py", line 799, in from_pydict
return cls(pa.Table.from_pydict(*args, **kwargs))
File "pyarrow/table.pxi", line 3725, in pyarrow.lib.Table.from_pydict
File "pyarrow/table.pxi", line 5254, in pyarrow.lib._from_pydict
File "pyarrow/array.pxi", line 350, in pyarrow.lib.asarray
File "pyarrow/array.pxi", line 236, in pyarrow.lib.array
File "pyarrow/array.pxi", line 110, in pyarrow.lib._handle_arrow_array_protocol
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/arrow_writer.py", line 204, in __arrow_array__
out = cast_array_to_feature(out, type, allow_number_to_str=not self.trying_type)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/table.py", line 1833, in wrapper
return func(array, *args, **kwargs)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/table.py", line 2091, in cast_array_to_feature
casted_values = _c(array.values, feature.feature)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/table.py", line 1833, in wrapper
return func(array, *args, **kwargs)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/table.py", line 2139, in cast_array_to_feature
return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/table.py", line 1833, in wrapper
return func(array, *args, **kwargs)
File "/home/qgallouedec/env/lib/python3.10/site-packages/datasets/table.py", line 1967, in array_cast
return pa_type.wrap_array(array)
File "pyarrow/types.pxi", line 879, in pyarrow.lib.BaseExtensionType.wrap_array
TypeError: Incompatible storage type for extension<arrow.py_extension_type<Array2DExtensionType>>: expected list<item: list<item: int8>>, got list<item: list<item: int64>>
```
### Expected behavior
Not to fail.
### Environment info
- Python 3.10.6
- datasets: master branch
- Numpy: 1.23.4 | 5,936 |
https://github.com/huggingface/datasets/issues/5931 | `datasets.map` not reusing cached copy by default | [
"This can happen when a map transform cannot be hashed deterministically (e.g., an object referenced by the transform changes its state after the first call - an issue with fast tokenizers). The solution is to provide `cache_file_name` in the `map` call to check this file for the cached result instead of relying on the default caching mechanism."
] | ### Describe the bug
When I load the dataset from local directory, it's cached copy is picked up after first time. However, for `map` operation, the operation is applied again and cached copy is not picked up. Is there any way to pick cached copy instead of processing it again? The only solution I could think of was to use `save_to_disk` after my last transform and then use that in my DataLoader pipeline. Are there any other solutions for the same?
One more thing, my dataset is occupying 6GB storage memory after I use `map`, is there any way I can reduce that memory usage?
### Steps to reproduce the bug
```
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = next(iter(self.raw_datasets.values())).features["audio"].sampling_rate
if dataset_sampling_rate != self.feature_extractor.sampling_rate:
self.raw_datasets = self.raw_datasets.cast_column(
"audio", datasets.features.Audio(sampling_rate=self.feature_extractor.sampling_rate)
)
vectorized_datasets = self.raw_datasets.map(
self.prepare_dataset,
remove_columns=next(iter(self.raw_datasets.values())).column_names,
num_proc=self.num_workers,
desc="preprocess datasets",
)
# filter data that is longer than max_input_length
self.vectorized_datasets = vectorized_datasets.filter(
self.is_audio_in_length_range,
num_proc=self.num_workers,
input_columns=["input_length"],
)
def prepare_dataset(self, batch):
# load audio
sample = batch["audio"]
inputs = self.feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
batch["labels"] = self.tokenizer(batch["target_text"]).input_ids
return batch
```
### Expected behavior
`map` to use cached copy and if possible an alternative technique to reduce memory usage after using `map`
### Environment info
- `datasets` version: 2.12.0
- Platform: Linux-3.10.0-1160.71.1.el7.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.16
- Huggingface_hub version: 0.15.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.2
| 5,931 |
https://github.com/huggingface/datasets/issues/5930 | loading private custom dataset script - authentication error | [
"This issue seems to have been resolved, so I'm closing it."
] | ### Describe the bug
Train model with my custom dataset stored in HuggingFace and loaded with the loading script requires authentication but I am not sure how ?
I am logged in in the terminal, in the browser. I receive this error:
/python3.8/site-packages/datasets/utils/file_utils.py", line 566, in get_from_cache
raise ConnectionError(f"Couldn't reach {url} ({repr(head_error)})")
ConnectionError: Couldn't reach https://huggingface.co/datasets/fkov/s/blob/main/data/s/train/labels `(ConnectionError('Unauthorized for URL `https://huggingface.co/datasets/fkov/s/blob/main/data/s/train/labels. Please use the parameter `**`use_auth_token=True`**` after logging in with `**`huggingface-cli login`**`'))
when I added: `use_auth_token=True` and logged in via terminal then I received error:
or the same error in different format:
raise ConnectionError(f"`Couldn't reach {url} (error {response.status_code}`)")
ConnectionError: Couldn't reach https://huggingface.co/datasets/fkov/s/blob/main/data/s/train/labels (`error 401`)
### Steps to reproduce the bug
1. cloned transformers library locally:
https://huggingface.co/docs/transformers/v4.15.0/examples :
> git clone https://github.com/huggingface/transformers
> cd transformers
> pip install .
> cd /transformers/examples/pytorch/audio-classification
> pip install -r requirements.txt
2. created **loading script**
> https://huggingface.co/docs/datasets/dataset_script added next to dataset:
3. uploaded **private custom dataset** with loading script to HuggingFace
> https://huggingface.co/docs/datasets/dataset_script
4. added dataset loading script to **local directory** in the above cloned transformers library:
> cd /transformers/examples/pytorch/audio-classification
5. logged in to HuggingFace on local terminal with :
> **huggingface-cli login**
6. run the model with the custom dataset stored on HuggingFace with code: https://github.com/huggingface/transformers/blob/main/examples/pytorch/audio-classification/README.md
cd /transformers/examples/pytorch/audio-classification
> python run_audio_classification.py \
> --model_name_or_path facebook/wav2vec2-base \
> --output_dir l/users/flck/outputs/wav2vec2-base-s \
> --overwrite_output_dir \
> --dataset_name s \
> --dataset_config_name s \
> --remove_unused_columns False \
> --do_train \
> --do_eval \
> --fp16 \
> --learning_rate 3e-5 \
> --max_length_seconds 1 \
> --attention_mask False \
> --warmup_ratio 0.1 \
> --num_train_epochs 5 \
> --per_device_train_batch_size 32 \
> --gradient_accumulation_steps 4 \
> --per_device_eval_batch_size 32 \
> --dataloader_num_workers 4 \
> --logging_strategy steps \
> --logging_steps 10 \
> --evaluation_strategy epoch \
> --save_strategy epoch \
> --load_best_model_at_end True \
> --metric_for_best_model accuracy \
> --save_total_limit 3 \
> --seed 0 \
> --push_to_hub \
> **--use_auth_token=True**
### Expected behavior
Be able to train a model the https://github.com/huggingface/transformers/blob/main/examples/pytorch/audio-classification/ run_audio_classification.py with private custom dataset stored on HuggingFace.
### Environment info
- datasets version: 2.12.0
- `transformers` version: 4.30.0.dev0
- Platform: Linux-5.4.204-ql-generic-12.0-19-x86_64-with-glibc2.17
- Python version: 3.8.12
- Huggingface_hub version: 0.15.1
- Safetensors version: 0.3.1
- PyTorch version (GPU?): 2.0.1+cu117 (True)
Versions of relevant libraries:
[pip3] numpy==1.24.3
[pip3] torch==2.0.1
[pip3] torchaudio==2.0.2
[conda] numpy 1.24.3 pypi_0 pypi
[conda] torch 2.0.1 pypi_0 pypi
[conda] torchaudio 2.0.2 pypi_0 pypi
| 5,930 |
https://github.com/huggingface/datasets/issues/5929 | Importing PyTorch reduces multiprocessing performance for map | [
"Hi! The times match when I run this code locally or on Colab.\r\n\r\nAlso, we use `multiprocess`, not `multiprocessing`, for parallelization, and torch's `__init__.py` (executed on `import torch` ) slightly modifies the latter.",
"Hey Mariosasko,\r\n\r\nThanks for looking into it. We further did some investigations after your comment and figured out it's only affecting some hardware/software configurations with the `pytorch` installation of `conda-forge`. Based on this we found the following issue in PyTorch: https://github.com/pytorch/pytorch/issues/102269 with a quick fix for now.\r\n\r\nSince it seems to be a deeper issue with forking processes, the difference between`multiprocess` and `multiprocessing` didn't make a difference.\r\n\r\nClosing this, since the issue comes from `pytorch` not `dataset`. \r\n"
] | ### Describe the bug
I noticed that the performance of my dataset preprocessing with `map(...,num_proc=32)` decreases when PyTorch is imported.
### Steps to reproduce the bug
I created two example scripts to reproduce this behavior:
```
import datasets
datasets.disable_caching()
from datasets import Dataset
import time
PROC=32
if __name__ == "__main__":
dataset = [True] * 10000000
dataset = Dataset.from_dict({'train': dataset})
start = time.time()
dataset.map(lambda x: x, num_proc=PROC)
end = time.time()
print(end - start)
```
Takes around 4 seconds on my machine.
While the same code, but with an `import torch`:
```
import datasets
datasets.disable_caching()
from datasets import Dataset
import time
import torch
PROC=32
if __name__ == "__main__":
dataset = [True] * 10000000
dataset = Dataset.from_dict({'train': dataset})
start = time.time()
dataset.map(lambda x: x, num_proc=PROC)
end = time.time()
print(end - start)
```
takes around 22 seconds.
### Expected behavior
I would expect that the import of torch to not have such a significant effect on the performance of map using multiprocessing.
### Environment info
- `datasets` version: 2.12.0
- Platform: Linux-5.15.0-56-generic-x86_64-with-glibc2.35
- Python version: 3.11.3
- Huggingface_hub version: 0.15.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.2
- torch: 2.0.1 | 5,929 |
https://github.com/huggingface/datasets/issues/5927 | `IndexError` when indexing `Sequence` of `Array2D` with `None` values | [
"Easy fix would be to add:\r\n\r\n```python\r\nnull_indices -= np.arange(len(null_indices))\r\n```\r\n\r\nbefore L279, but I'm not sure it's the most intuitive way to fix it.",
"Same issue here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/7fcbe5b1575c8d162b65b9397b3dfda995a4e048/src/datasets/features/features.py#L1398\r\n\r\nFixed in #5948 "
] | ### Describe the bug
Having `None` values in a `Sequence` of `ArrayND` fails.
### Steps to reproduce the bug
```python
from datasets import Array2D, Dataset, Features, Sequence
data = [
[
[[0]],
None,
None,
]
]
feature = Sequence(Array2D((1, 1), dtype="int64"))
dataset = Dataset.from_dict({"a": data}, features=Features({"a": feature}))
dataset[0] # error raised only when indexing
```
```
Traceback (most recent call last):
File "/Users/quentingallouedec/gia/c.py", line 13, in <module>
dataset[0] # error raised only when indexing
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/arrow_dataset.py", line 2658, in __getitem__
return self._getitem(key)
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/arrow_dataset.py", line 2643, in _getitem
formatted_output = format_table(
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 634, in format_table
return formatter(pa_table, query_type=query_type)
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 406, in __call__
return self.format_row(pa_table)
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 441, in format_row
row = self.python_arrow_extractor().extract_row(pa_table)
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 144, in extract_row
return _unnest(pa_table.to_pydict())
File "pyarrow/table.pxi", line 4146, in pyarrow.lib.Table.to_pydict
File "pyarrow/table.pxi", line 1312, in pyarrow.lib.ChunkedArray.to_pylist
File "pyarrow/array.pxi", line 1521, in pyarrow.lib.Array.to_pylist
File "pyarrow/scalar.pxi", line 675, in pyarrow.lib.ListScalar.as_py
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/features/features.py", line 760, in to_pylist
return self.to_numpy(zero_copy_only=zero_copy_only).tolist()
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/datasets/features/features.py", line 725, in to_numpy
numpy_arr = np.insert(numpy_arr.astype(np.float64), null_indices, np.nan, axis=0)
File "<__array_function__ internals>", line 200, in insert
File "/Users/quentingallouedec/gia/env/lib/python3.10/site-packages/numpy/lib/function_base.py", line 5426, in insert
old_mask[indices] = False
IndexError: index 3 is out of bounds for axis 0 with size 3
```
AFAIK, the problem only occurs when you use a `Sequence` of `ArrayND`.
I strongly suspect that the problem comes from this line, or `np.insert` is misused:
https://github.com/huggingface/datasets/blob/02ee418831aba68d0be93227bce8b3f42ef8980f/src/datasets/features/features.py#L729
To put t simply, you want something that do that:
```python
import numpy as np
numpy_arr = np.zeros((1, 1, 1))
null_indices = np.array([1, 2])
np.insert(numpy_arr, null_indices, np.nan, axis=0)
# raise an error, instead of outputting
# array([[[ 0.]],
# [[nan]],
# [[nan]]])
```
### Expected behavior
The previous code should not raise an error.
### Environment info
- Python 3.10.11
- datasets 2.10.0
- pyarrow 12.0.0 | 5,927 |
https://github.com/huggingface/datasets/issues/5926 | Uncaught exception when generating the splits from a dataset that miss data | [
"Thanks for reporting, @severo.\r\n\r\nThis is a known issue with `fsspec`:\r\n- #5862\r\n- https://github.com/fsspec/filesystem_spec/issues/1265"
] | ### Describe the bug
Dataset https://huggingface.co/datasets/blog_authorship_corpus has an issue with its hosting platform, since https://drive.google.com/u/0/uc?id=1cGy4RNDV87ZHEXbiozABr9gsSrZpPaPz&export=download returns 404 error.
But when trying to generate the split names, we get an exception which is now correctly caught.
Seen originally in https://github.com/huggingface/datasets-server/blob/adbdcd6710ffed4e2eb2e4cd905b5e0dff530a15/services/worker/src/worker/job_runners/config/parquet_and_info.py#L435
### Steps to reproduce the bug
```python
>>> from datasets import StreamingDownloadManager, load_dataset_builder
>>> builder = load_dataset_builder(path="blog_authorship_corpus")
Downloading builder script: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5.60k/5.60k [00:00<00:00, 23.1MB/s]
Downloading metadata: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2.81k/2.81k [00:00<00:00, 14.7MB/s]
Downloading readme: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7.30k/7.30k [00:00<00:00, 30.8MB/s]
>>> dl_manager = StreamingDownloadManager(base_path=builder.base_path)
>>> builder._split_generators(dl_manager)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/slesage/.cache/huggingface/modules/datasets_modules/datasets/blog_authorship_corpus/6f5d78241afd8313111956f877a57db7a0e9fc6718255dc85df0928197feb683/blog_authorship_corpus.py", line 79, in _split_generators
data = dl_manager.download_and_extract(_DATA_URL)
File "/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py", line 1087, in download_and_extract
return self.extract(self.download(url_or_urls))
File "/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py", line 1039, in extract
urlpaths = map_nested(self._extract, url_or_urls, map_tuple=True)
File "/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 435, in map_nested
return function(data_struct)
File "/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py", line 1044, in _extract
protocol = _get_extraction_protocol(urlpath, use_auth_token=self.download_config.use_auth_token)
File "/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py", line 433, in _get_extraction_protocol
with fsspec.open(urlpath, **kwargs) as f:
File "/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/fsspec/core.py", line 439, in open
return open_files(
File "/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/fsspec/core.py", line 194, in __getitem__
out = super().__getitem__(item)
IndexError: list index out of range
```
### Expected behavior
We should have an Exception raised by the datasets library.
### Environment info
- `datasets` version: 2.12.0
- Platform: Linux-5.19.0-1026-aws-x86_64-with-glibc2.35
- Python version: 3.9.15
- Huggingface_hub version: 0.15.1
- PyArrow version: 11.0.0
- Pandas version: 2.0.2 | 5,926 |
https://github.com/huggingface/datasets/issues/5925 | Breaking API change in datasets.list_datasets caused by change in HfApi.list_datasets | [] | ### Describe the bug
Hi all,
after an update of the `datasets` library, we observer crashes in our code. We relied on `datasets.list_datasets` returning a `list`. Now, after the API of the HfApi.list_datasets was changed and it returns a `list` instead of an `Iterable`, the `datasets.list_datasets` now sometimes returns a `list` and somesimes an `Iterable`.
It would be helpful to indicate that by the return type of the `datasets.list_datasets` function.
Thanks,
Martin
### Steps to reproduce the bug
Here, the code crashed after we updated the `datasets` library:
```python
# list_datasets no longer returns a list, which leads to an error when one tries to slice it
for datasets.list_datasets(with_details=True)[:limit]:
...
```
### Expected behavior
It would be helpful to indicate that by the return type of the `datasets.list_datasets` function.
### Environment info
Ubuntu 22.04
datasets 2.12.0 | 5,925 |
https://github.com/huggingface/datasets/issues/5923 | Cannot import datasets - ValueError: pyarrow.lib.IpcWriteOptions size changed, may indicate binary incompatibility | [
"Based on https://github.com/rapidsai/cudf/issues/10187, this probably means your `pyarrow` installation is not compatible with `datasets`.\r\n\r\nCan you please execute the following commands in the terminal and paste the output here?\r\n```\r\nconda list | grep arrow\r\n``` \r\n```\r\npython -c \"import pyarrow; print(pyarrow.__file__)\"\r\n```\r\n\r\n\r\n",
"> Based on [rapidsai/cudf#10187](https://github.com/rapidsai/cudf/issues/10187), this probably means your `pyarrow` installation is not compatible with `datasets`.\r\n> \r\n> Can you please execute the following commands in the terminal and paste the output here?\r\n> \r\n> ```\r\n> conda list | grep arrow\r\n> ```\r\n> \r\n> ```\r\n> python -c \"import pyarrow; print(pyarrow.__file__)\"\r\n> ```\r\n\r\n\r\nHere is the output to the first command:\r\n```\r\narrow-cpp 11.0.0 py39h7f74497_0 \r\npyarrow 12.0.0 pypi_0 pypi\r\n```\r\nand the second:\r\n```\r\n/Users/edward/opt/anaconda3/envs/cs235/lib/python3.9/site-packages/pyarrow/__init__.py\r\n```\r\nThanks!\r\n\r\n\r\n\r\n",
"after installing pytesseract 0.3.10, I got the above error. FYI ",
"RuntimeError: Failed to import transformers.trainer because of the following error (look up to see its traceback):\r\npyarrow.lib.IpcWriteOptions size changed, may indicate binary incompatibility. Expected 88 from C header, got 72 from PyObject",
"I got the same error, pyarrow 12.0.0 released May/2023 (https://pypi.org/project/pyarrow/) is not compatible, running `pip install pyarrow==11.0.0` to force install the previous version solved the problem.\r\n\r\nDo we need to update dependencies? ",
"Please note that our CI properly passes all tests with `pyarrow-12.0.0`, for Python 3.7 and Python 3.10, for Ubuntu and Windows: see for example https://github.com/huggingface/datasets/actions/runs/5157324334/jobs/9289582291",
"For conda with python3.8.16 this solved my problem! thanks!\r\n\r\n> I got the same error, pyarrow 12.0.0 released May/2023 (https://pypi.org/project/pyarrow/) is not compatible, running `pip install pyarrow==11.0.0` to force install the previous version solved the problem.\r\n> \r\n> Do we need to update dependencies? I can work on that if no one else is working on it.\r\n\r\n",
"Thanks for replying. I am not sure about those environments but it seems like pyarrow-12.0.0 does not work for conda with python 3.8.16. \r\n\r\n> Please note that our CI properly passes all tests with `pyarrow-12.0.0`, for Python 3.7 and Python 3.10, for Ubuntu and Windows: see for example https://github.com/huggingface/datasets/actions/runs/5157324334/jobs/9289582291\r\n\r\n",
"Got the same error with:\r\n\r\n```\r\narrow-cpp 11.0.0 py310h7516544_0 \r\npyarrow 12.0.0 pypi_0 pypi\r\n\r\npython 3.10.11 h7a1cb2a_2 \r\n\r\ndatasets 2.13.0 pyhd8ed1ab_0 conda-forge\r\n```",
"> I got the same error, pyarrow 12.0.0 released May/2023 (https://pypi.org/project/pyarrow/) is not compatible, running `pip install pyarrow==11.0.0` to force install the previous version solved the problem.\r\n> \r\n> Do we need to update dependencies?\r\n\r\nThis solved the issue for me as well.",
"> I got the same error, pyarrow 12.0.0 released May/2023 (https://pypi.org/project/pyarrow/) is not compatible, running `pip install pyarrow==11.0.0` to force install the previous version solved the problem.\r\n> \r\n> Do we need to update dependencies?\r\n\r\nSolved it for me also",
"> 基于 [rapidsai/cudf#10187](https://github.com/rapidsai/cudf/issues/10187)，这可能意味着您的安装与 不兼容。`pyarrow``datasets`\r\n> \r\n> 您能否在终端中执行以下命令并将输出粘贴到此处？\r\n> \r\n> ```\r\n> conda list | grep arrow\r\n> ```\r\n> \r\n> ```\r\n> python -c \"import pyarrow; print(pyarrow.__file__)\"\r\n> ```\r\n\r\narrow-cpp 11.0.0 py310h7516544_0 \r\npyarrow 12.0.1 pypi_0 pypi\r\n\r\n/root/miniconda3/lib/python3.10/site-packages/pyarrow/__init__.py",
"Got the same problem with\r\n\r\narrow-cpp 11.0.0 py310h1fc3239_0 \r\npyarrow 12.0.1 pypi_0 pypi\r\n\r\nminiforge3/envs/mlp/lib/python3.10/site-packages/pyarrow/__init__.py\r\n\r\nReverting back to pyarrow 11 solved the problem.\r\n",
"Solved with `pip install pyarrow==11.0.0`",
"I got different. Solved with\r\npip install pyarrow==12.0.1\r\npip install cchardet\r\n\r\nenv:\r\nPython 3.9.16\r\ntransformers 4.32.1",
"> I got the same error, pyarrow 12.0.0 released May/2023 (https://pypi.org/project/pyarrow/) is not compatible, running `pip install pyarrow==11.0.0` to force install the previous version solved the problem.\r\n> \r\n> Do we need to update dependencies?\r\n\r\nThis works for me as well",
"> I got different. Solved with pip install pyarrow==12.0.1 pip install cchardet\r\n> \r\n> env: Python 3.9.16 transformers 4.32.1\r\n\r\nI guess it also depends on the Python version. I got Python 3.11.5 and pyarrow==12.0.0. \r\nIt works! ",
"Hi, if this helps anyone, pip install pyarrow==11.0.0 did not work for me (I'm using Colab) but this worked: \r\n!pip install --extra-index-url=https://pypi.nvidia.com cudf-cu11",
"> Hi, if this helps anyone, pip install pyarrow==11.0.0 did not work for me (I'm using Colab) but this worked: !pip install --extra-index-url=https://pypi.nvidia.com cudf-cu11\r\n\r\nthanks! I met the same problem and your suggestion solved it.",
"(I was doing quiet install so I didn't notice it initially)\r\nI've been loading the same dataset for months on Colab, just now I got this error as well. I think Colab has changed their image recently (I had some errors regarding CUDA previously as well). beware of this and restart runtime if you're doing quite pip installs.\r\nmoreover installing stable version of datasets on pypi gives this:\r\n\r\n```\r\nERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\r\nibis-framework 7.1.0 requires pyarrow<15,>=2, but you have pyarrow 15.0.0 which is incompatible.\r\nSuccessfully installed datasets-2.17.0 dill-0.3.8 multiprocess-0.70.16 pyarrow-15.0.0\r\nWARNING: The following packages were previously imported in this runtime:\r\n [pyarrow]\r\nYou must restart the runtime in order to use newly installed versions.\r\n``` \r\n",
"for colab - pip install pyarrow==11.0.0",
"The above methods didn't help me. So I installed an older version: `!pip install datasets==2.16.1`\r\nand `import datasets` worked!!",
"@rasith1998 @PennlaineChu You can avoid this issue by restarting the session after the `datasets` installation (see https://github.com/huggingface/datasets/issues/6661 for more info)\r\n\r\nAlso, we've contacted Google Colab folks to update the default PyArrow installation, so the issue should soon be \"officially\" resolved on their side.",
"> Also, we've contacted Google Colab folks to update the default PyArrow installation, so the issue should soon be \"officially\" resolved on their side.\r\n\r\nThis has been done! Google Colab now pre-installs PyArrow 14.0.2, which makes this issue unlikely to happen, so I'm closing it.",
"I am facing this issue outside of Colab, in a normal Python (3.10.14) environment:\r\n```\r\npyarrow==11.0.0\r\ndatasets=2.20.0\r\ntransformers==4.41.2\r\n```\r\n\r\nWhat can I do to solve it?\r\n\r\nI am somewhat bound to `pyarrow==11.0.0`. Is there a version of `datasets` that supports this?"
] | ### Describe the bug
When trying to import datasets, I get a pyarrow ValueError:
Traceback (most recent call last):
File "/Users/edward/test/test.py", line 1, in <module>
import datasets
File "/Users/edward/opt/anaconda3/envs/cs235/lib/python3.9/site-packages/datasets/__init__.py", line 43, in <module>
from .arrow_dataset import Dataset
File "/Users/edward/opt/anaconda3/envs/cs235/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 65, in <module>
from .arrow_reader import ArrowReader
File "/Users/edward/opt/anaconda3/envs/cs235/lib/python3.9/site-packages/datasets/arrow_reader.py", line 28, in <module>
import pyarrow.parquet as pq
File "/Users/edward/opt/anaconda3/envs/cs235/lib/python3.9/site-packages/pyarrow/parquet/__init__.py", line 20, in <module>
from .core import *
File "/Users/edward/opt/anaconda3/envs/cs235/lib/python3.9/site-packages/pyarrow/parquet/core.py", line 45, in <module>
from pyarrow.fs import (LocalFileSystem, FileSystem, FileType,
File "/Users/edward/opt/anaconda3/envs/cs235/lib/python3.9/site-packages/pyarrow/fs.py", line 49, in <module>
from pyarrow._gcsfs import GcsFileSystem # noqa
File "pyarrow/_gcsfs.pyx", line 1, in init pyarrow._gcsfs
ValueError: pyarrow.lib.IpcWriteOptions size changed, may indicate binary incompatibility. Expected 88 from C header, got 72 from PyObject
### Steps to reproduce the bug
`import datasets`
### Expected behavior
Successful import
### Environment info
Conda environment, MacOS
python 3.9.12
datasets 2.12.0
| 5,923 |
https://github.com/huggingface/datasets/issues/5922 | Length of table does not accurately reflect the split | [
"As already replied by @lhoestq (private channel):\r\n> `.train_test_split` (as well as `.shard`, `.select`) doesn't create a new arrow table to save time and disk space. Instead, it uses an indices mapping on top of the table that locate which examples are part of train or test.",
"This is an optimization that we don't plan to \"fix\", so I'm closing this issue."
] | ### Describe the bug
I load a Huggingface Dataset and do `train_test_split`. I'm expecting the underlying table for the dataset to also be split, but it's not.
### Steps to reproduce the bug
### Expected behavior
The expected behavior is when `len(hf_dataset["train"].data)` should match the length of the train split, and not be the entire unsplit dataset.
### Environment info
datasets 2.10.1
python 3.10.11 | 5,922 |
https://github.com/huggingface/datasets/issues/5918 | File not found for audio dataset | [
"load_dataset () did not work for loading local files either "
] | ### Describe the bug
After loading an audio dataset, and looking at a sample entry, the `path` element, which is supposed to be the path to the audio file, doesn't actually exist.
### Steps to reproduce the bug
Run bug.py:
```py
import os.path
from datasets import load_dataset
def run() -> None:
cv13 = load_dataset(
"mozilla-foundation/common_voice_13_0",
"hi",
split="train",
)
print(cv13[0])
audio_file = cv13[0]["path"]
if not os.path.exists(audio_file):
raise ValueError(f'File {audio_file} does not exist.')
if __name__ == "__main__":
run()
```
The result (on my machine):
```json
{'client_id': '0f018a99663f33afbb7d38aee281fb1afcfd07f9e7acd00383f604e1e17c38d6ed8adf1bd2ccbf927a52c5adefb8ac4b158ce27a7c2ed9581e71202eb302dfb3', 'path': 'C:\\Users\\rober\\.cache\\huggingface\\datasets\\downloads\\extracted\\8d1479bc09b4609bc2675bd02d6869a4d5e09f7e6616f540bd55eacef46c6e2b\\common_voice_hi_26008353.mp3', 'audio': {'path': 'C:\\Users\\rober\\.cache\\huggingface\\datasets\\downloads\\extracted\\8d1479bc09b4609bc2675bd02d6869a4d5e09f7e6616f540bd55eacef46c6e2b\\common_voice_hi_26008353.mp3', 'array': array([ 6.46234854e-26, -1.35709319e-25, -8.07793567e-26, ...,
1.06425944e-07, 4.46417090e-08, 2.61451660e-09]), 'sampling_rate': 48000}, 'sentence': 'हमने उसका जन्मदिन मनाया।', 'up_votes': 2, 'down_votes': 0, 'age': '', 'gender': '', 'accent': '', 'locale': 'hi', 'segment': '' ', 'variant': ''}
```
```txt
Traceback (most recent call last):
File "F:\eo-reco\bug.py", line 18, in <module>
run()
File "F:\eo-reco\bug.py", line 15, in run
raise ValueError(f'File {audio_file} does not exist.')
ValueError: File C:\Users\rober\.cache\huggingface\datasets\downloads\extracted\8d1479bc09b4609bc2675bd02d6869a4d5e09f7e6616f540bd55eacef46c6e2b\common_voice_hi_26008353.mp3 does not exist.
```
### Expected behavior
The `path` element points to the correct file, which happens to be:
```
C:\Users\rober\.cache\huggingface\datasets\downloads\extracted\8d1479bc09b4609bc2675bd02d6869a4d5e09f7e6616f540bd55eacef46c6e2b\hi_train_0\common_voice_hi_26008353.mp3
```
That is, there's an extra directory `hi_train_0` that is not in the `path` element.
### Environment info
- `datasets` version: 2.12.0
- Platform: Windows-10-10.0.22621-SP0
- Python version: 3.11.3
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.1
- | 5,918 |
https://github.com/huggingface/datasets/issues/5914 | array is too big; `arr.size * arr.dtype.itemsize` is larger than the maximum possible size in Datasets | [] | ### Describe the bug
When using the `filter` or `map` function to preprocess a dataset, a ValueError is encountered with the error message "array is too big; arr.size * arr.dtype.itemsize is larger than the maximum possible size."
Detailed error message:
Traceback (most recent call last):
File "data_processing.py", line 26, in <module>
processed_dataset[split] = samromur_children[split].map(prepare_dataset, cache_file_name=cache_dict[split],writer_batch_size = 50)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 2405, in map
desc=desc,
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 557, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 524, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/fingerprint.py", line 480, in wrapper
out = func(self, *args, **kwargs)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 2756, in _map_single
example = apply_function_on_filtered_inputs(example, i, offset=offset)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 2655, in apply_function_on_filtered_inputs
processed_inputs = function(*fn_args, *additional_args, **fn_kwargs)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 2347, in decorated
result = f(decorated_item, *args, **kwargs)
File "data_processing.py", line 11, in prepare_dataset
audio = batch["audio"]
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 123, in __getitem__
value = decode_nested_example(self.features[key], value) if value is not None else None
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/features/features.py", line 1260, in decode_nested_example
return schema.decode_example(obj, token_per_repo_id=token_per_repo_id) if obj is not None else None
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/features/audio.py", line 156, in decode_example
array, sampling_rate = self._decode_non_mp3_path_like(path, token_per_repo_id=token_per_repo_id)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/datasets/features/audio.py", line 257, in _decode_non_mp3_path_like
array, sampling_rate = librosa.load(f, sr=self.sampling_rate, mono=self.mono)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/librosa/core/audio.py", line 176, in load
y, sr_native = __soundfile_load(path, offset, duration, dtype)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/librosa/core/audio.py", line 222, in __soundfile_load
y = sf_desc.read(frames=frame_duration, dtype=dtype, always_2d=False).T
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/soundfile.py", line 891, in read
out = self._create_empty_array(frames, always_2d, dtype)
File "/projects/zhwa3087/software/anaconda/envs/mycustomenv/lib/python3.7/site-packages/soundfile.py", line 1323, in _create_empty_array
return np.empty(shape, dtype, order='C')
ValueError: array is too big; `arr.size * arr.dtype.itemsize` is larger than the maximum possible size.
### Steps to reproduce the bug
```python
from datasets import load_dataset, DatasetDict
from transformers import WhisperFeatureExtractor
from transformers import WhisperTokenizer
samromur_children= load_dataset("language-and-voice-lab/samromur_children")
feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-small")
tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-small", language="icelandic", task="transcribe")
def prepare_dataset(batch):
# load and resample audio data from 48 to 16kHz
audio = batch["audio"]
# compute log-Mel input features from input audio array
batch["input_features"] = feature_extractor(audio["array"], sampling_rate=16000).input_features[0]
# encode target text to label ids
batch["labels"] = tokenizer(batch["normalized_text"]).input_ids
return batch
cache_dict = {"train": "./cache/audio_train.cache", \
"validation": "./cache/audio_validation.cache", \
"test": "./cache/audio_test.cache"}
filter_cache_dict = {"train": "./cache/filter_train.arrow", \
"validation": "./cache/filter_validation.arrow", \
"test": "./cache/filter_test.arrow"}
print("before filtering")
print(samromur_children)
#filter the dataset to only include examples with more than 2 seconds of audio
samromur_children = samromur_children.filter(lambda example: example["audio"]["array"].shape[0] > 16000*2, cache_file_names=filter_cache_dict)
print("after filtering")
print(samromur_children)
processed_dataset = DatasetDict()
# processed_dataset = samromur_children.map(prepare_dataset, cache_file_names=cache_dict, num_proc=10,)
for split in ["train", "validation", "test"]:
processed_dataset[split] = samromur_children[split].map(prepare_dataset, cache_file_name=cache_dict[split])
```
### Expected behavior
The dataset is successfully processed and ready to train the model.
### Environment info
Python version: 3.7.13
datasets package version: 2.4.0
librosa package version: 0.10.0.post2 | 5,914 |
https://github.com/huggingface/datasets/issues/5913 | I tried to load a custom dataset using the following statement: dataset = load_dataset('json', data_files=data_files). The dataset contains 50 million text-image pairs, but an error occurred. | [
"Thanks for reporting, @cjt222.\r\n\r\nWhat is the structure of your JSON files. Please note that it is normally simpler if the data file format is JSON-Lines instead. ",
"> Thanks for reporting, @cjt222.\r\n> \r\n> What is the structure of your JSON files. Please note that it is normally simpler if the data file format is JSON-Lines instead.\r\n\r\nThanks! I have encountered similar problems. I modify the json format from list to line and works!"
] | ### Describe the bug
File "/home/kas/.conda/envs/diffusers/lib/python3.7/site-packages/datasets/builder.py", line 1858, in _prepare_split_single
Downloading and preparing dataset json/default to /home/kas/diffusers/examples/dreambooth/cache_data/datasets/json/default-acf423d8c6ef99d0/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4...
Downloading data files: 0%| | 0/1 [00:00<?, ?it/s] Downloading data files: 100%|██████████| 1/1 [00:00<00:00, 84.35it/s]
Extracting data files: 0%| | 0/1 [00:00<?, ?it/s] for _, table in generator:
File "/home/kas/.conda/envs/diffusers/lib/python3.7/site-packages/datasets/packaged_modules/json/json.py", line 114, in _generate_tables
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
File "pyarrow/_json.pyx", line 258, in pyarrow._json.read_json
Extracting data files: 100%|██████████| 1/1 [00:00<00:00, 27.72it/s]
Generating train split: 0 examples [00:00, ? examples/s] File "pyarrow/error.pxi", line 144, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 125, in pyarrow.lib.check_status
pyarrow.lib.ArrowCapacityError: array cannot contain more than 2147483646 bytes, have 2390448764
### Steps to reproduce the bug
1、data_files = ["1.json", "2.json", "3.json"]
2、dataset = load_dataset('json', data_files=data_files)
### Expected behavior
Read the dataset normally.
### Environment info
- `datasets` version: 2.12.0
- Platform: Linux-4.15.0-29-generic-x86_64-with-debian-buster-sid
- Python version: 3.7.16
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 1.3.5 | 5,913 |
https://github.com/huggingface/datasets/issues/5912 | Missing elements in `map` a batched dataset | [
"Hi ! in your code batching is **only used within** `map`, to process examples in batch. The dataset itself however is not batched and returns elements one by one.\r\n\r\nTo iterate on batches, you can do\r\n```python\r\nfor batch in dataset.iter(batch_size=8):\r\n ...\r\n```"
] | ### Describe the bug
As outlined [here](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Flength-error-using-map-with-datasets%2F40969%2F3%3Fu%3Dsachin)%2C the following collate function drops 5 out of possible 6 elements in the batch (it is 6 because out of the eight, two are bad links in laion). A reproducible [kaggle kernel ](https://www.kaggle.com/sachin/laion-hf-dataset/edit) can be found here.
The weirdest part is when inspecting the sizes of the tensors as shown below, both `tokenized_captions["input_ids"]` and `image_features` show the correct shapes. Simply the output only has one element (with the batch dimension squeezed out).
```python
class CollateFn:
def get_image(self, url):
try:
response = requests.get(url)
return Image.open(io.BytesIO(response.content)).convert("RGB")
except PIL.UnidentifiedImageError:
logger.info(f"Reading error: Could not transform f{url}")
return None
except requests.exceptions.ConnectionError:
logger.info(f"Connection error: Could not transform f{url}")
return None
def __call__(self, batch):
images = [self.get_image(url) for url in batch["url"]]
captions = [caption for caption, image in zip(batch["caption"], images) if image is not None]
images = [image for image in images if image is not None]
tokenized_captions = tokenizer(
captions,
padding="max_length",
truncation=True,
max_length=tokenizer.model_max_length,
return_tensors="pt",
)
image_features = torch.stack([torch.Tensor(feature_extractor(image)["pixel_values"][0]) for image in images])
# import pdb; pdb.set_trace()
return {"input_ids": tokenized_captions["input_ids"], "images": image_features}
collate_fn = CollateFn()
laion_ds = datasets.load_dataset("laion/laion400m", split="train", streaming=True)
laion_ds_batched = laion_ds.map(collate_fn, batched=True, batch_size=8, remove_columns=next(iter(laion_ds)).keys())
```
### Steps to reproduce the bug
A reproducible [kaggle kernel ](https://www.kaggle.com/sachin/laion-hf-dataset/edit) can be found here.
### Expected behavior
Would expect `next(iter(laion_ds_batched))` to produce two tensors of shape `(batch_size, 77)` and `batch_size, image_shape`.
### Environment info
datasets==2.12.0
python==3.10 | 5,912 |
https://github.com/huggingface/datasets/issues/5910 | Cannot use both set_format and set_transform | [
"Currently, it's not possible to chain `set_format`/`set_transform` calls (plus, this is a breaking change if we decide to implement it), so I see two possible solutions:\r\n* using `set_format`/`set_transform` for the 1st transform and then passing the transformed example/batch to the 2nd transform\r\n* implementing and registering a custom formatter (the relevant code is [here](https://github.com/huggingface/datasets/tree/main/src/datasets/formatting))\r\n\r\nBtw, your example requires a single `set_format` call:\r\n```python\r\nds.set_format(\"torch\", columns=[\"image\"], output_all_columns=True, dtype=torch.double)\r\n```",
"Hey Mario,\r\nThanks, for getting back to me. the toDouble was just an example my real life case requires many more transforms.\r\n\r\nWhat do you mean by:\r\n> using set_format/set_transform for the 1st transform and then passing the transformed example/batch to the 2nd transform\r\n\r\nHow would that go, I thought you can't chain them?\r\n\r\nAs for the custom formatter, is it possible to reference an existing formatter, in my case `torch_formatter` inside of my custom formatter?\r\n\r\nmaybe I can inherit from it and just call `super.recursive_tensorize()`?",
"> How would that go, I thought you can't chain them?\r\n\r\nYes, they cannot be chained. This is what I meant:\r\n```python\r\nds.set_transform(first_transform)\r\n# calling the 2nd transform on each accessed batch\r\nsecond_transform(ds[2:3])\r\n```\r\n\r\n> As for the custom formatter, is it possible to reference an existing formatter, in my case torch_formatter inside of my custom formatter?\r\n>\r\n>maybe I can inherit from it and just call super.recursive_tensorize()?\r\n\r\nYes, subclassing makes the most sense.",
"Great, thank you for the details.",
"https://github.com/huggingface/datasets/issues/6012"
] | ### Describe the bug
I need to process some data using the set_transform method but I also need the data to be formatted for pytorch before processing it.
I don't see anywhere in the documentation something that says that both methods cannot be used at the same time.
### Steps to reproduce the bug
```
from datasets import load_dataset
ds = load_dataset("mnist", split="train")
ds.set_format(type="torch")
def transform(entry):
return entry["image"].double()
ds.set_transform(transform)
print(ds[0])
```
### Expected behavior
It should print the pytorch tensor image as a double, but it errors because "entry" in the transform function doesn't receive a pytorch tensor to begin with, it receives a PIL Image -> entry.double() errors because entry isn't a pytorch tensor.
### Environment info
Latest versions.
### Note:
It would be at least handy to have access to a function that can do the dataset.set_format in the set_transform function.
Something like:
```
from datasets import load_dataset, do_format
ds = load_dataset("mnist", split="train")
def transform(entry):
entry = do_format(entry, type="torch")
return entry["image"].double()
ds.set_transform(transform)
print(ds[0])
``` | 5,910 |
https://github.com/huggingface/datasets/issues/5908 | Unbearably slow sorting on big mapped datasets | [
"Hi ! `shard` currently returns a slow dataset by default, with examples evenly distributed in the dataset.\r\n\r\nYou can get a fast dataset using `contiguous=True` (which should be the default imo):\r\n\r\n```python\r\ndataset = dataset.shard(10, 0, contiguous=True)\r\n```\r\n\r\nThis way you don't need to flatten_indices() and sort should be fast as well",
"@lhoestq \r\n\r\n> contiguous=True (which should be the default imo)\r\n\r\nFor `IterableDataset`, it's not possible to implement contiguous sharding without knowing the number of examples in advance, so setting the default value to `contiguous=True` would result in an inconsistency between `Dataset` and `IterableDataset` (when we add `IterableDataset.shard`)",
"Actually sharded iterable datasets are made of sub iterables that generally yield contiguous data no ? So in a way it's possible to shard an iterable dataset contiguously.\r\n\r\nIf the dataset is made of one shard it's indeed not possible to shard it contiguously though",
"> Actually sharded iterable datasets are made of sub iterables that generally yield contiguous data no ? So in a way it's possible to shard an iterable dataset contiguously.\r\n\r\nBut sharding an iterable dataset by sharding its `gen_kwargs` would still yield approximate shards(not equal to `Dataset.shard`), no? ",
"Yes indeed !",
"I understand the issue doesn't exist with non-mapped datasets, but if flattening is so much more efficient than sorting the indices, that's an issue in itself.\n\nThere are plenty of issues people posted for which the root cause turns out to be the same. It seems like mapped datasets are terribly inefficient. I think I saw some issue like that somewhere (about the mapped datasets in general), but can't find it now.\n\nMaybe indices should be flattened before any additional processing, then."
] | ### Describe the bug
For me, with ~40k lines, sorting took 3.5 seconds on a flattened dataset (including the flatten operation) and 22.7 seconds on a mapped dataset (right after sharding), which is about x5 slowdown. Moreover, it seems like it slows down exponentially with bigger datasets (wasn't able to sort 700k lines at all, with flattening takes about a minute).
### Steps to reproduce the bug
```Python
from datasets import load_dataset
import time
dataset = load_dataset("xnli", "en", split="train")
dataset = dataset.shard(10, 0)
print(len(dataset))
t = time.time()
# dataset = dataset.flatten_indices() # uncomment this line and it's fast
dataset = dataset.sort("label", reverse=True, load_from_cache_file=False)
print(f"finished in {time.time() - t:.4f} seconds")
```
### Expected behavior
Expect sorting to take the same or less time than flattening and then sorting.
### Environment info
- `datasets` version: 2.12.1.dev0 (same with 2.12.0 too)
- Platform: Windows-10-10.0.22621-SP0
- Python version: 3.10.10
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.1 | 5,908 |
https://github.com/huggingface/datasets/issues/5906 | Could you unpin responses version? | [] | ### Describe the bug
Could you unpin [this](https://github.com/huggingface/datasets/blob/main/setup.py#L139) or move it to test requirements? This is a testing library and we also use it for our tests as well. We do not want to use a very outdated version.
### Steps to reproduce the bug
could not install this library due to dependency conflict.
### Expected behavior
can install datasets
### Environment info
linux 64 | 5,906 |
https://github.com/huggingface/datasets/issues/5905 | Offer an alternative to Iterable Dataset that allows lazy loading and processing while skipping batches efficiently | [
"We plan to improve this eventually (see https://github.com/huggingface/datasets/issues/5454 and https://github.com/huggingface/datasets/issues/5380).\r\n\r\n> Is it possible to lazily load samples of a mapped dataset ? I'm used to [dataset scripts](https://huggingface.co/docs/datasets/dataset_script), maybe something can be done there.\r\nIf not, I could do it using a plain Pytorch dataset. Then I would need to convert it to a datasets' dataset to get all the features of datasets. Is it something possible ?\r\n\r\nYes, by creating a mapped dataset that stores audio URLs. Indexing a dataset in such format only downloads and decodes the bytes of the accessed samples (without storing them on disk).\r\n\r\nYou can do the following to create this dataset:\r\n```python\r\n\r\ndef gen():\r\n # Generator that yields (audio URL, text) pairs as dict\r\n ...\r\n yield {\"audio\": \"audio_url\", \"text\": \"some text\"}\r\n\r\nfeatures = Features({\"audio\": datasets.Audio(), \"text\": datasets.Value(\"string\")})\r\nds = Dataset.from_generator(gen, features=features)\r\nds[2:5] # downloads and decodes the samples each time they are accessed\r\n```"
] | ### Feature request
I would like a way to resume training from a checkpoint without waiting for a very long time when using an iterable dataset.
### Motivation
I am training models on the speech-recognition task. I have very large datasets that I can't comfortably store on a disk and also quite computationally intensive audio processing to do. As a result I want to load data from my remote when it is needed and perform all processing on the fly.
I am currently using the iterable dataset feature of _datasets_. It does everything I need with one exception. My issue is that when resuming training at a step n, we have to download all the data and perform the processing of steps < n, just to get the iterable at the right step. In my case it takes almost as long as training for the same steps, which make resuming training from a checkpoint useless in practice.
I understand that the nature of iterators make it probably nearly impossible to quickly resume training.
I thought about a possible solution nonetheless :
I could in fact index my large dataset and make it a mapped dataset. Then I could use set_transform to perform the processing on the fly. Finally, if I'm not mistaken, the _accelerate_ package allows to [skip steps efficiently](https://github.com/huggingface/accelerate/blob/a73898027a211c3f6dc4460351b0ec246aa824aa/src/accelerate/data_loader.py#L827) for a mapped dataset.
Is it possible to lazily load samples of a mapped dataset ? I'm used to [dataset scripts](https://huggingface.co/docs/datasets/dataset_script), maybe something can be done there.
If not, I could do it using a plain _Pytorch_ dataset. Then I would need to convert it to a _datasets_' dataset to get all the features of _datasets_. Is it something possible ?
### Your contribution
I could provide a PR to allow lazy loading of mapped dataset or the conversion of a mapped _Pytorch_ dataset into a _Datasets_ dataset if you think it is an useful new feature. | 5,905 |
https://github.com/huggingface/datasets/issues/5898 | Loading The flores data set for specific language | [
"got that the syntax is like this\r\n\r\ndataset = load_dataset(\"facebook/flores\", \"ace_Arab\")"
] | ### Describe the bug
I am trying to load the Flores data set
the code which is given is
```
from datasets import load_dataset
dataset = load_dataset("facebook/flores")
```
This gives the error of config name
""ValueError: Config name is missing"
Now if I add some config it gives me the some error
"HFValidationError: Repo id must use alphanumeric chars or '-', '_', '.', '--' and '..' are forbidden, '-' and '.' cannot start or end the name, max length is 96: 'facebook/flores, 'ace_Arab''.
"
How I can load the data of the specific language ?
Couldn't find any tutorial
any one can help me out?
### Steps to reproduce the bug
step one load the data set
`from datasets import load_dataset
dataset = load_dataset("facebook/flores")`
it gives the error of config
once config is given
it gives the error of
"HFValidationError: Repo id must use alphanumeric chars or '-', '_', '.', '--' and '..' are forbidden, '-' and '.' cannot start or end the name, max length is 96: 'facebook/flores, 'ace_Arab''.
"
### Expected behavior
Data set should be loaded but I am receiving error
### Environment info
Datasets , python , | 5,898 |
https://github.com/huggingface/datasets/issues/5896 | HuggingFace does not cache downloaded files aggressively/early enough | [
"I also faced this. Any update?",
"We've dropped the `apache-beam` dependency in https://huggingface.co/datasets/wikipedia/discussions/19, so you should no longer get this error."
] | ### Describe the bug
I wrote the following script:
```
import datasets
dataset = datasets.load.load_dataset("wikipedia", "20220301.en", split="train[:10000]")
```
I ran it and spent 90 minutes downloading a 20GB file. Then I saw:
```
Downloading: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20.3G/20.3G [1:30:29<00:00, 3.73MB/s]
Traceback (most recent call last):
File "/home/jack/Code/Projects/Transformers/Codebase/main.py", line 5, in <module>
dataset = datasets.load.load_dataset("wikipedia", "20220301.en", split="train[:10000]")
File "/home/jack/.local/lib/python3.10/site-packages/datasets/load.py", line 1782, in load_dataset
builder_instance.download_and_prepare(
File "/home/jack/.local/lib/python3.10/site-packages/datasets/builder.py", line 883, in download_and_prepare
self._save_info()
File "/home/jack/.local/lib/python3.10/site-packages/datasets/builder.py", line 2037, in _save_info
import apache_beam as beam
ModuleNotFoundError: No module named 'apache_beam'
```
And the 20GB of data was seemingly instantly gone forever, because when I ran the script again, it had to do the download again.
### Steps to reproduce the bug
See above
### Expected behavior
See above
### Environment info
datasets 2.10.1
Python 3.10 | 5,896 |
https://github.com/huggingface/datasets/issues/5895 | The dir name and split strings are confused when loading ArmelR/stack-exchange-instruction dataset | [
"Thanks for reporting, @DongHande.\r\n\r\nI think the issue is caused by the metadata in the dataset card: in the header of the `README.md`, they state that the dataset has 4 splits (\"finetune\", \"reward\", \"rl\", \"evaluation\"). \r\n```yaml\r\n splits:\r\n - name: finetune\r\n num_bytes: 6674567576\r\n num_examples: 3000000\r\n - name: reward\r\n num_bytes: 6674341521\r\n num_examples: 3000000\r\n - name: rl\r\n num_bytes: 6679279968\r\n num_examples: 3000000\r\n - name: evaluation\r\n num_bytes: 4022714493\r\n num_examples: 1807695\r\n```\r\n\r\n\r\nI guess the user wanted to define these as configs, instead of splits. This is not yet supported for no-script datasets, but will be soon supported. See:\r\n- #5331\r\n\r\nI think we should contact the dataset author to inform about the issue with the split names, as you already did: https://huggingface.co/datasets/ArmelR/stack-exchange-instruction/discussions/1\r\nLet's continue the discussion there!",
"Thank you! It has been fixed. "
] | ### Describe the bug
When I load the ArmelR/stack-exchange-instruction dataset, I encounter a bug that may be raised by confusing the dir name string and the split string about the dataset.
When I use the script "datasets.load_dataset('ArmelR/stack-exchange-instruction', data_dir="data/finetune", split="train", use_auth_token=True)", it fails. But it succeeds when I add the "streaming = True" parameter.
The website of the dataset is https://huggingface.co/datasets/ArmelR/stack-exchange-instruction/ .
The traceback logs are as below:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/load.py", line 1797, in load_dataset
builder_instance.download_and_prepare(
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/builder.py", line 890, in download_and_prepare
self._download_and_prepare(
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/builder.py", line 985, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/builder.py", line 1706, in _prepare_split
split_info = self.info.splits[split_generator.name]
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/splits.py", line 530, in __getitem__
instructions = make_file_instructions(
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/arrow_reader.py", line 112, in make_file_instructions
name2filenames = {
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/arrow_reader.py", line 113, in <dictcomp>
info.name: filenames_for_dataset_split(
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/naming.py", line 70, in filenames_for_dataset_split
prefix = filename_prefix_for_split(dataset_name, split)
File "/home/xxx/miniconda3/envs/code/lib/python3.9/site-packages/datasets/naming.py", line 54, in filename_prefix_for_split
if os.path.basename(name) != name:
File "/home/xxx/miniconda3/envs/code/lib/python3.9/posixpath.py", line 142, in basename
p = os.fspath(p)
TypeError: expected str, bytes or os.PathLike object, not NoneType
### Steps to reproduce the bug
1. import datasets library function: ```from datasets import load_dataset```
2. load dataset: ```ds=load_dataset('ArmelR/stack-exchange-instruction', data_dir="data/finetune", split="train", use_auth_token=True)```
### Expected behavior
The dataset can be loaded successfully without the streaming setting.
### Environment info
Linux,
python=3.9
datasets=2.12.0 | 5,895 |
https://github.com/huggingface/datasets/issues/5892 | User access requests with manual review do not notify the dataset owner | [
"cc @SBrandeis",
"I think this has been addressed.\r\n\r\nPlease open a new issue if you are still not getting notified."
] | ### Describe the bug
When a user access requests are enabled, and new requests are set to Manual Review, the dataset owner should be notified of the pending requests. However, instead, currently nothing happens, and so the dataset request can go unanswered for quite some time until the owner happens to check that particular dataset's Settings pane.
### Steps to reproduce the bug
1. Enable a dataset's user access requests
2. Set to Manual Review
3. Ask another HF user to request access to the dataset
4. Dataset owner is not notified
### Expected behavior
The dataset owner should receive some kind of notification, perhaps in their HF site inbox, or by email, when a dataset access request is made and manual review is enabled.
### Environment info
n/a | 5,892 |
https://github.com/huggingface/datasets/issues/5889 | Token Alignment for input and output data over train and test batch/dataset. | [] | `data`
> DatasetDict({
train: Dataset({
features: ['input', 'output'],
num_rows: 4500
})
test: Dataset({
features: ['input', 'output'],
num_rows: 500
})
})
**# input (in-correct sentence)**
`data['train'][0]['input']`
**>>** 'We are meet sunday 10am12pmET in Crown Heights Brooklyn New York'
**# output (correct sentence)**
`data['train'][0]['output']`
**>>** 'We meet Sundays 10am-12pmET in Crown Heights, Brooklyn, New York.'
**I Want to align the output tokens with input**
```
`# tokenize both inputs and targets
def tokenize_fn(batch):
# tokenize the input sequence first
# this populates input_ids, attention_mask, etc.
tokenized_inputs = tokenizer(
batch['input']
)
labels_batch = tokenizer.tokenize(batch['output']) # original targets
aligned_labels_batch = []
for i, labels in enumerate(labels_batch):
word_ids = tokenized_inputs[i].word_ids()
aligned_labels_batch.append(align_targets(labels, word_ids)) # align_targets is another user defined function which is been called here
# recall: the 'target' must be stored in key called 'labels'
tokenized_inputs['labels'] = aligned_labels_batch
return tokenized_inputs`
```
```
data.map(
tokenize_fn,
batched=True,
remove_columns=data['train'].column_names,
)
```
When this user defined function is mapped to every records of train and test batch am getting following error:
**1.** **raise DatasetTransformationNotAllowedError(
3457 "Using `.map` in batched mode on a dataset with attached indexes is allowed only if it doesn't create or remove existing examples. You can first run `.drop_index() to remove your index and then re-add it."**
**2.** **TypeError: TextEncodeInput must be Union[TextInputSequence, Tuple[InputSequence, InputSequence]]** | 5,889 |
https://github.com/huggingface/datasets/issues/5887 | HuggingsFace dataset example give error | [
"Nice catch @donhuvy, that's because some models don't need the `token_type_ids`, as in this case, as the example is using `distilbert-base-cased`, and according to the DistilBert documentation at https://huggingface.co/transformers/v3.0.2/model_doc/distilbert.html, `DistilBert doesn’t have token_type_ids, you don’t need to indicate which token belongs to which segment. Just separate your segments with the separation token tokenizer.sep_token (or [SEP])`. `token_type_ids` are neither required in some other well known models such as RoBERTa. \r\n\r\nHere the issue comes due to a mismatch between the tokenizer and the model, as the Colab is using a BERT tokenizer (`bert-base-cased`), while the model is a DistilBERT (`distilbert-base-cased`), so aligning the tokenizer and the model solves it!",
"#self-assign",
"@donhuvy I've created https://github.com/huggingface/datasets/pull/5902 to solve it! 🤗",
"This has been addressed in #5902.\r\n\r\nThe Quicktour notebook is deprecated now - please use the notebook version of the [Quickstart doc page](https://huggingface.co/docs/datasets/main/en/quickstart) instead (\"Open in Colab\" button)."
] | ### Describe the bug
### Steps to reproduce the bug
Use link as reference document written https://colab.research.google.com/github/huggingface/datasets/blob/main/notebooks/Overview.ipynb#scrollTo=biqDH9vpvSVz
```python
# Now let's train our model
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model.train().to(device)
for i, batch in enumerate(dataloader):
batch.to(device)
outputs = model(**batch)
loss = outputs.loss
loss.backward()
optimizer.step()
model.zero_grad()
print(f'Step {i} - loss: {loss:.3}')
if i > 5:
break
```
Error
```python
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
[<ipython-input-44-7040b885f382>](https://localhost:8080/#) in <cell line: 5>()
5 for i, batch in enumerate(dataloader):
6 batch.to(device)
----> 7 outputs = model(**batch)
8 loss = outputs.loss
9 loss.backward()
[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in _call_impl(self, *args, **kwargs)
1499 or _global_backward_pre_hooks or _global_backward_hooks
1500 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501 return forward_call(*args, **kwargs)
1502 # Do not call functions when jit is used
1503 full_backward_hooks, non_full_backward_hooks = [], []
TypeError: DistilBertForQuestionAnswering.forward() got an unexpected keyword argument 'token_type_ids'
```
https://github.com/huggingface/datasets/assets/1328316/5d8b1d61-9337-4d59-8423-4f37f834c156
### Expected behavior
Run success on Google Colab (free)
### Environment info
Windows 11 x64, Google Colab free (my Google Drive just empty about 200 MB, but I don't think it cause problem) | 5,887 |
https://github.com/huggingface/datasets/issues/5886 | Use work-stealing algorithm when parallel computing | [
"Alternatively we could set the number of shards to be a factor than the number of processes (current they're equal) - this way it will be less likely to end up with a shard that is significantly slower than all the other ones."
] | ### Feature request
when i used Dataset.map api to process data concurrently, i found that
it gets slower and slower as it gets closer to completion. Then i read the source code of arrow_dataset.py and found that it shard the dataset and use multiprocessing pool to execute each shard.It may cause the slowest task to drag out the entire program's execution time,especially when processing huge dataset.
### Motivation
using work-stealing algorithm instead of sharding and parallel computing to optimize performance.
### Your contribution
just an idea. | 5,886 |
https://github.com/huggingface/datasets/issues/5888 | A way to upload and visualize .mp4 files (millions of them) as part of a dataset | [
"Hi! \r\n\r\nYou want to use `push_to_hub` (creates Parquet files) instead of `save_to_disk` (creates Arrow files) when creating a Hub dataset. Parquet is designed for long-term storage and takes less space than the Arrow format, and, most importantly, `load_dataset` can parse it, which should fix the viewer. \r\n\r\nRegarding the dataset generation, `Dataset.from_generator` with the video data represented as `datasets.Value(\"binary\")` followed by `push_to_hub` should work (if the `push_to_hub` step times out, restart it to resume uploading)\r\n\r\nPS: Once the dataset is uploaded, to make working with the dataset easier, it's a good idea to add a [transform](https://huggingface.co/docs/datasets/main/en/process#format-transform) to the README that shows how to decode the binary video data into something a model can understand. Also, if you get an `ArrowInvalid` error (can happen when working with large binary data) in `Dataset.from_generator`, reduce the value of `writer_batch_size` (the default is 1000) to fix it.",
"One issue here is that Dataset.from_generator can work well for the non 'infinite sampling' version of the dataset. The training set for example is often sampled dynamically given the video files that I have uploaded. I worry that storing the video data as binary means that I'll end up duplicating a lot of the data. Furthermore, storing video data as anything but .mp4 would quickly make the dataset size from 1.9TB to 1PB. ",
"> storing video data as anything but .mp4\r\n\r\nWhat I mean by storing as `datasets.Value(\"binary\")` is embedding raw MP4 bytes in the Arrow table, but, indeed, this would waste a lot of space if there are duplicates.\r\n\r\nSo I see two options:\r\n* if one video is not mapped to too many samples, you can embed the video bytes and do \"group by\" on the rest of the columns (this would turn them into lists) to avoid duplicating them (then, it should be easy to define a `map` in the README that samples the video data to \"unpack\" the samples)\r\n* you can create a dataset script that downloads the video files and embeds their file paths into the Arrow file\r\n\r\nAlso, I misread MP4 as MP3. We need to add a `Video` feature to the `datasets` lib to support MP4 files in the viewer (a bit trickier to implement than the `Image` feature due to the Arrow limitations).",
"I'm transferring this issue to the `datasets` repo, as it's not related to `huggingface_hub`",
"@mariosasko Right. If I want my dataset to be streamable, what are the necessary requirements to achieve that within the context of .mp4 binaries like we have here? I guess your second point here would not support that right?",
"The streaming would work, but the video paths would require using `fsspec.open` to get the content.",
"Are there any plans to make video playable on the hub?",
"Not yet. The (open source) tooling for video is not great in terms of ease of use/performance, so we are discussing internally the best way to support it (one option is creating a new library for video IO, but this will require a lot of work)",
"True. I spend a good 4 months just mixing and matching existing solutions so I could get performance that would not IO bound my model training. \r\n\r\nThis is what I ended up with, in case it's useful\r\n\r\nhttps://github.com/AntreasAntoniou/TALI/blob/045cf9e5aa75b1bf2c6d5351fb910fa10e3ff32c/tali/data/data_plus.py#L85"
] | **Is your feature request related to a problem? Please describe.**
I recently chose to use huggingface hub as the home for a large multi modal dataset I've been building. https://huggingface.co/datasets/Antreas/TALI
It combines images, text, audio and video. Now, I could very easily upload a dataset made via datasets.Dataset.from_generator, as long as it did not include video files. I found that including .mp4 files in the entries would not auto-upload those files.
Hence I tried to upload them myself. I quickly found out that uploading many small files is a very bad way to use git lfs, and that it would take ages, so, I resorted to using 7z to pack them all up. But then I had a new problem.
My dataset had a size of 1.9TB. Trying to upload such a large file with the default huggingface_hub API always resulted in time outs etc. So I decided to split the large files into chunks of 5GB each and reupload.
So, eventually it all worked out. But now the dataset can't be properly and natively used by the datasets API because of all the needed preprocessing -- and furthermore the hub is unable to visualize things.
**Describe the solution you'd like**
A native way to upload large datasets that include .mp4 or other video types.
**Describe alternatives you've considered**
Already explained earlier
**Additional context**
https://huggingface.co/datasets/Antreas/TALI
| 5,888 |
https://github.com/huggingface/datasets/issues/5884 | `Dataset.to_tf_dataset` fails when strings cannot be encoded as `np.bytes_` | [
"May eventually be solved in #5883 ",
"#self-assign"
] | ### Describe the bug
When loading any dataset that contains a column with strings that are not ASCII-compatible, looping over those records raises the following exception e.g. for `é` character `UnicodeEncodeError: 'ascii' codec can't encode character '\xe9' in position 0: ordinal not in range(128)`.
### Steps to reproduce the bug
Running the following script will eventually fail, when reaching to the batch that contains non-ASCII compatible strings.
```python
from datasets import load_dataset
ds = load_dataset("imdb", split="train")
tfds = ds.to_tf_dataset(batch_size=16)
for batch in tfds:
print(batch)
>>> UnicodeEncodeError: 'ascii' codec can't encode character '\xe9' in position 0: ordinal not in range(128)
```
### Expected behavior
The following script to run properly, making sure that the strings are either `numpy.unicode_` or `numpy.string` instead of `numpy.bytes_` since some characters are not ASCII compatible and that would lead to an issue when applying the `map`.
```python
from datasets import load_dataset
ds = load_dataset("imdb", split="train")
tfds = ds.to_tf_dataset(batch_size=16)
for batch in tfds:
print(batch)
```
### Environment info
- `datasets` version: 2.12.1.dev0
- Platform: macOS-13.3.1-arm64-arm-64bit
- Python version: 3.10.11
- Huggingface_hub version: 0.14.1
- PyArrow version: 11.0.0
- Pandas version: 1.5.3 | 5,884 |
https://github.com/huggingface/datasets/issues/5881 | Split dataset by node: index error when sharding iterable dataset | [
"cc @lhoestq in case you have any ideas here! Might need a multi-host set-up to debug (can give you access to a JAX one if you need)",
"I am also facing the same problem. Could you let me know if you found a solution for this?",
"I couldn't reproduce with the latest version of `datasets` 2.16.1, can you update `datasets` and try again ?"
] | ### Describe the bug
Context: we're splitting an iterable dataset by node and then passing it to a torch data loader with multiple workers
When we iterate over it for 5 steps, we don't get an error
When we instead iterate over it for 8 steps, we get an `IndexError` when fetching the data if we have too many workers
### Steps to reproduce the bug
Here, we have 2 JAX processes (`jax.process_count() = 2`) which we split the dataset over. The dataset loading script can be found here: https://huggingface.co/datasets/distil-whisper/librispeech_asr/blob/c6a1e805cbfeed5057400ac5937327d7e30281b8/librispeech_asr.py#L310
<details>
<summary> Code to reproduce </summary>
```python
from datasets import load_dataset
import jax
from datasets.distributed import split_dataset_by_node
from torch.utils.data import DataLoader
from tqdm import tqdm
# load an example dataset (https://huggingface.co/datasets/distil-whisper/librispeech_asr)
dataset = load_dataset("distil-whisper/librispeech_asr", "all", split="train.clean.100", streaming=True)
# just keep the text column -> no need to define a collator
dataset_text = dataset.remove_columns(set(dataset.features.keys()) - {"text"})
# define some constants
batch_size = 256
num_examples = 5 # works for 5 examples, doesn't for 8
num_workers = dataset_text.n_shards
# try with multiple workers
dataloader = DataLoader(dataset_text, batch_size=batch_size, num_workers=num_workers, drop_last=True)
for i, batch in tqdm(enumerate(dataloader), total=num_examples, desc="Multiple workers"):
if i == num_examples:
break
# try splitting by node (we can't do this with `dataset_text` since `split_dataset_by_node` expects the Audio column for an ASR dataset)
dataset = split_dataset_by_node(dataset, rank=jax.process_index(), world_size=jax.process_count())
# remove the text column again
dataset_text = dataset.remove_columns(set(dataset.features.keys()) - {"text"})
dataloader = DataLoader(dataset_text, batch_size=16, num_workers=num_workers // 2, drop_last=True)
for i, batch in tqdm(enumerate(dataloader), total=num_examples, desc="Split by node"):
if i == num_examples:
break
# too many workers
dataloader = DataLoader(dataset_text, batch_size=256, num_workers=num_workers, drop_last=True)
for i, batch in tqdm(enumerate(dataloader), total=num_examples, desc="Too many workers"):
if i == num_examples:
break
```
</details>
<details>
<summary> With 5 examples: </summary>
```
Multiple workers: 100%|███████████████████████████████████████████████████████████████████| 5/5 [00:16<00:00, 3.33s/it]
Assigning 7 shards (or data sources) of the dataset to each node.
Split by node: 100%|██████████████████████████████████████████████████████████████████████| 5/5 [00:13<00:00, 2.76s/it]
Assigning 7 shards (or data sources) of the dataset to each node.
Too many dataloader workers: 14 (max is dataset.n_shards=7). Stopping 7 dataloader workers.
To parallelize data loading, we give each process some shards (or data sources) to process. Therefore it's unnecessary t
o have a number of workers greater than dataset.n_shards=7. To enable more parallelism, please split the dataset in more
files than 7.
Too many workers: 100%|███████████████████████████████████████████████████████████████████| 5/5 [00:15<00:00, 3.03s/it]
```
</details>
<details>
<summary> With 7 examples: </summary>
```
Multiple workers: 100%|███████████████████████████████████████████████████████████████████| 8/8 [00:13<00:00, 1.71s/it]
Assigning 7 shards (or data sources) of the dataset to each node.
Split by node: 100%|██████████████████████████████████████████████████████████████████████| 8/8 [00:11<00:00, 1.38s/it]
Assigning 7 shards (or data sources) of the dataset to each node.
Too many dataloader workers: 14 (max is dataset.n_shards=7). Stopping 7 dataloader workers.
To parallelize data loading, we give each process some shards (or data sources) to process. Therefore it's unnecessary to have a number of workers greater than dataset.n_shards=7. To enable more parallelism, please split the dataset in more files than 7.
Too many workers: 88%|██████████████████████████████████████████████████████████▋ | 7/8 [00:13<00:01, 1.89s/it]
Traceback (most recent call last):
File "distil-whisper/test_librispeech.py", line 36, in <module>
for i, batch in tqdm(enumerate(dataloader), total=num_examples, desc="Too many workers"):
File "/home/sanchitgandhi/hf/lib/python3.8/site-packages/tqdm/std.py", line 1178, in __iter__
for obj in iterable:
File "/home/sanchitgandhi/hf/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 633, in __next__
data = self._next_data()
File "/home/sanchitgandhi/hf/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1325, in _next_data
return self._process_data(data)
File "/home/sanchitgandhi/hf/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1371, in _process_data
data.reraise()
File "/home/sanchitgandhi/hf/lib/python3.8/site-packages/torch/_utils.py", line 644, in reraise
raise exception
IndexError: Caught IndexError in DataLoader worker process 7.
Original Traceback (most recent call last):
File "/home/sanchitgandhi/hf/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 308, in _worker_loop
data = fetcher.fetch(index)
File "/home/sanchitgandhi/hf/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 32, in fetch
data.append(next(self.dataset_iter))
File "/home/sanchitgandhi/datasets/src/datasets/iterable_dataset.py", line 986, in __iter__
yield from self._iter_pytorch(ex_iterable)
File "/home/sanchitgandhi/datasets/src/datasets/iterable_dataset.py", line 920, in _iter_pytorch
for key, example in ex_iterable.shard_data_sources(worker_info.id, worker_info.num_workers):
File "/home/sanchitgandhi/datasets/src/datasets/iterable_dataset.py", line 540, in shard_data_sources
self.ex_iterable.shard_data_sources(worker_id, num_workers),
File "/home/sanchitgandhi/datasets/src/datasets/iterable_dataset.py", line 796, in shard_data_sources
self.ex_iterable.shard_data_sources(worker_id, num_workers),
File "/home/sanchitgandhi/datasets/src/datasets/iterable_dataset.py", line 126, in shard_data_sources
requested_gen_kwargs = _merge_gen_kwargs([gen_kwargs_list[i] for i in shard_indices])
File "/home/sanchitgandhi/datasets/src/datasets/utils/sharding.py", line 76, in _merge_gen_kwargs
for key in gen_kwargs_list[0]
IndexError: list index out of range
```
</details>
### Expected behavior
Should pass for both 5 and 7 examples
### Environment info
- `datasets` version: 2.12.1.dev0
- Platform: Linux-5.13.0-1023-gcp-x86_64-with-glibc2.29
- Python version: 3.8.10
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.1 | 5,881 |
https://github.com/huggingface/datasets/issues/5880 | load_dataset from s3 file system through streaming can't not iterate data | [
"This sounds related to #5281.\r\n\r\nCan you try passing `storage_options=s3_client.storage_options` instead passing it to `use_auth_token=` ?",
"I tried `storage_options` before, but it doesn't work, I checked our source code and I found that we even didn't pass this parameter to the following process. if I use `storage_options` instead of `use_auth_token`, then I also need to change another place of the code. the last line of `streaming_download_manager.py`. our code only passes the `use_auth_token` to the following handler, but does nothing to the `storage_options`\r\n<img width=\"1050\" alt=\"image\" src=\"https://github.com/huggingface/datasets/assets/59083384/5be90933-3331-4ecf-9e11-34f9852d8f92\">\r\n",
"Cloud storage support is still experimental indeed and you can expect some bugs.\r\n\r\nI think we need to pass the storage options anywhere use_auth_token is passed in indeed. Let me know if you'd be interested in contributing a fix !",
"Oh, that's great, I really like to fix it. because datasets is really useful and most of our projects need to use it, but we can store our data on the internet due to security reasons. fix it not only make our own work more efficient but also can benefit others who use it."
] | ### Describe the bug
I have a JSON file in my s3 file system(minio), I can use load_dataset to get the file link, but I can't iterate it
<img width="816" alt="image" src="https://github.com/huggingface/datasets/assets/59083384/cc0778d3-36f3-45b5-ac68-4e7c664c2ed0">
<img width="1144" alt="image" src="https://github.com/huggingface/datasets/assets/59083384/76872af3-8b3c-42ff-9f55-528c920a7af1">
we can change 4 lines to fix this bug, you can check whether it is ok for us.
<img width="941" alt="image" src="https://github.com/huggingface/datasets/assets/59083384/5a22155a-ece7-496c-8506-047e5c235cd3">
### Steps to reproduce the bug
1. storage a file in you s3 file system
2. use load_dataset to read it through streaming
3. iterate it
### Expected behavior
can iterate it successfully
### Environment info
- `datasets` version: 2.12.0
- Platform: macOS-10.16-x86_64-i386-64bit
- Python version: 3.8.16
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.1
| 5,880 |
https://github.com/huggingface/datasets/issues/5878 | Prefetching for IterableDataset | [
"Very cool! Do you have a link to the code that you're using to eagerly fetch the data? Would also be interested in hacking around something here for pre-fetching iterable datasets",
"I ended up just switching back to the pytorch dataloader and using it's multiprocessing functionality to handle this :(. I'm just not that familiar with python multiprocessing to get something to work in jupyter (kept having weird behaviors happening with zombies living after the cell finished).",
"Ultimately settled on using webdataset to circumvent huggingface datasets entirely. Would definitely switch back if: https://github.com/huggingface/datasets/issues/5337 was resolved.",
"Hi! You can combine `datasets` with `torchdata` to prefetch `IterableDataset`'s samples:\r\n```python\r\nfrom datasets import load_dataset\r\nfrom torchdata.datapipes.iter import IterableWrapper, HuggingFaceHubReader\r\nfrom torch.utils.data import DataLoader\r\n\r\nds = load_dataset(\"sst\", split=\"train\", streaming=True)\r\n# processing...\r\ndp = IterableWrapper(ds)\r\ndp = dp.prefetch(100)\r\ndl = DataLoader(dp, batch_size=8)\r\n\r\ni = iter(dl)\r\nnext(i)\r\n```",
"Hey @mariosasko! Thanks for the tip here - introducing prefetch with `torchdata` didn't really give me any performance difference vs not prefetching, but the concept is definitely one that could be really beneficial. Are there any benchmarks that show the speed-up you can get with `torchdata`'s prefetch just for comparison?"
] | ### Feature request
Add support for prefetching the next n batches through iterabledataset to reduce batch loading bottleneck in training loop.
### Motivation
The primary motivation behind this is to use hardware accelerators alongside a streaming dataset. This is required when you are in a low ram or low disk space setting as well as quick iteration where you're iterating though different accelerator environments (e.x changing ec2 instances quickly to figure out batch/sec for a particular architecture).
Currently, using the IterableDataset results in accelerators becoming basically useless due to the massive bottleneck induced by the dataset lazy loading/transform/mapping.
I've considered two alternatives:
PyTorch dataloader that handles this. However, I'm using jax, and I believe this is a piece of functionality that should live in the stream class.
Replicating the "num_workers" part of the PyTorch DataLoader to eagerly load batches and apply the transform so Arrow caching will automatically cache results and make them accessible.
### Your contribution
I may or may not have time to do this. Currently, I've written the basic multiprocessor approach to handle the eager DataLoader for my own use case with code that's not integrated to datasets. I'd definitely see this as being the default over the regular Dataset for most people given that they wouldn't have to wait on the datasets while also not worrying about performance. | 5,878 |
https://github.com/huggingface/datasets/issues/5877 | Request for text deduplication feature | [
"The \"exact match\" deduplication will be possible when we resolve https://github.com/huggingface/datasets/issues/2514 (first, https://github.com/apache/arrow/issues/30950 needs to be addressed on the Arrow side). In the meantime, you can use Polars or DuckDB (e.g., via [datasets-sql](https://github.com/mariosasko/datasets_sql)).\r\n\r\nFuzzy deduplication is out-of-scope for now ([splink](https://github.com/moj-analytical-services/splink) is probably the best tool for it).",
"This library can be an intermediate solution : https://github.com/ChenghaoMou/text-dedup/tree/main",
"I have been using polars to remove duplicates but it would be nice to do it directly in pyarrow.\r\n\r\nFor example,\r\n\r\n1. Read dataset with pyarrow\r\n2. Use scan_pyarrow_dataset() with Polars to create a LazyFrame\r\n3. Use sort and unique to remove duplicates based on a subset of columns\r\n4. Convert to table and save data with ds.write_dataset()\r\n\r\nThere are times where that workflow makes perfect sense because I do additional transformations with Polars. Most of the time I am simply just reading dataset A and writing dataset B without duplicates though, and I wish I could use a pyarrow scanner or table directly. ",
"Hi\r\nsee this new release from hf [datatrove](https://github.com/huggingface/datatrove)\r\nDataTrove is a library to process, filter and deduplicate text data at a very large scale. It provides a set of prebuilt commonly used processing blocks with a framework to easily add custom functionality"
] | ### Feature request
It would be great if there would be support for high performance, highly scalable text deduplication algorithms as part of the datasets library.
### Motivation
Motivated by this blog post https://huggingface.co/blog/dedup and this library https://github.com/google-research/deduplicate-text-datasets, but slightly frustrated by how its not very easy to work with these tools I am proposing this feature.
### Your contribution
I would be happy to contribute to the development effort of this feature. would love to collaborate with others in the development effort. | 5,877 |
https://github.com/huggingface/datasets/issues/5876 | Incompatibility with DataLab | [
"Indeed, `clobber=True` (with a warning if the existing protocol will be overwritten) should fix the issue, but maybe a better solution is to register our compression filesystem before the script is executed and unregister them afterward. WDYT @lhoestq @albertvillanova?",
"I think we should use clobber and show a warning if it overwrote a registered filesystem indeed ! This way the user can re-register the filesystems if needed. Though they should probably be compatible (and maybe do the exact same thing) so I wouldn't de-register the `datasets` filesystems"
] | ### Describe the bug
Hello,
I am currently working on a project where both [DataLab](https://github.com/ExpressAI/DataLab) and [datasets](https://github.com/huggingface/datasets) are subdependencies.
I noticed that I cannot import both libraries, as they both register FileSystems in `fsspec`, expecting the FileSystems not being registered before.
When running the code below, I get the following error:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\datalabs\__init__.py", line 28, in <module>
from datalabs.arrow_dataset import concatenate_datasets, Dataset
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\datalabs\arrow_dataset.py", line 60, in <module>
from datalabs.arrow_writer import ArrowWriter, OptimizedTypedSequence
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\datalabs\arrow_writer.py", line 28, in <module>
from datalabs.features import (
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\datalabs\features\__init__.py", line 2, in <module>
from datalabs.features.audio import Audio
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\datalabs\features\audio.py", line 21, in <module>
from datalabs.utils.streaming_download_manager import xopen
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\datalabs\utils\streaming_download_manager.py", line 16, in <module>
from datalabs.filesystems import COMPRESSION_FILESYSTEMS
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\datalabs\filesystems\__init__.py", line 37, in <module>
fsspec.register_implementation(fs_class.protocol, fs_class)
File "C:\Users\Bened\anaconda3\envs\ner-eval-dashboard2\lib\site-packages\fsspec\registry.py", line 51, in register_implementation
raise ValueError(
ValueError: Name (bz2) already in the registry and clobber is False
```
I think as simple solution would be to just set `clobber=True` in https://github.com/huggingface/datasets/blob/main/src/datasets/filesystems/__init__.py#L28. This allows the register to discard previous registrations. This should work, as the datalabs FileSystems are copies of the datasets FileSystems. However, I don't know if it is guaranteed to be compatible with other libraries that might use the same protocols.
I am linking the symmetric issue on [DataLab](https://github.com/ExpressAI/DataLab/issues/425) as ideally the issue is solved in both libraries the same way. Otherwise, it could lead to different behaviors depending on which library gets imported first.
### Steps to reproduce the bug
1. Run `pip install datalabs==0.4.15 datasets==2.12.0`
2. Run the following python code:
```
import datalabs
import datasets
```
### Expected behavior
It should be possible to import both libraries without getting a Value Error
### Environment info
datalabs==0.4.15
datasets==2.12.0
| 5,876 |
https://github.com/huggingface/datasets/issues/5875 | Why split slicing doesn't behave like list slicing ? | [
"A duplicate of https://github.com/huggingface/datasets/issues/1774"
] | ### Describe the bug
If I want to get the first 10 samples of my dataset, I can do :
```
ds = datasets.load_dataset('mnist', split='train[:10]')
```
But if I exceed the number of samples in the dataset, an exception is raised :
```
ds = datasets.load_dataset('mnist', split='train[:999999999]')
```
> ValueError: Requested slice [:999999999] incompatible with 60000 examples.
### Steps to reproduce the bug
```
ds = datasets.load_dataset('mnist', split='train[:999999999]')
```
### Expected behavior
I would expect it to behave like python lists (no exception raised, the whole list is kept) :
```
d = list(range(1000))[:999999]
print(len(d)) # > 1000
```
### Environment info
- `datasets` version: 2.9.0
- Platform: macOS-12.6-arm64-arm-64bit
- Python version: 3.9.12
- PyArrow version: 11.0.0
- Pandas version: 1.5.3 | 5,875 |
https://github.com/huggingface/datasets/issues/5874 | Using as_dataset on a "parquet" builder | [
"Hi! You can refer to [this doc](https://huggingface.co/docs/datasets/filesystems#load-and-save-your-datasets-using-your-cloud-storage-filesystem) to see the intended usage (basically, it skips the Arrow -> Parquet conversion step in `ds = load_dataset(...); ds.to_parquet(\"path/to/parquet\")`) and allows writing Parquet to remote storage unlike `to_parquet`).\r\n\r\n> I guess I'd expect as_dataset to generate the dataset in arrow format if it has to, or to suggest an alternative way to load the dataset (I've also tried other methods with load_dataset to no avail, probably due to misunderstandings on my part).\r\n\r\n`as_dataset` does not work with `file_format=\"parquet\"` files as Parquet files cannot be memory-mapped, so I think we should just raise an error in that case.\r\n"
] | ### Describe the bug
I used a custom builder to ``download_and_prepare`` a dataset. The first (very minor) issue is that the doc seems to suggest ``download_and_prepare`` will return the dataset, while it does not ([builder.py](https://github.com/huggingface/datasets/blob/main/src/datasets/builder.py#L718-L738)).
```
>>> from datasets import load_dataset_builder
>>> builder = load_dataset_builder("rotten_tomatoes")
>>> ds = builder.download_and_prepare("./output_dir", file_format="parquet")
```
The main issue I am facing is loading the dataset from those parquet files. I used the `as_dataset` method suggested by the doc, however it returns:
`
FileNotFoundError: [Errno 2] Failed to open local file 'output_dir/__main__-train-00000-of-00245.arrow'. Detail:
[errno 2] No such file or directory.
`
### Steps to reproduce the bug
1. Create a custom builder of some sort: `builder = CustomBuilder()`.
2. Run `download_and_prepare` with the parquet format: `builder.download_and_prepare("./output_dir", file_format="parquet")`.
3. Run `dataset = builder.as_dataset()`.
### Expected behavior
I guess I'd expect `as_dataset` to generate the dataset in arrow format if it has to, or to suggest an alternative way to load the dataset (I've also tried other methods with `load_dataset` to no avail, probably due to misunderstandings on my part).
### Environment info
```
- `datasets` version: 2.12.0
- Platform: Linux-5.15.0-1027-gcp-x86_64-with-glibc2.31
- Python version: 3.10.0
- Huggingface_hub version: 0.14.1
- PyArrow version: 8.0.0
- Pandas version: 1.5.3
``` | 5,874 |
https://github.com/huggingface/datasets/issues/5873 | Allow setting the environment variable for the lock file path | [] | ### Feature request
Add an environment variable to replace the default lock file path.
### Motivation
Usually, dataset path is a read-only path while the lock file needs to be modified each time. It would be convenient if the path can be reset individually.
### Your contribution
```/src/datasets/utils/filelock.py
class UnixFileLock(BaseFileLock):
def __init__(self, lock_file, timeout=-1, max_filename_length=None):
#-------------------
if os.getenv('DS_TMP_PATH'):
file_name = str(lock_file).split('/')[-1]
dataset_tmp_path = os.getenv('DS_TMP_PATH')
lock_file = os.path.join(dataset_tmp_path, file_name)
#-------------------
max_filename_length = os.statvfs(os.path.dirname(lock_file)).f_namemax
super().__init__(lock_file, timeout=timeout, max_filename_length=max_filename_length)
```
A simple demo is as upper. Thanks. | 5,873 |
https://github.com/huggingface/datasets/issues/5871 | data configuration hash suffix depends on uncanonicalized data_dir | [
"It could even use `os.path.realpath` to resolve symlinks.",
"Indeed, it makes sense to normalize `data_dir`. Feel free to submit a PR (this can be \"fixed\" [here](https://github.com/huggingface/datasets/blob/89f775226321ba94e5bf4670a323c0fb44f5f65c/src/datasets/builder.py#L173))",
"#self-assign"
] | ### Describe the bug
I am working with the `recipe_nlg` dataset, which requires manual download. Once it's downloaded, I've noticed that the hash in the custom data configuration is different if I add a trailing `/` to my `data_dir`. It took me a while to notice that the hashes were different, and to understand that that was the cause of my dataset being processed anew instead of the cached version being used.
### Steps to reproduce the bug
1. Follow the steps to manually download the `recipe_nlg` dataset to `/data/recipenlg`.
2. Load it using `load_dataset`, once without a trailing slash and once with one:
```python
>>> ds = load_dataset("recipe_nlg", data_dir="/data/recipenlg")
Using custom data configuration default-082278caeea85765
Downloading and preparing dataset recipe_nlg/default to /home/kyle/.cache/huggingface/datasets/recipe_nlg/default-082278caeea85765/1.0.0/aa4f120223637bedf7360cecb70a9bd108acfd64e38207ca90c9f385d21e5e74...
Dataset recipe_nlg downloaded and prepared to /home/kyle/.cache/huggingface/datasets/recipe_nlg/default-082278caeea85765/1.0.0/aa4f120223637bedf7360cecb70a9bd108acfd64e38207ca90c9f385d21e5e74. Subsequent calls will reuse this data.
100%|███████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.10s/it]
DatasetDict({
train: Dataset({
features: ['id', 'title', 'ingredients', 'directions', 'link', 'source', 'ner'],
num_rows: 2231142
})
})
>>> ds = load_dataset("recipe_nlg", data_dir="/data/recipenlg/")
Using custom data configuration default-83e87680785d0493
Downloading and preparing dataset recipe_nlg/default to /home/user/.cache/huggingface/datasets/recipe_nlg/default-83e87680785d0493/1.0.0/aa4f120223637bedf7360cecb70a9bd108acfd64e38207ca90c9f385d21e5e74...
Generating train split: 1%| | 12701/2231142 [00:04<13:15, 2790.25 examples/s
^C
```
3. Observe that the hash suffix in the custom data configuration changes due to the altered string.
### Expected behavior
I think I would expect the hash to remain constant if it actually points to the same location on disk. I would expect the use of `os.path.normpath` to canonicalize the paths.
### Environment info
- `datasets` version: 2.8.0
- Platform: Linux-5.4.0-147-generic-x86_64-with-glibc2.31
- Python version: 3.10.8
- PyArrow version: 10.0.1
- Pandas version: 1.5.2 | 5,871 |
https://github.com/huggingface/datasets/issues/5870 | Behaviour difference between datasets.map and IterableDatasets.map | [
"PS - some work is definitely needed for 'special cases' docs, not explanations, just usages of 'functions' under mixture of special cases, like a combination of custom databuilder + iterable dataset for large size + dynamic .map() application."
] | ### Describe the bug
All the examples in all the docs mentioned throughout huggingface datasets correspond to datasets object, and not IterableDatasets object. At one point of time, they might have been in sync, but the code for datasets version >=2.9.0 is very different as compared to the docs.
I basically need to .map() a transform on images in an iterable dataset, which was made using a custom databuilder config.
This works very good in map-styles datasets, but the .map() fails in IterableDatasets, show behvaiour as such:
"pixel_values" key not found, KeyError in examples object/dict passed into transform function for map, which works fine with map style, even as batch.
In iterable style, the object/dict passed into map() paramter callable function is completely different as what is mentioned in all examples.
Please look into this. Thank you
My databuilder class is inherited as such:
def _info(self):
print ("Config: ",self.config.__dict__.keys())
return datasets.DatasetInfo(
description=_DESCRIPTION,
features=datasets.Features(
{
"labels": datasets.Sequence(datasets.Value("uint16")),
# "labels_name": datasets.Value("string"),
# "pixel_values": datasets.Array3D(shape=(3, 1280, 960), dtype="float32"),
"pixel_values": datasets.Array3D(shape=(1280, 960, 3), dtype="uint8"),
"image_s3_path": datasets.Value("string"),
}
),
supervised_keys=None,
homepage="none",
citation="",
)
def _split_generators(self, dl_manager):
records_train = list(db.mini_set.find({'split':'train'},{'image_s3_path':1, 'ocwen_template_name':1}))[:10000]
records_val = list(db.mini_set.find({'split':'val'},{'image_s3_path':1, 'ocwen_template_name':1}))[:1000]
# print (len(records),self.config.num_shards)
# shard_size_train = len(records_train)//self.config.num_shards
# sharded_records_train = [records_train[i:i+shard_size_train] for i in range(0,len(records_train),shard_size_train)]
# shard_size_val = len(records_val)//self.config.num_shards
# sharded_records_val = [records_val[i:i+shard_size_val] for i in range(0,len(records_val),shard_size_val)]
return [
datasets.SplitGenerator(
name=datasets.Split.TRAIN, gen_kwargs={"records":records_train} # passing list of records, for sharding to take over
),
datasets.SplitGenerator(
name=datasets.Split.VALIDATION, gen_kwargs={"records":records_val} # passing list of records, for sharding to take over
),
]
def _generate_examples(self, records):
# print ("Generating examples for [{}] shards".format(len(shards)))
# initiate_db_connection()
# records = list(db.mini_set.find({'split':split},{'image_s3_path':1, 'ocwen_template_name':1}))[:10]
id_ = 0
# for records in shards:
for i,rec in enumerate(records):
img_local_path = fetch_file(rec['image_s3_path'],self.config.buffer_dir)
# t = self.config.processor(Image.open(img_local_path), random_padding=True, return_tensors="np").pixel_values.squeeze()
# print (t.shape, type(t),type(t[0][0][0]))
# sys.exit()
pvs = np.array(Image.open(img_local_path).resize((1280,960))) # image object is wxh, so resize as per that, numpy array of it is hxwxc, transposing to cxwxh
# pvs = self.config.processor(Image.open(img_local_path), random_padding=True, return_tensors="np").pixel_values.astype(np.float16).squeeze()
# print (type(pvs[0][0][0]))
lblids = self.config.processor.tokenizer('<s_class>'+rec['ocwen_template_name']+'</s_class>'+'</s>', add_special_tokens=False, padding=False, truncation=False, return_tensors="np")["input_ids"].squeeze(0) # take padding later, as per batch collating
# print (len(lblids),type(lblids[0]))
# print (type(pvs),pvs.shape,type(pvs[0][0][0]), type(lblids))
yield id_, {"labels":lblids,"pixel_values":pvs,"image_s3_path":rec['image_s3_path']}
id_+=1
os.remove(img_local_path)
and I load it inside my trainer script as such
`ds = load_dataset("/tmp/DonutDS/dataset/", split="train", streaming=True) # iterable dataset, where .map() falls`
or also as
`ds = load_from_disk('/tmp/DonutDS/dataset/') #map style dataset`
Thank you to the team for having such a great library, and for this bug fix in advance!
### Steps to reproduce the bug
Above config can allow one to reproduce the said bug
### Expected behavior
.map() should show some consistency b/w map-style and iterable-style datasets, or atleast the docs should address iterable-style datasets behaviour and examples. I honestly do not figure the use of such docs.
### Environment info
datasets==2.9.0
transformers==4.26.0 | 5,870 |
https://github.com/huggingface/datasets/issues/5869 | Image Encoding Issue when submitting a Parquet Dataset | [
"Hi @PhilippeMoussalli thanks for opening a detailed issue. It seems the issue is more related to the `datasets` library so I'll ping @lhoestq @mariosasko on this one :) \n\n(edit: also can one of you move the issue to the datasets repo? Thanks in advance 🙏)",
"Hi ! The `Image()` info is stored in the **schema metadata**. More precisely there should be a \"huggingface\" field in the schema metadata that contains the `datasets` feature type of each column.\r\n\r\nTo fix your issue, you can use the same schema as the original Parquet files to write the new ones. You can also get the schema with metadata from a `Features` object, e.g.\r\n\r\n```python\r\nfrom datasets import Features, Image, Value\r\n\r\nfeatures = Features({\"image\": Image(), \"text\": Value(\"string\")})\r\nschema = features.arrow_schema\r\nprint(schema.metadata)\r\n# {b'huggingface': b'{\"info\": {\"features\": {\"image\": {\"_type\": \"Image\"}, \"text\": {\"dtype\": \"string\", \"_type\": \"Value\"}}}}'}\r\n```",
"It appears that the parquet files at `hf://datasets/lambdalabs/pokemon-blip-captions` don't have this metadata, and it is defined in the dataset_infos.json instead (legacy).\r\n\r\nYou can get the right schema with the HF metadata this way:\r\n\r\n```python\r\nfrom datasets import load_dataset_builder\r\n\r\nfeatures = load_dataset_builder(\"lambdalabs/pokemon-blip-captions\").info.features\r\nschema = features.arrow_schema\r\n```",
"Btw in the future we might add support for an dedicated Image extension type in Arrow so that you won't need to add the schema metadata anymore ;)",
"Thanks @Wauplin @lhoestq for the quick reply :)! \r\n\r\nI tried your approach by passing the huggingface schema to the dask writer \r\n\r\n```\r\nfrom datasets import Features, Image, Value\r\ndf = dd.read_parquet(f\"hf://datasets/lambdalabs/pokemon-blip-captions\",index=False)\r\nfeatures = Features({\"image\": Image(), \"text\": Value(\"string\")})\r\nschema = features.arrow_schema\r\ndd.to_parquet(df, path = \"hf://datasets/philippemo/dummy_dataset/data\", schema=schema)\r\n```\r\nAt first it didn't work as I was not able to visualize the images, so then I manually added the `dataset_infos.json` from the example dataset and it worked :)\r\n\r\nHowever, It's not very ideal since there are some metadata in that file that need to be computed in order to load the data properly such as `num_of_bytes` and `num_examples` which might be unknown in my use case. \r\n\r\n\r\n\r\nDo you have any pointers there? you mentioned that `datasets_info.json` will be deprecated/legacy. Could you point me to some example image datasets on the hub that are stored as parquet and don't have the `datasets_info.json`?\r\n\r\n",
"You don't need the dataset_infos.json file as long as you have the schema with HF metadata ;)\r\nI could also check that it works fine myself on the git revision without the dataset_infos.json file.\r\n\r\nWhat made you think it didn't work ?",
"> You don't need the dataset_infos.json file as long as you have the schema with HF metadata ;) I could also check that it works fine myself on the git revision without the dataset_infos.json file.\r\n> \r\n> What made you think it didn't work ?\r\n\r\nThose are two identical dataset repos where both were pushed with dask with the specified schema you mentioned above. I then uploaded the `dataset_infos.json` manually taken from the original example dataset into one of them. \r\n\r\n* **With schema**: https://huggingface.co/datasets/philippemo/dummy_dataset_with_schema\r\n* **Without schema**: https://huggingface.co/datasets/philippemo/dummy_dataset_without_schema\r\n\r\nYou can see that in the examples without schema the images fail to render properly. When loaded with `datasets` they return an dict and not a Pillow Image ",
"I see ! I think it's a bug on our side - it should work without the metadata - let me investigate",
"Alright, it's fixed: https://huggingface.co/datasets/philippemo/dummy_dataset_without_schema\r\n\r\nIt shows the image correctly now - even without the extra metadata :)",
"Thanks @lhoestq! \r\nI tested pushing a dataset again without the metadata and it works perfectly! \r\nI appreciate the help",
"Hi @lhoestq, \r\n\r\nI'v tried pushing another dataset again and I think the issue reappeared again: \r\n\r\n```\r\ndf = dd.read_parquet(f\"hf://datasets/lambdalabs/pokemon-blip-captions\")\r\nfeatures = datasets.Features({\"image\": datasets.Image(), \"text\": datasets.Value(\"string\")})\r\nschema = features.arrow_schema\r\ndd.to_parquet(df, path = \"hf://datasets/philippemo/dummy_dataset_without_schema_12_06/data\", schema=schema)\r\n```\r\n\r\nHere is the dataset: \r\n https://huggingface.co/datasets/philippemo/dummy_dataset_without_schema_12_06\r\nThe one that was working 2 weeks ago still seems to be intact though, it might be that It rendered properly when it was initially submitted and after this something was reverted from your side:\r\nhttps://huggingface.co/datasets/philippemo/dummy_dataset_without_schema\r\n\r\nIt's weird because nothing really changed from the implementation, might be another issue in the hub backend. Do you have any pointers on how to resolve this? ",
"We're doing some changes in the way we're handling image parquet datasets right now. We'll include the fix from https://github.com/huggingface/datasets/pull/5921 in the new datasets-server version in the coming days",
"alright thanks for the update :), would that be part of the new release of datasets or is it something separate? if so, where can I track it? ",
"Once the new version of `datasets` is released (tomorrow probably) we'll open an issue on https://github.com/huggingface/datasets-server to update to this version :)",
"Alright we did the update :) This is fixed for good now",
"Yes thanks 🎉🎉🎉"
] | ### Describe the bug
Hello,
I'd like to report an issue related to pushing a dataset represented as a Parquet file to a dataset repository using Dask. Here are the details:
We attempted to load an example dataset in Parquet format from the Hugging Face (HF) filesystem using Dask with the following code snippet:
```
import dask.dataframe as dd
df = dd.read_parquet("hf://datasets/lambdalabs/pokemon-blip-captions",index=False)
```
In this dataset, the "image" column is represented as a dictionary/struct with the format:
```
df = df.compute()
df["image"].iloc[0].keys()
-> dict_keys(['bytes', 'path'])
```
I think this is the format encoded by the [`Image`](https://huggingface.co/docs/datasets/v2.0.0/en/package_reference/main_classes#datasets.Image) feature extractor from datasets to format suitable for Arrow.
The next step was to push the dataset to a repository that I created:
```
dd.to_parquet(dask_df, path = "hf://datasets/philippemo/dummy_dataset/data")
```
However, after pushing the dataset using Dask, the "image" column is now represented as the encoded dictionary `(['bytes', 'path'])`, and the images are not properly visualized. You can find the dataset here: [Link to the problematic dataset](https://huggingface.co/datasets/philippemo/dummy_dataset).
It's worth noting that both the original dataset and the one submitted with Dask have the same schema with minor alterations related to metadata:
**[ Schema of original dummy example.](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions/blob/main/data/train-00000-of-00001-566cc9b19d7203f8.parquet)**
```
image: struct<bytes: binary, path: null>
child 0, bytes: binary
child 1, path: null
text: string
```
**[ Schema of pushed dataset with dask](https://huggingface.co/datasets/philippemo/dummy_dataset/blob/main/data/part.0.parquet)**
```
image: struct<bytes: binary, path: null>
child 0, bytes: binary
child 1, path: null
text: string
```
This issue seems to be related to an encoding type that occurs when pushing a model to the hub. Normally, models should be represented as an HF dataset before pushing, but we are working with an example where we need to push large datasets using Dask.
Could you please provide clarification on how to resolve this issue?
Thank you!
### Reproduction
To get the schema I downloaded the parquet files and used pyarrow.parquet to read the schema
```
import pyarrow.parquet
pyarrow.parquet.read_schema(<path_to_parquet>, memory_map=True)
```
### Logs
_No response_
### System info
```shell
- huggingface_hub version: 0.14.1
- Platform: Linux-5.19.0-41-generic-x86_64-with-glibc2.35
- Python version: 3.10.6
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Token path ?: /home/philippe/.cache/huggingface/token
- Has saved token ?: True
- Who am I ?: philippemo
- Configured git credential helpers: cache
- FastAI: N/A
- Tensorflow: N/A
- Torch: N/A
- Jinja2: 3.1.2
- Graphviz: N/A
- Pydot: N/A
- Pillow: 9.4.0
- hf_transfer: N/A
- gradio: N/A
- ENDPOINT: https://huggingface.co
- HUGGINGFACE_HUB_CACHE: /home/philippe/.cache/huggingface/hub
- HUGGINGFACE_ASSETS_CACHE: /home/philippe/.cache/huggingface/assets
- HF_TOKEN_PATH: /home/philippe/.cache/huggingface/token
- HF_HUB_OFFLINE: False
- HF_HUB_DISABLE_TELEMETRY: False
- HF_HUB_DISABLE_PROGRESS_BARS: None
- HF_HUB_DISABLE_SYMLINKS_WARNING: False
- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False
- HF_HUB_DISABLE_IMPLICIT_TOKEN: False
- HF_HUB_ENABLE_HF_TRANSFER: False
```
| 5,869 |
https://github.com/huggingface/datasets/issues/5868 | Is it possible to change a cached file and 're-cache' it instead of re-generating? | [
"Arrow files/primitives (tables and arrays) are immutable, so re-generating them is the only option, I'm afraid.",
"> \r\n\r\nGot it, thanks for your reply"
] | ### Feature request
Hi,
I have a huge cached file using `map`(over 500GB), and I want to change an attribution of each element, is there possible to do it using some method instead of re-generating, because `map` takes over 24 hours
### Motivation
For large datasets, I think it is very important because we always face the problem which is changing something in the original cache without re-generating it.
### Your contribution
For now, I can't help, sorry. | 5,868 |
https://github.com/huggingface/datasets/issues/5866 | Issue with Sequence features | [
"Thanks for reporting! I've opened a PR with a fix."
] | ### Describe the bug
Sequences features sometimes causes errors when the specified length is not -1
### Steps to reproduce the bug
```python
import numpy as np
from datasets import Features, ClassLabel, Sequence, Value, Dataset
feats = Features(**{'target': ClassLabel(names=[0, 1]),'x': Sequence(feature=Value(dtype='float64',id=None), length=2, id=None)})
Dataset.from_dict({"target": np.ones(2000).astype(int), "x": np.random.rand(2000,2)},features = feats).flatten_indices()
```
Throws:
```
TypeError: Couldn't cast array of type
fixed_size_list<item: double>[2]
to
Sequence(feature=Value(dtype='float64', id=None), length=2, id=None)
```
The same code works without any issues when `length = -1`
EDIT: The error seems to happen only when the length of the dataset is bigger than 1000 for some reason
### Expected behavior
No exception
### Environment info
- `datasets` version: 2.10.1
- Python version: 3.9.5
- PyArrow version: 11.0.0
- Pandas version: 1.4.1 | 5,866 |
https://github.com/huggingface/datasets/issues/5864 | Slow iteration over Torch tensors | [
"I am highly interested performance of dataset so I ran your example as a curious user.\r\n```python\r\ntrain_dataset.cast_column(\"x\", Array3D(shape=img_shape, dtype=\"float32\"))\r\n```\r\nhave return values and \"x\" is a new column, it shoulde be\r\n```python\r\nds=train_dataset.cast_column(\"img\", Array3D(shape=(3,32,32), dtype=\"float32\"))\r\n```\r\nI rewrite your example as\r\n```python\r\ntrain_dataset = load_dataset(\r\n 'cifar100',\r\n split='train',\r\n use_auth_token=True,\r\n)\r\ntransform_func = torchvision.transforms.Compose([\r\n ToTensor(), \r\n Normalize(mean=[0.485, 0.456, 0.406], std= [0.229, 0.224, 0.225]),] \r\n)\r\n \r\ntrain_dataset = train_dataset.map(\r\n desc=f\"Preprocessing samples\",\r\n function=lambda x: {\"img\": transform_func(x[\"img\"])},\r\n)\r\nds=train_dataset.cast_column(\"img\", Array3D(shape=(3,32,32), dtype=\"float32\"))\r\nfor i in tqdm(ds):\r\n pass\r\n```\r\nthat require ~11s in my environment. While\r\n```python\r\nds = load_dataset(\r\n 'cifar100',\r\n split='train',\r\n use_auth_token=True,\r\n)\r\n\r\nfor i in tqdm(ds):\r\n pass\r\n```\r\nonly need ~6s. (So I guess it's still undesirable)"
] | ### Describe the bug
I have a problem related to this [issue](https://github.com/huggingface/datasets/issues/5841): I get a way slower iteration when using a Torch dataloader if I use vanilla Numpy tensors or if I first apply a ToTensor transform to the input. In particular, it takes 5 seconds to iterate over the vanilla input and ~30s after the transformation.
### Steps to reproduce the bug
Here is the minimum code to reproduce the problem
```python
import numpy as np
from datasets import Dataset, DatasetDict, load_dataset, Array3D, Image, Features
from torch.utils.data import DataLoader
from tqdm import tqdm
import torchvision
from torchvision.transforms import ToTensor, Normalize
#################################
# Without transform
#################################
train_dataset = load_dataset(
'cifar100',
split='train',
use_auth_token=True,
)
train_dataset.set_format(type="numpy", columns=["img", "fine_label"])
train_loader= DataLoader(
train_dataset,
batch_size=100,
pin_memory=False,
shuffle=True,
num_workers=8,
)
for batch in tqdm(train_loader, desc="Loading data, no transform"):
pass
#################################
# With transform
#################################
transform_func = torchvision.transforms.Compose([
ToTensor(),
Normalize(mean=[0.485, 0.456, 0.406], std= [0.229, 0.224, 0.225]),]
)
train_dataset = train_dataset.map(
desc=f"Preprocessing samples",
function=lambda x: {"img": transform_func(x["img"])},
)
train_dataset.set_format(type="numpy", columns=["img", "fine_label"])
train_loader= DataLoader(
train_dataset,
batch_size=100,
pin_memory=False,
shuffle=True,
num_workers=8,
)
for batch in tqdm(train_loader, desc="Loading data after transform"):
pass
```
I have also tried converting the Image column to an Array3D
```python
img_shape = train_dataset[0]["img"].shape
features = train_dataset.features.copy()
features["x"] = Array3D(shape=img_shape, dtype="float32")
train_dataset = train_dataset.map(
desc=f"Preprocessing samples",
function=lambda x: {"x": np.array(x["img"], dtype=np.uint8)},
features=features,
)
train_dataset.cast_column("x", Array3D(shape=img_shape, dtype="float32"))
train_dataset.set_format(type="numpy", columns=["x", "fine_label"])
```
but to no avail. Any clue?
### Expected behavior
The iteration should take approximately the same time with or without the transformation, as it doesn't change the shape of the input. What may be the issue here?
### Environment info
```
- `datasets` version: 2.12.0
- Platform: Linux-5.4.0-137-generic-x86_64-with-glibc2.31
- Python version: 3.9.16
- Huggingface_hub version: 0.14.1
- PyArrow version: 12.0.0
- Pandas version: 2.0.1
``` | 5,864 |
https://github.com/huggingface/datasets/issues/5862 | IndexError: list index out of range with data hosted on Zenodo | [
"This error is also raised when data is hosted on Google Drive:\r\n- https://huggingface.co/datasets/docred/discussions/5\r\n- https://huggingface.co/datasets/linnaeus/discussions/3\r\n- https://huggingface.co/datasets/poleval2019_mt/discussions/3\r\n- https://huggingface.co/datasets/reddit_tifu/discussions/2\r\n- https://huggingface.co/datasets/species_800/discussions/3\r\n- https://huggingface.co/datasets/wiki_lingua/discussions/1\r\n- https://huggingface.co/datasets/yoruba_text_c3/discussions/1"
] | The dataset viewer sometimes raises an `IndexError`:
```
IndexError: list index out of range
```
See:
- huggingface/datasets-server#1151
- https://huggingface.co/datasets/reddit/discussions/5
- huggingface/datasets-server#1118
- https://huggingface.co/datasets/krr-oxford/OntoLAMA/discussions/1
- https://huggingface.co/datasets/hyperpartisan_news_detection/discussions/3
- https://huggingface.co/datasets/um005/discussions/2
- https://huggingface.co/datasets/tapaco/discussions/2
- https://huggingface.co/datasets/common_language/discussions/3
- https://huggingface.co/datasets/pass/discussions/1
After investigation:
- This happens with data files hosted on Zenodo
- Indeed, there is an underlying 429 HTTP error: Too Many Requests
Note that some time ago, it also happened with data files hosted on Google Drive. See:
- #4581
- #4580
The reason then was that there was a 403 HTTP error: Forbidden
| 5,862 |
https://github.com/huggingface/datasets/issues/5858 | Throw an error when dataset improperly indexed | [
"Thanks for reporting, @sarahwie.\r\n\r\nPlease note that in `datasets` we do not have vectorized operation like `pandas`. Therefore, your equality comparisons above are `False`:\r\n- For example: `squad['question']` returns a `list`, and this list is not equal to `\"Who was the Norse leader?\"`\r\n\r\nThe `False` value is equivalent to `0` when indexing a dataset, thus the reason why you get the first element (with index 0): \r\n- For example: `squad[False]` is equivalent to `squad[0]`\r\n\r\nMaybe we should an exception instead of assuming that `False` is equivalent to `0` (and `True` is equivalent to `1`) in the context of indexing."
] | ### Describe the bug
Pandas-style subset indexing on dataset does not throw an error, when maybe it should. Instead returns the first instance of the dataset regardless of index condition.
### Steps to reproduce the bug
Steps to reproduce the behavior:
1. `squad = datasets.load_dataset("squad_v2", split="validation")`
2. `item = squad[squad['question'] == "Who was the Norse leader?"]`
or `it = squad[squad['id'] == '56ddde6b9a695914005b962b']`
3. returns the first item in the dataset, which does not satisfy the above conditions:
`{'id': '56ddde6b9a695914005b9628', 'title': 'Normans', 'context': 'The Normans (Norman: Nourmands; French: Normands; Latin: Normanni) were the people who in the 10th and 11th centuries gave their name to Normandy, a region in France. They were descended from Norse ("Norman" comes from "Norseman") raiders and pirates from Denmark, Iceland and Norway who, under their leader Rollo, agreed to swear fealty to King Charles III of West Francia. Through generations of assimilation and mixing with the native Frankish and Roman-Gaulish populations, their descendants would gradually merge with the Carolingian-based cultures of West Francia. The distinct cultural and ethnic identity of the Normans emerged initially in the first half of the 10th century, and it continued to evolve over the succeeding centuries.', 'question': 'In what country is Normandy located?', 'answers': {'text': ['France', 'France', 'France', 'France'], 'answer_start': [159, 159, 159, 159]}}`
### Expected behavior
Should either throw an error message, or return the dataset item that satisfies the condition.
### Environment info
- `datasets` version: 2.9.0
- Platform: macOS-13.3.1-arm64-arm-64bit
- Python version: 3.10.8
- PyArrow version: 10.0.1
- Pandas version: 1.5.3 | 5,858 |
https://github.com/huggingface/datasets/issues/5857 | Adding chemistry dataset/models in huggingface | [
"Hi! \r\n\r\nThis would be a nice addition to the Hub! You can find the existing chemistry datasets/models on the Hub (using the `chemistry` tag) [here](https://huggingface.co/search/full-text?q=chemistry&type=model&type=dataset).\r\n\r\nFeel free to ping us here on the Hub if you need help adding the datasets.\r\n"
] | ### Feature request
Huggingface is really amazing platform for open science.
In addition to computer vision, video and NLP, would it be of interest to add chemistry/materials science dataset/models in Huggingface? Or, if its already done, can you provide some pointers.
We have been working on a comprehensive benchmark on this topic: [JARVIS-Leaderboard](https://pages.nist.gov/jarvis_leaderboard/) and I am wondering if we could contribute/integrate this project as a part of huggingface.
### Motivation
Similar to the main stream AI field, there is need of large scale benchmarks/models/infrastructure for chemistry/materials data.
### Your contribution
We can start adding datasets as our [benchmarks](https://github.com/usnistgov/jarvis_leaderboard/tree/main/jarvis_leaderboard/benchmarks) should be easily convertible to the dataset format. | 5,857 |
https://github.com/huggingface/datasets/issues/5856 | Error loading natural_questions | [
"Hi! You can avoid this error by using the preprocessed version:\r\n```python\r\nimport datasets\r\nds = datasets.load_dataset('natural_questions')\r\n```\r\n\r\nPS: Once we finish https://github.com/huggingface/datasets/pull/5364, this error will no longer be a problem.",
"> Hi! You can avoid this error by using the preprocessed version:\r\n> \r\n> ```python\r\n> import datasets\r\n> ds = datasets.load_dataset('natural_questions')\r\n> ```\r\n> \r\n> PS: Once we finish #5364, this error will no longer be a problem.\r\n\r\nThanks, wish #5364 finish early"
] | ### Describe the bug
When try to load natural_questions through datasets == 2.12.0 with python == 3.8.9:
```python
import datasets
datasets.load_dataset('natural_questions',beam_runner='DirectRunner')
```
It failed with following info:
`pyarrow.lib.ArrowNotImplementedError: Nested data conversions not implemented for chunked array outputs`
### Steps to reproduce the bug
In python console:
```python
import datasets
datasets.load_dataset('natural_questions',beam_runner='DirectRunner')
```
Then the trace is:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/nlp/.cache/pypoetry/virtualenvs/drg-W3LF4Ol9-py3.8/lib/python3.8/site-packages/datasets/load.py", line 1797, in load_dataset
builder_instance.download_and_prepare(
File "/home/nlp/.cache/pypoetry/virtualenvs/drg-W3LF4Ol9-py3.8/lib/python3.8/site-packages/datasets/builder.py", line 890, in download_and_prepare
self._download_and_prepare(
File "/home/nlp/.cache/pypoetry/virtualenvs/drg-W3LF4Ol9-py3.8/lib/python3.8/site-packages/datasets/builder.py", line 2019, in _download_and_prepare
num_examples, num_bytes = beam_writer.finalize(metrics.query(m_filter))
File "/home/nlp/.cache/pypoetry/virtualenvs/drg-W3LF4Ol9-py3.8/lib/python3.8/site-packages/datasets/arrow_writer.py", line 694, in finalize
shard_num_bytes, _ = parquet_to_arrow(source, destination)
File "/home/nlp/.cache/pypoetry/virtualenvs/drg-W3LF4Ol9-py3.8/lib/python3.8/site-packages/datasets/arrow_writer.py", line 737, in parquet_to_arrow
for record_batch in parquet_file.iter_batches():
File "pyarrow/_parquet.pyx", line 1323, in iter_batches
File "pyarrow/error.pxi", line 121, in pyarrow.lib.check_status
pyarrow.lib.ArrowNotImplementedError: Nested data conversions not implemented for chunked array outputs
```
### Expected behavior
load natural_question questions
### Environment info
```
- `datasets` version: 2.12.0
- Platform: Linux-3.10.0-1160.42.2.el7.x86_64-x86_64-with-glibc2.2.5
- Python version: 3.8.9
- Huggingface_hub version: 0.14.1
- PyArrow version: 11.0.0
- Pandas version: 2.0.1
``` | 5,856 |
https://github.com/huggingface/datasets/issues/5855 | `to_tf_dataset` consumes too much memory | [
"Cc @amyeroberts @Rocketknight1 \r\n\r\nIndded I think it's because it does something like this under the hood when there's no multiprocessing:\r\n\r\n```python\r\ntf_dataset = tf_dataset.shuffle(len(dataset))\r\n```\r\n\r\nPS: with multiprocessing it appears to be different:\r\n\r\n```python\r\nindices = np.arange(len(dataset))\r\nif shuffle:\r\n np.random.shuffle(indices)\r\n```",
"Hi @massquantity, the dataset being shuffled there is not the full dataset. If you look at [the line above](https://github.com/huggingface/datasets/blob/main/src/datasets/utils/tf_utils.py#L182), the dataset is actually just a single indices array at that point, and that array is the only thing that gets fully loaded into memory and shuffled. We then load samples from the dataset by applying a transform function to the shuffled dataset, which fetches samples based on the indices it receives.\r\n\r\nIf your dataset is **really** gigantic, then this index tensor might be a memory issue, but since it's just an int64 tensor it will only use 1GB of memory per 125 million samples.\r\n\r\nStill, if you're encountering memory issues, there might be another cause here - can you share some code to reproduce the error, or does it depend on some internal/proprietary dataset?",
"Hi @Rocketknight1, you're right and I also noticed that only indices are used in shuffling. My data has shape (50000000, 10), but really the problem doesn't relate to a specific dataset. Simply running the following code costs me 10GB of memory.\r\n\r\n```python\r\nfrom datasets import Dataset\r\n\r\ndef gen():\r\n for i in range(50000000):\r\n yield {\"data\": i}\r\n\r\nds = Dataset.from_generator(gen, cache_dir=\"./huggingface\")\r\n\r\ntf_ds = ds.to_tf_dataset(\r\n batch_size=1,\r\n shuffle=True,\r\n drop_remainder=False,\r\n prefetch=True,\r\n)\r\ntf_ds = iter(tf_ds)\r\nnext(tf_ds)\r\n# {'data': <tf.Tensor: shape=(1,), dtype=int64, numpy=array([0])>}\r\n```\r\n\r\nI just realized maybe it was an issue from tensorflow (I'm using tf 2.12). So I tried the following code, and it used 10GB of memory too.\r\n```python\r\nimport numpy as np\r\nimport tensorflow as tf\r\n\r\ndata_size = 50000000\r\ntf_dataset = tf.data.Dataset.from_tensor_slices(np.arange(data_size))\r\ntf_dataset = iter(tf_dataset.shuffle(data_size))\r\nnext(tf_dataset)\r\n# <tf.Tensor: shape=(), dtype=int64, numpy=24774043>\r\n```\r\n\r\nBy the way, as @lhoestq mentioned, multiprocessing uses numpy shuffling, and it uses less than 1 GB of memory:\r\n```python\r\ntf_ds_mp = ds.to_tf_dataset(\r\n batch_size=1,\r\n shuffle=True,\r\n drop_remainder=False,\r\n prefetch=True,\r\n num_workers=2,\r\n)\r\n```",
"Thanks for that reproduction script - I've confirmed the same issue is occurring for me. Investigating it now!",
"Update: The memory usage is occurring in creation of the index and shuffle buffer. You can reproduce it very simply with:\r\n\r\n```python\r\nimport tensorflow as tf\r\nindices = tf.range(50_000_000, dtype=tf.int64)\r\ndataset = tf.data.Dataset.from_tensor_slices(indices)\r\ndataset = dataset.shuffle(len(dataset))\r\nprint(next(iter(dataset))\r\n```\r\nWhen I wrote this code I thought `tf.data` had an optimization for shuffling an entire tensor that wouldn't create the entire shuffle buffer, but evidently it's just creating the enormous buffer in memory. I'll see if I can find a more efficient way to do this - we might end up moving everything to the `numpy` multiprocessing path to avoid it.",
"I opened a PR to fix this - will continue the discussion there!"
] | ### Describe the bug
Hi, I'm using `to_tf_dataset` to convert a _large_ dataset to `tf.data.Dataset`. I observed that the data loading *before* training took a lot of time and memory, even with `batch_size=1`.
After some digging, i believe the reason lies in the shuffle behavior. The [source code](https://github.com/huggingface/datasets/blob/main/src/datasets/utils/tf_utils.py#L185) uses `len(dataset)` as the `buffer_size`, which may load all the data into the memory, and the [tf.data doc](https://www.tensorflow.org/guide/data#randomly_shuffling_input_data) also states that "While large buffer_sizes shuffle more thoroughly, they can take a lot of memory, and significant time to fill".
### Steps to reproduce the bug
```python
from datasets import Dataset
def gen(): # some large data
for i in range(50000000):
yield {"data": i}
ds = Dataset.from_generator(gen, cache_dir="./huggingface")
tf_ds = ds.to_tf_dataset(
batch_size=64,
shuffle=False, # no shuffle
drop_remainder=False,
prefetch=True,
)
# fast and memory friendly 🤗
for batch in tf_ds:
...
tf_ds_shuffle = ds.to_tf_dataset(
batch_size=64,
shuffle=True,
drop_remainder=False,
prefetch=True,
)
# slow and memory hungry for simple iteration 😱
for batch in tf_ds_shuffle:
...
```
### Expected behavior
Shuffling should not load all the data into the memory. Would adding a `buffer_size` parameter in the `to_tf_dataset` API alleviate the problem?
### Environment info
- `datasets` version: 2.11.0
- Platform: Linux-5.17.1-051701-generic-x86_64-with-glibc2.17
- Python version: 3.8.13
- Huggingface_hub version: 0.13.4
- PyArrow version: 11.0.0
- Pandas version: 1.4.3
| 5,855 |
https://github.com/huggingface/datasets/issues/5854 | Can not load audiofolder dataset on kaggle | [
"Hi! `audiofolder` requires `datasets>=2.5.0`, so please update the `datasets`' installation (`pip install -U datasets`) in the environment (and restart the env for the update to take effect) to resolve the issue.",
"> Hi! `audiofolder` requires `datasets>=2.5.0`, so please update the `datasets`' installation (`pip install -U datasets`) in the environment to resolve the issue.\r\n\r\nI don't think it is a problem of the version. It runs ok on colab or local machine. Only on kaggle will has this bug.",
"Based on your dataset info, the installed version is `2.1.0`, which does not include `audiofolder`.\r\n\r\nBy default, Kaggle preinstalls `datasets` into a new env, but the version it installs is outdated and does not contain newer features such as `audiofolder`"
] | ### Describe the bug
It's crash log:
FileNotFoundError: Couldn't find a dataset script at /kaggle/working/audiofolder/audiofolder.py or any data file in the same directory. Couldn't find 'audiofolder' on the Hugging Face Hub either: FileNotFoundError: Couldn't find file at https://raw.githubusercontent.com/huggingface/datasets/master/datasets/audiofolder/audiofolder.py
### Steps to reproduce the bug
common_voice = load_dataset("audiofolder", data_dir="/kaggle/working/data")
### Expected behavior
load dataset without error. It works ok on colab, but on kaggle it happends.
### Environment info
- `datasets` version: 2.1.0
- Platform: Linux-5.15.109+-x86_64-with-glibc2.31
- Python version: 3.10.10
- PyArrow version: 9.0.0
- Pandas version: 1.5.3 | 5,854 |
https://github.com/huggingface/datasets/issues/5849 | CSV datasets should only read the CSV data files in the repo | [] | When a no-script dataset has many CSV files and a JPG file, the library infers to use the Csv builder, but tries to read as CSV all files in the repo, also the JPG file.
I think the Csv builder should filter out non-CSV files when reading.
An analogue solution should be implemented for other packaged builders.
Related to:
- https://huggingface.co/datasets/abidlabs/img2text/discussions/1
- https://github.com/gradio-app/gradio/pull/3973#issuecomment-1545409061
CC: @abidlabs @severo | 5,849 |
https://github.com/huggingface/datasets/issues/5847 | Streaming IterableDataset not working with translation pipeline | [
"I wasn't sure to file this against transformers or datasets.",
"[`KeyDataset`](https://github.com/huggingface/transformers/blob/7f8b909189547944617741d8d3c6c84504701693/src/transformers/pipelines/pt_utils.py#L296) doesn't support iterable datasets, so you either need to implement a version that does (and also indexing nested (translation) fields):\r\n\r\n```python\r\nfrom torch.utils.data import Dataset, IterableDataset\r\n\r\ndef build_key_fetcher(key: str):\r\n def _key_fetcher(item):\r\n for sub_key in key.split(\".\"):\r\n item = item[sub_key]\r\n return item\r\n return _key_fetcher\r\n\r\nclass KeyDataset(Dataset):\r\n def __new__(cls, dataset: Dataset, key: str):\r\n cls = _KeyIterableDataset if isinstance(dataset, IterableDataset) else _KeyMapDataset\r\n self = object.__new__(cls)\r\n self.dataset = dataset\r\n self.key = key\r\n self._key_fetcher = build_key_fetcher(key)\r\n return self\r\n\r\nclass _KeyMapDataset(KeyDataset):\r\n def __getitem__(self, i):\r\n return self._key_fetcher(self.dataset[i])\r\n \r\n def __len__(self):\r\n return len(self.dataset)\r\n\r\n\r\nclass _KeyIterableDataset(KeyDataset):\r\n def __iter__(self):\r\n for ex in self.dataset:\r\n yield self._key_fetcher(ex)\r\n\r\nks = KeyDataset(ds, \"translation.en\")\r\n```\r\n\r\nor use `IterableDataset`'s `map`:\r\n```python\r\ndef fetch_en_translation(ex):\r\n return {\"en\": ex[\"translation\"][\"en\"]}\r\nks = ds.map(fetch_en_translation, remove_columns=ds.column_names) \r\n```\r\n\r\ncc @sgugger: Perhaps the `KeyDataset` + PyTorch `IterableDataset` case should be supported by Transformers",
"@mariosasko The map snippet didn't quite work, but gave me enough of a clue to get it working. The following snippet does work:\r\n```\r\ndef en_translation(x):\r\n return {\"en\":x['translation']['en']}\r\nks = ds.map(en_translation, remove_columns=['translation'])\r\ntest=[]\r\nfor x in iter(ks):\r\n test.append(x['en'])\r\nxx= mt(test)\r\nfor x in xx:\r\n print(x)\r\n```\r\n\r\nI tried just returning `x['translation']['en`]` in the helper function instead of the dict, but that didn't give me an iterator over strings that pipeline would work with either.\r\n\r\n\r\nThe snippet as is gives the following error:\r\n```\r\nTraceback (most recent call last):\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/pdb.py\", line 1704, in main\r\n pdb._runscript(mainpyfile)\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/pdb.py\", line 1573, in _runscript\r\n self.run(statement)\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/bdb.py\", line 580, in run\r\n exec(cmd, globals, locals)\r\n File \"<string>\", line 1, in <module>\r\n File \"/home/jlquinn/models/hf/ende.t5.pipe.py\", line 1, in <module>\r\n from transformers import pipeline\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/text2text_generation.py\", line 335, in __call__\r\n return super().__call__(*args, **kwargs)\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/text2text_generation.py\", line 138, in __call__\r\n result = super().__call__(*args, **kwargs)\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/base.py\", line 1027, in __call__\r\n return self.run_single(inputs, preprocess_params, forward_params, postprocess_params)\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/base.py\", line 1033, in run_single\r\n model_inputs = self.preprocess(inputs, **preprocess_params)\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/text2text_generation.py\", line 287, in preprocess\r\n return super()._parse_and_tokenize(*args, truncation=truncation)\r\n File \"/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/text2text_generation.py\", line 100, in _parse_and_tokenize\r\n raise ValueError(\r\nValueError: `args[0]`: <datasets.iterable_dataset.IterableDataset object at 0x7f5fd38ef1c0> have the wrong format. The should be either of type `str` or type `list`\r\nUncaught exception. Entering post mortem debugging\r\nRunning 'cont' or 'step' will restart the program\r\n```\r\n",
"So perhaps there's no bug exactly, but I would love to see two things: 1) improve the documentation to better understand what's really getting returned. 2) update the example provided of using transformer pipeline with a dataset to include the oddball case that translation appears to be.",
"cc @Narsil ",
"Hi,\r\n\r\nfor the original snippet, the issue is that `streaming` datasets are not countable (they have no len) and therefore `KeyDataset` cannot work with them ( KeyDataset is a dataset and therefore requires a length).\r\n\r\nI modified slightly the original snippet to make it work:\r\n\r\n```python\r\nfrom transformers import pipeline\r\nfrom transformers.pipelines.pt_utils import KeyDataset\r\nfrom datasets import load_dataset\r\n\r\nds = load_dataset(path=\"wmt14\", name=\"fr-en\", split=\"test\", streaming=True)\r\nbs = 1\r\nmt = pipeline(\r\n \"translation_en_to_fr\", model=\"hf-internal-testing/tiny-random-T5ForConditionalGeneration\", batch_size=bs\r\n)\r\n\r\n\r\ndef ks(ds):\r\n for item in ds:\r\n yield item[\"translation\"][\"en\"]\r\n\r\n\r\n# print(f\"{ks}\")\r\nxx = mt(ks(ds))\r\nfor x in xx:\r\n print(x)\r\n```\r\n\r\nThis is what the first example in the docs suggests to use (as it's the most flexible): https://huggingface.co/docs/transformers/v4.29.1/en/pipeline_tutorial#using-pipelines-on-a-dataset\r\n\r\n`KeyDataset` really exists only to get a `sized` dataset to work nicer with `tqdm` for instance.\r\n\r\n@sgugger should we update the docs to remove `KeyDataset` entirely ? (We can add a note to pass manually the length of the data to tqdm so that the progress bar option can still be easy to use ?)\r\n",
"Maybe moving `KeyDataset` later on in the guide and specify it's mostly for streaming then? Or is it also necessary for batch_size>1 (which is what the current doc implies)?",
"Hmm\r\n\r\nIterator (`yield`) :\r\n- Not countable\r\n- Super flexible\r\n- Cannot use `num_workers>1` (threading requires indexing at the correct location, iterators require to iterate in order,so each thread would iterate over the full thing being genuinely a bad idea)\r\n- Can batch\r\n- tqdm doesn't show a nice progress bar (it has no total)\r\n\r\nKeyDataset (Or any PyTorch like Dataset returning the correct object for the pipeline):\r\n- Countable\r\n- Less flexible (not applicable to datasets with streaming), can only work on single keys. But should be easy to read and write your own (like @mariosasko did)\r\n- Works with `num_workers > 1` (Every worker can fetch exactly what's needed)\r\n- Can batch \r\n- tqdm shows a nice progress bar\r\n\r\nIn the docs, if we update all the examples to use iterators, and include an example with\r\n\r\n```\r\nfor item in tqdm.tqdm(pipe(iterator(), total=len(dataset))))\r\n```\r\n\r\nWe can save the biggest feature that doesn't work out of the box with iterators which is the tqdm progress bar.\r\n\r\n`num_workers>1` we can mention it, but it tends to be an issues only on CPU intensive loads, like image (and maybe audio)\r\n"
] | ### Describe the bug
I'm trying to use a streaming dataset for translation inference to avoid downloading the training data.
I'm using a pipeline and a dataset, and following the guidance in the tutorial.
Instead I get an exception that IterableDataset has no len().
### Steps to reproduce the bug
CODE:
```
from transformers import pipeline
from transformers.pipelines.pt_utils import KeyDataset
from datasets import load_dataset
ds = load_dataset(path="wmt14", name="fr-en", split="test", streaming=True)
bs=1
mt = pipeline("translation_en_to_fr", model="t5-base", batch_size=bs)
#print(mt("hello")) THIS WORKS
ks = KeyDataset(ds, "translation")
print(f"{ks}")
xx= mt(ks)
for x in xx:
print(x)
```
RUN:
```
(watnlp) [jlquinn@bertdev01 hf]$ python ende.t5.pipe.py
2023-05-11 16:48:08.817572: I tensorflow/core/util/util.cc:169] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2023-05-11 16:48:08.821388: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libcudart.so.11.0'; dlerror: libcudart.so.11.0: cannot open shared object file: No such file or directory
2023-05-11 16:48:08.821407: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.
<transformers.pipelines.pt_utils.KeyDataset object at 0x7f61ed5da9d0>
Traceback (most recent call last):
File "/home/jlquinn/models/hf/ende.t5.pipe.py", line 11, in <module>
for x in xx:
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/pt_utils.py", line 111, in __next__
item = next(self.iterator)
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/pt_utils.py", line 111, in __next__
item = next(self.iterator)
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 681, in __next__
data = self._next_data()
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 720, in _next_data
index = self._next_index() # may raise StopIteration
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 671, in _next_index
return next(self._sampler_iter) # may raise StopIteration
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/torch/utils/data/sampler.py", line 247, in __iter__
for idx in self.sampler:
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/torch/utils/data/sampler.py", line 76, in __iter__
return iter(range(len(self.data_source)))
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/pt_utils.py", line 13, in __len__
return len(self.dataset)
File "/home/jlquinn/miniconda3/envs/watnlp/lib/python3.9/site-packages/transformers/pipelines/pt_utils.py", line 289, in __len__
return len(self.dataset)
TypeError: object of type 'IterableDataset' has no len()
```
### Expected behavior
I'm expecting french translations of the english test set to be printed.
### Environment info
Run on CPU with no GPU.
RHEL 8.7 x86_64
python 3.9.0
transformers 4.17.0
datasets 2.0.0
tokenizers 0.12.1
```
(watnlp) [jlquinn@bertdev01 hf]$ datasets-cli env
Copy-and-paste the text below in your GitHub issue.
- `datasets` version: 2.0.0
- Platform: Linux-4.18.0-372.19.1.el8_6.x86_64-x86_64-with-glibc2.28
- Python version: 3.9.0
- PyArrow version: 8.0.0
- Pandas version: 1.4.4
```
| 5,847 |
https://github.com/huggingface/datasets/issues/5851 | Error message not clear in interleaving datasets | [] | ### System Info
standard env
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [X] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [X] My own task or dataset (give details below)
### Reproduction
I'm trying to interleave 'sciq', 'wiki' and the 'pile-enron' dataset. I think the error I made was that I loaded the train split of one, but for the other but the error is not too helpful-
```
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
[/home/suryahari/Vornoi/save_model_ops.py](https://vscode-remote+ssh-002dremote-002bthomsonlab-002d2-002ejamesgornet-002ecom.vscode-resource.vscode-cdn.net/home/suryahari/Vornoi/save_model_ops.py) in line 3
[41](file:///home/suryahari/Vornoi/save_model_ops.py?line=40) # %%
----> [43](file:///home/suryahari/Vornoi/save_model_ops.py?line=42) dataset = interleave_datasets(datasets, stopping_strategy="all_exhausted")
File [~/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py:124](https://vscode-remote+ssh-002dremote-002bthomsonlab-002d2-002ejamesgornet-002ecom.vscode-resource.vscode-cdn.net/home/suryahari/~/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py:124), in interleave_datasets(datasets, probabilities, seed, info, split, stopping_strategy)
[122](file:///home/suryahari/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py?line=121) for dataset in datasets[1:]:
[123](file:///home/suryahari/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py?line=122) if (map_style and not isinstance(dataset, Dataset)) or (iterable and not isinstance(dataset, IterableDataset)):
--> [124](file:///home/suryahari/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py?line=123) raise ValueError(
[125](file:///home/suryahari/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py?line=124) f"Unable to interleave a {type(datasets[0])} with a {type(dataset)}. Expected a list of Dataset objects or a list of IterableDataset objects."
[126](file:///home/suryahari/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py?line=125) )
[127](file:///home/suryahari/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py?line=126) if stopping_strategy not in ["first_exhausted", "all_exhausted"]:
[128](file:///home/suryahari/miniconda3/envs/vornoi/lib/python3.10/site-packages/datasets/combine.py?line=127) raise ValueError(f"{stopping_strategy} is not supported. Please enter a valid stopping_strategy.")
ValueError: Unable to interleave a with a . Expected a list of Dataset objects or a list of IterableDataset objects.
```
### Expected behavior
the error message should hopefully be more clear | 5,851 |
https://github.com/huggingface/datasets/issues/5846 | load_dataset('bigcode/the-stack-dedup', streaming=True) very slow! | [
"This is due to the slow resolution of the data files: https://github.com/huggingface/datasets/issues/5537.\r\n\r\nWe plan to switch to `huggingface_hub`'s `HfFileSystem` soon to make the resolution faster (will be up to 20x faster once we merge https://github.com/huggingface/huggingface_hub/pull/1443)\r\n\r\n",
"You're right, when I try to parse more than 50GB of text data, I also get very slow, usually taking hours or even tens of hours.",
"> You're right, when I try to parse more than 50GB of text data, I also get very slow, usually taking hours or even tens of hours.\r\n\r\nThat's unrelated to the problem discussed in this issue. ",
"> > You're right, when I try to parse more than 50GB of text data, I also get very slow, usually taking hours or even tens of hours.\r\n> \r\n> That's unrelated to the problem discussed in this issue.\r\n\r\nSorry, I misunderstood it.",
"Closing this issue as it has been addressed in `huggingface_hub`!\r\n\r\n(This now takes 25s to execute on my machine.)",
"Thanks for the improvements! 🎉🎉\n\n25 seconds is better but still about 2500x slower than this _should_ be! Loading a tiny 1-2KB metadata file is all that would be necessary with a better design.",
"Once we merge https://github.com/huggingface/huggingface_hub/pull/2103, this should only take a few seconds. \r\n\r\nFor the 2500x speed-up (without metadata files with pre-cached results), we wouldn't even be allowed to use `os.path` functions or `requests`/`aiohttp` for HTTP requests, so I don't think this is feasible for us as it would make the code unreadable.\r\n\r\nThe HF Datasets Hub is (almost) platform-agnostic, so you are free to implement your own library (in a faster language than Python) to achieve this kind of performance, and we would be happy to support it 🙂. "
] | ### Describe the bug
Running
```
import datasets
ds = datasets.load_dataset('bigcode/the-stack-dedup', streaming=True)
```
takes about 2.5 minutes!
I would expect this to be near instantaneous. With other datasets, the runtime is one or two seconds.
### Environment info
- `datasets` version: 2.11.0
- Platform: macOS-13.3.1-arm64-arm-64bit
- Python version: 3.10.10
- Huggingface_hub version: 0.13.4
- PyArrow version: 11.0.0
- Pandas version: 2.0.0 | 5,846 |
https://github.com/huggingface/datasets/issues/5844 | TypeError: Couldn't cast array of type struct<answer: struct<unanswerable: bool, answerType: string, free_form_answer: string, evidence: list<item: string>, evidenceAnnotate: list<item: string>, highlighted_evidence: list<item: string>>> to ... | [] | ### Describe the bug
TypeError: Couldn't cast array of type struct<answer: struct<unanswerable: bool, answerType: string, free_form_answer: string, evidence: list<item: string>, evidenceAnnotate: list<item: string>, highlighted_evidence: list<item: string>>> to {'answer': {'unanswerable': Value(dtype='bool', id=None), 'answerType': Value(dtype='string', id=None), 'free_form_answer': Value(dtype='string', id=None), 'evidence': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), 'evidenceAnnotate': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), 'highlighted_evidence': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None)}, 'unanswerable': Value(dtype='bool', id=None), 'answerType': Value(dtype='string', id=None), 'free_form_answer': Value(dtype='string', id=None), 'evidence': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), 'evidenceAnnotate': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), 'highlighted_evidence': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None)}
When I use _load_dataset()_ I get the error
`from datasets import load_dataset
datafiles = {'train': './data/train.json', 'validation': './data/validation.json', 'test': './data/test.json'}
raw_data = load_dataset("json", data_files=datafiles, cache_dir="./cache")
`
Detailed error information is as follows：
Traceback (most recent call last):
File "C:/Users/CHENJIALEI/Desktop/NLPCC2023/NLPCC23_SciMRC-main/test2.py", line 9, in <module>
raw_data = load_dataset("json", data_files=datafiles, cache_dir="./cache")
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\load.py", line 1747, in load_dataset
builder_instance.download_and_prepare(
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\builder.py", line 814, in download_and_prepare
self._download_and_prepare(
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\builder.py", line 905, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\builder.py", line 1521, in _prepare_split
writer.write_table(table)
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\arrow_writer.py", line 540, in write_table
pa_table = table_cast(pa_table, self._schema)
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 2069, in table_cast
return cast_table_to_schema(table, schema)
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 2031, in cast_table_to_schema
arrays = [cast_array_to_feature(table[name], feature) for name, feature in features.items()]
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 2031, in <listcomp>
arrays = [cast_array_to_feature(table[name], feature) for name, feature in features.items()]
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1740, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1740, in <listcomp>
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1867, in cast_array_to_feature
casted_values = _c(array.values, feature[0])
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1742, in wrapper
return func(array, *args, **kwargs)
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1862, in cast_array_to_feature
arrays = [_c(array.field(name), subfeature) for name, subfeature in feature.items()]
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1862, in <listcomp>
arrays = [_c(array.field(name), subfeature) for name, subfeature in feature.items()]
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1742, in wrapper
return func(array, *args, **kwargs)
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1867, in cast_array_to_feature
casted_values = _c(array.values, feature[0])
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1742, in wrapper
return func(array, *args, **kwargs)
File "D:\Environment\anaconda3\envs\test\lib\site-packages\datasets\table.py", line 1913, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{array.type}\nto\n{feature}")
It is successful when I load the data separately
`raw_data = load_dataset("json", data_files="./data/train.json", cache_dir="./cache")`
### Steps to reproduce the bug
1.from datasets import load_dataset
2.datafiles = {'train': './data/train.json', 'validation': './data/validation.json', 'test': './data/test.json'}
3.raw_data = load_dataset("json", data_files=datafiles, cache_dir="./cache")
### Expected behavior
Successfully load dataset
### Environment info
datasets == 2.6.1
pyarrow == 8.0.0
python == 3.8
platform:windows11 | 5,844 |
https://github.com/huggingface/datasets/issues/5841 | Abusurdly slow on iteration | [
"Hi ! You can try to use the [Image](https://huggingface.co/docs/datasets/v2.12.0/en/package_reference/main_classes#datasets.Image) type which [decodes images on-the-fly](https://huggingface.co/docs/datasets/v2.12.0/en/about_dataset_features#image-feature) into pytorch tensors :)\r\n\r\n```python\r\nds = Dataset.from_dict({\"tensor\":a}).with_format(\"torch\")\r\n%time sum(1 for _ in ds)\r\n# CPU times: user 5.04 s, sys: 96.5 ms, total: 5.14 s\r\n# Wall time: 5.14 s\r\n# 10000\r\n```\r\n\r\n```python\r\nfeatures = Features({\"tensor\": Image()})\r\nds = Dataset.from_dict({\"tensor\":a}, features=features).with_format(\"torch\")\r\n%time sum(1 for _ in ds)\r\n# CPU times: user 1.86 s, sys: 49 ms, total: 1.91 s\r\n# Wall time: 1.9 s\r\n# 10000\r\n```\r\n\r\n-> Speed x2.7\r\n\r\nAnd if you want to keep using arrays of integers, consider using the [Array2D](https://huggingface.co/docs/datasets/v2.12.0/en/package_reference/main_classes#datasets.Array2D) or [Array3D](https://huggingface.co/docs/datasets/v2.12.0/en/package_reference/main_classes#datasets.Array3D) types which are even faster (since it doesn't decode images):\r\n\r\n```python\r\nfeatures = Features({\"tensor\": Array2D(shape=(100, 224), dtype=\"float32\")})\r\nds = Dataset.from_dict({\"tensor\":a}, features=features).with_format(\"torch\")\r\n%time sum(1 for _ in ds)\r\n# CPU times: user 828 ms, sys: 68.4 ms, total: 896 ms\r\n# Wall time: 897 ms\r\n# 10000\r\n```\r\n\r\n-> Speed x5.7\r\n\r\nBatching also speeds up a lot\r\n\r\n```python\r\nfrom torch.utils.data import DataLoader\r\ndl = DataLoader(ds, batch_size=100)\r\n%time sum(1 for _ in dl)\r\n# CPU times: user 564 ms, sys: 83.5 ms, total: 648 ms\r\n# Wall time: 579 ms\r\n# 100\r\n```\r\n\r\n-> Speed x8.9\r\n\r\n```python\r\n%time sum(1 for _ in ds.iter(batch_size=100))\r\n# CPU times: user 119 ms, sys: 96.8 ms, total: 215 ms\r\n# Wall time: 117 ms\r\n# 100\r\n```\r\n\r\n-> Speed x46",
"Anyway, regarding the speed difference between numpy and pytorch, I think the issue is that we first convert numpy sub-arrays to pytorch and then consolidate into one tensor, while we should to the opposite. Indeed converting a numpy array to pytorch has a fix cost that seems to cause a slow down. The current pipeline is\r\n\r\n```\r\narrow -> nested numpy arrays -> lists of torch tensors -> one torch tensor\r\n```\r\n\r\nand we should do\r\n\r\n```\r\narrow -> nested numpy arrays -> one numpy array -> one torch tensor\r\n```",
"I have a similar issue: iterating over a dataset takes 5s without applying any transform, but takes ~30s after applying a transform.\r\nHere is the minimum code to reproduce the problem\r\n\r\n```python\r\nimport numpy as np\r\nfrom datasets import Dataset, DatasetDict, load_dataset, Array3D, Image, Features\r\nfrom torch.utils.data import DataLoader\r\nfrom tqdm import tqdm\r\nimport torchvision \r\nfrom torchvision.transforms import ToTensor, Normalize\r\n\r\n\r\n#################################\r\n# Without transform\r\n#################################\r\n \r\ntrain_dataset = load_dataset(\r\n 'cifar100',\r\n split='train',\r\n use_auth_token=True,\r\n)\r\n\r\ntrain_dataset.set_format(type=\"numpy\", columns=[\"img\", \"fine_label\"])\r\n\r\ntrain_loader= DataLoader(\r\n train_dataset,\r\n batch_size=100,\r\n pin_memory=False,\r\n shuffle=True,\r\n num_workers=8,\r\n)\r\n\r\nfor batch in tqdm(train_loader, desc=\"Loading data, no transform\"):\r\n pass\r\n\r\n\r\n#################################\r\n# With transform\r\n#################################\r\n\r\ntransform_func = torchvision.transforms.Compose([\r\n ToTensor(), \r\n Normalize(mean=[0.485, 0.456, 0.406], std= [0.229, 0.224, 0.225]),] \r\n)\r\n \r\ntrain_dataset = train_dataset.map(\r\n desc=f\"Preprocessing samples\",\r\n function=lambda x: {\"img\": transform_func(x[\"img\"])},\r\n)\r\n\r\ntrain_dataset.set_format(type=\"numpy\", columns=[\"img\", \"fine_label\"])\r\n\r\n\r\ntrain_loader= DataLoader(\r\n train_dataset,\r\n batch_size=100,\r\n pin_memory=False,\r\n shuffle=True,\r\n num_workers=8,\r\n)\r\n\r\n\r\nfor batch in tqdm(train_loader, desc=\"Loading data after transform\"):\r\n pass \r\n```\r\n\r\nI have also tried converting the Image column to an Array3D\r\n```python\r\nimg_shape = train_dataset[0][\"img\"].shape\r\n\r\nfeatures = train_dataset.features.copy()\r\nfeatures[\"x\"] = Array3D(shape=img_shape, dtype=\"float32\")\r\n\r\ntrain_dataset = train_dataset.map(\r\n desc=f\"Preprocessing samples\",\r\n function=lambda x: {\"x\": np.array(x[\"img\"], dtype=np.uint8)},\r\n features=features,\r\n)\r\ntrain_dataset.cast_column(\"x\", Array3D(shape=img_shape, dtype=\"float32\"))\r\ntrain_dataset.set_format(type=\"numpy\", columns=[\"x\", \"fine_label\"])\r\n```\r\nbut to no avail. Any clue?",
"Thanks! I convert my dataset feature to Array3D and this speed became awesome!"
] | ### Describe the bug
I am attempting to iterate through an image dataset, but I am encountering a significant slowdown in the iteration speed. In order to investigate this issue, I conducted the following experiment:
```python
a=torch.randn(100,224)
a=torch.stack([a] * 10000)
a.shape
# %%
ds=Dataset.from_dict({"tensor":a})
for i in tqdm(ds.with_format("numpy")):
pass
for i in tqdm(ds.with_format("torch")):
pass
```
I noticed that the dataset in numpy format performs significantly faster than the one in torch format. My hypothesis is that the dataset undergoes a transformation process of torch->python->numpy(torch) in the background, which might be causing the slowdown. Is there any way to expedite the process by bypassing such transformations?
Furthermore, if I increase the size of a to an image shape, like:
```python
a=torch.randn(3,224,224)
```
the iteration speed becomes absurdly slow, around 100 iterations per second, whereas the speed with numpy format is approximately 250 iterations per second. This level of speed would be unacceptable for large image datasets, as it could take several hours just to iterate through a single epoch.
### Steps to reproduce the bug
```python
a=torch.randn(100,224)
a=torch.stack([a] * 10000)
a.shape
# %%
ds=Dataset.from_dict({"tensor":a})
for i in tqdm(ds.with_format("numpy")):
pass
for i in tqdm(ds.with_format("torch")):
pass
```
### Expected behavior
iteration faster
### Environment info
- `datasets` version: 2.11.0
- Platform: Linux-5.4.0-148-generic-x86_64-with-glibc2.10
- Python version: 3.8.16
- Huggingface_hub version: 0.13.4
- PyArrow version: 11.0.0
- Pandas version: 2.0.0 | 5,841 |
https://github.com/huggingface/datasets/issues/5840 | load model error. | [
"Please report this in the `transformers` repo, as it's not related to `datasets`"
] | ### Describe the bug
I had trained one model use deepspeed, when I load the final load I get the follow error:
OSError: Can't load tokenizer for '/XXX/DeepSpeedExamples/applications/DeepSpeed-Chat/output/step3-models/1.3b/actor'. If you were trying to load it from 'https://huggingface.co/models', make sure you don't have a local directory with the same name. Otherwise, make sure '/home/fm001/hzl/Project/DeepSpeedExamples/applications/DeepSpeed-Chat/output/step3-models/1.3b/actor' is the correct path to a directory containing all relevant files for a BloomTokenizerFast tokenizer.
my load code is : python chat.py --path /XXX/DeepSpeedExamples/applications/DeepSpeed-Chat/output/step3-models/1.3b/actor/
### Steps to reproduce the bug
。。。
### Expected behavior
。。。
### Environment info
。。。 | 5,840 |
https://github.com/huggingface/datasets/issues/5842 | Remove columns in interable dataset | [
"Transferring this issue as it's related to the 🤗 Datasets library ",
"Hi @surya-narayanan! Could you provide some code snippet?",
"This method has been recently added to the `IterableDataset`, so you need to update the `datasets`' installation (`pip install -U datasets`) to use it."
] | ### Feature request
Right now, remove_columns() produces a NotImplementedError for iterable style datasets
### Motivation
It would be great to have the same functionality irrespective of whether one is using an iterable or a map-style dataset
### Your contribution
hope and courage. | 5,842 |
https://github.com/huggingface/datasets/issues/5843 | Can't add iterable datasets to a Dataset Dict. | [
"Transferring as this is relating to the 🤗 Datasets library",
"You need to use `IterableDatasetDict` instead of `DatasetDict` for iterable datasets."
] | ### System Info
standard env
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [X] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
Get the following error:
TypeError: Values in `DatasetDict` should be of type `Dataset` but got type '<class 'datasets.iterable_dataset.IterableDataset'>'
### Expected behavior
should be able to add iterable datasets to a dataset dict | 5,843 |
https://github.com/huggingface/datasets/issues/5839 | Make models/functions optimized with `torch.compile` hashable | [] | As reported in https://github.com/huggingface/datasets/issues/5819, hashing functions/transforms that reference a model, or a function, optimized with `torch.compile` currently fails due to them not being picklable (the concrete error can be found in the linked issue).
The solutions to consider:
1. hashing/pickling the original, uncompiled version of a compiled model/function (attributes `_orig_mod`/`_torchdynamo_orig_callable`) (less precise than the 2nd option as it ignores the other params of `torch.compute`)
2. wait for https://github.com/pytorch/pytorch/issues/101107 to be resolved
| 5,839 |
https://github.com/huggingface/datasets/issues/5838 | Streaming support for `load_from_disk` | [
"As the name says, `load_from_disk` load the data from your disk. If the data is hosted on S3, it is first downloaded locally and then loaded from your disk.\r\n\r\nThere is a discussion on streaming data from S3 here though: #5281 ",
"@lhoestq \r\nThanks for your comment. I have checked out the discussion before and attempted at replicating the mentioned changes in the main branch (#5580). What I found was that if a dataset is saved using `save_to_disk`, it cannot be read by `load_dataset`. The error message asks me to to use `load_from_disk` instead. What would be the correct way of saving the data in this scenario?",
"Using `push_to_hub` you can save the dataset on the HF Hub as parquet files, and reload it / stream it using `load_dataset` :)\r\n\r\nIf you want to save your dataset somewhere else you can use `.to_parquet` to get a parquet file. If your dataset is big it's usually recommended to shard it into multi parquet files (around 1GB each).",
"@lhoestq \r\nThanks for the explanation. Appreciate it. I'll try this out.",
"@lhoestq\r\nI tried the method you mentioned. This the current scenario I'm facing:\r\n\r\n- The parquet file can be read from disk and streaming can be enabled.\r\n- The parquet file can be read from `s3` (local MinIO).\r\n- When `streaming=True` is enabled for `s3`, I get the error mentioned below:\r\n\r\n```\r\nFile ~/.../lib/python3.8/site-packages/s3fs/core.py:502, in S3FileSystem.set_session(self, refresh, kwargs)\r\n 500 conf = AioConfig(**config_kwargs)\r\n 501 if self.session is None:\r\n--> 502 self.session = aiobotocore.session.AioSession(**self.kwargs)\r\n 504 for parameters in (config_kwargs, self.kwargs, init_kwargs, client_kwargs):\r\n 505 for option in (\"region_name\", \"endpoint_url\"):\r\n\r\nTypeError: __init__() got an unexpected keyword argument 'headers'\r\n```\r\n\r\nDoes this mean there is a bug in the main branch?",
"Streaming from S3 is still experimental, there might be a few bugs unfortunately.\r\n\r\nCan you share the full stack trace ?",
"@lhoestq \r\nSure, here you go:\r\n\r\n```python\r\nTypeError Traceback (most recent call last)\r\nCell In[8], line 1\r\n----> 1 dataset = load_dataset(\"parquet\", data_files=[\"s3://<bucket name>/<data folder>/data-parquet\"], storage_options=fs.storage_options, streaming=True)\r\n\r\nFile ~/.../datasets/src/datasets/load.py:1790, in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, verification_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, num_proc, storage_options, **config_kwargs)\r\n 1788 # Return iterable dataset in case of streaming\r\n 1789 if streaming:\r\n-> 1790 return builder_instance.as_streaming_dataset(split=split)\r\n 1792 # Some datasets are already processed on the HF google storage\r\n 1793 # Don't try downloading from Google storage for the packaged datasets as text, json, csv or pandas\r\n 1794 try_from_hf_gcs = path not in _PACKAGED_DATASETS_MODULES\r\n\r\nFile ~/.../datasets/src/datasets/builder.py:1264, in DatasetBuilder.as_streaming_dataset(self, split, base_path)\r\n 1257 dl_manager = StreamingDownloadManager(\r\n 1258 base_path=base_path or self.base_path,\r\n 1259 download_config=DownloadConfig(use_auth_token=self.use_auth_token, storage_options=self.storage_options),\r\n 1260 dataset_name=self.name,\r\n 1261 data_dir=self.config.data_dir,\r\n 1262 )\r\n 1263 self._check_manual_download(dl_manager)\r\n-> 1264 splits_generators = {sg.name: sg for sg in self._split_generators(dl_manager)}\r\n 1265 # By default, return all splits\r\n 1266 if split is None:\r\n\r\nFile ~/.../datasets/src/datasets/packaged_modules/parquet/parquet.py:34, in Parquet._split_generators(self, dl_manager)\r\n 32 if not self.config.data_files:\r\n 33 raise ValueError(f\"At least one data file must be specified, but got data_files={self.config.data_files}\")\r\n---> 34 data_files = dl_manager.download_and_extract(self.config.data_files)\r\n 35 if isinstance(data_files, (str, list, tuple)):\r\n 36 files = data_files\r\n\r\nFile ~/.../datasets/src/datasets/download/streaming_download_manager.py:1087, in StreamingDownloadManager.download_and_extract(self, url_or_urls)\r\n 1069 def download_and_extract(self, url_or_urls):\r\n 1070 \"\"\"Prepare given `url_or_urls` for streaming (add extraction protocol).\r\n 1071 \r\n 1072 This is the lazy version of `DownloadManager.download_and_extract` for streaming.\r\n (...)\r\n 1085 url(s): (`str` or `list` or `dict`), URL(s) to stream data from matching the given input `url_or_urls`.\r\n 1086 \"\"\"\r\n-> 1087 return self.extract(self.download(url_or_urls))\r\n\r\nFile ~/.../datasets/src/datasets/download/streaming_download_manager.py:1039, in StreamingDownloadManager.extract(self, url_or_urls)\r\n 1020 def extract(self, url_or_urls):\r\n 1021 \"\"\"Add extraction protocol for given url(s) for streaming.\r\n 1022 \r\n 1023 This is the lazy version of `DownloadManager.extract` for streaming.\r\n (...)\r\n 1037 ```\r\n 1038 \"\"\"\r\n-> 1039 urlpaths = map_nested(self._extract, url_or_urls, map_tuple=True)\r\n 1040 return urlpaths\r\n\r\nFile ~/.../datasets/src/datasets/utils/py_utils.py:443, in map_nested(function, data_struct, dict_only, map_list, map_tuple, map_numpy, num_proc, parallel_min_length, types, disable_tqdm, desc)\r\n 441 num_proc = 1\r\n 442 if num_proc <= 1 or len(iterable) < parallel_min_length:\r\n--> 443 mapped = [\r\n 444 _single_map_nested((function, obj, types, None, True, None))\r\n 445 for obj in logging.tqdm(iterable, disable=disable_tqdm, desc=desc)\r\n 446 ]\r\n 447 else:\r\n 448 num_proc = num_proc if num_proc <= len(iterable) else len(iterable)\r\n\r\nFile ~/.../datasets/src/datasets/utils/py_utils.py:444, in <listcomp>(.0)\r\n 441 num_proc = 1\r\n 442 if num_proc <= 1 or len(iterable) < parallel_min_length:\r\n 443 mapped = [\r\n--> 444 _single_map_nested((function, obj, types, None, True, None))\r\n 445 for obj in logging.tqdm(iterable, disable=disable_tqdm, desc=desc)\r\n 446 ]\r\n 447 else:\r\n 448 num_proc = num_proc if num_proc <= len(iterable) else len(iterable)\r\n\r\nFile ~/.../datasets/src/datasets/utils/py_utils.py:363, in _single_map_nested(args)\r\n 361 return {k: _single_map_nested((function, v, types, None, True, None)) for k, v in pbar}\r\n 362 else:\r\n--> 363 mapped = [_single_map_nested((function, v, types, None, True, None)) for v in pbar]\r\n 364 if isinstance(data_struct, list):\r\n 365 return mapped\r\n\r\nFile ~/.../datasets/src/datasets/utils/py_utils.py:363, in <listcomp>(.0)\r\n 361 return {k: _single_map_nested((function, v, types, None, True, None)) for k, v in pbar}\r\n 362 else:\r\n--> 363 mapped = [_single_map_nested((function, v, types, None, True, None)) for v in pbar]\r\n 364 if isinstance(data_struct, list):\r\n 365 return mapped\r\n\r\nFile ~/.../datasets/src/datasets/utils/py_utils.py:346, in _single_map_nested(args)\r\n 344 # Singleton first to spare some computation\r\n 345 if not isinstance(data_struct, dict) and not isinstance(data_struct, types):\r\n--> 346 return function(data_struct)\r\n 348 # Reduce logging to keep things readable in multiprocessing with tqdm\r\n 349 if rank is not None and logging.get_verbosity() < logging.WARNING:\r\n\r\nFile ~/.../datasets/src/datasets/download/streaming_download_manager.py:1044, in StreamingDownloadManager._extract(self, urlpath)\r\n 1042 def _extract(self, urlpath: str) -> str:\r\n 1043 urlpath = str(urlpath)\r\n-> 1044 protocol = _get_extraction_protocol(urlpath, use_auth_token=self.download_config.use_auth_token)\r\n 1045 # get inner file: zip://train-00000.json.gz::https://foo.bar/data.zip -> zip://train-00000.json.gz\r\n 1046 path = urlpath.split(\"::\")[0]\r\n\r\nFile ~/.../datasets/src/datasets/download/streaming_download_manager.py:433, in _get_extraction_protocol(urlpath, use_auth_token)\r\n 431 else:\r\n 432 urlpath, kwargs = urlpath, {}\r\n--> 433 with fsspec.open(urlpath, **kwargs) as f:\r\n 434 return _get_extraction_protocol_with_magic_number(f)\r\n\r\nFile ~/.../lib/python3.8/site-packages/fsspec/core.py:102, in OpenFile.__enter__(self)\r\n 99 def __enter__(self):\r\n 100 mode = self.mode.replace(\"t\", \"\").replace(\"b\", \"\") + \"b\"\r\n--> 102 f = self.fs.open(self.path, mode=mode)\r\n 104 self.fobjects = [f]\r\n 106 if self.compression is not None:\r\n\r\nFile ~/.../lib/python3.8/site-packages/fsspec/spec.py:1199, in AbstractFileSystem.open(self, path, mode, block_size, cache_options, compression, **kwargs)\r\n 1197 else:\r\n 1198 ac = kwargs.pop(\"autocommit\", not self._intrans)\r\n-> 1199 f = self._open(\r\n 1200 path,\r\n 1201 mode=mode,\r\n 1202 block_size=block_size,\r\n 1203 autocommit=ac,\r\n 1204 cache_options=cache_options,\r\n 1205 **kwargs,\r\n 1206 )\r\n 1207 if compression is not None:\r\n 1208 from fsspec.compression import compr\r\n\r\nFile ~/.../lib/python3.8/site-packages/s3fs/core.py:659, in S3FileSystem._open(self, path, mode, block_size, acl, version_id, fill_cache, cache_type, autocommit, requester_pays, cache_options, **kwargs)\r\n 656 if cache_type is None:\r\n 657 cache_type = self.default_cache_type\r\n--> 659 return S3File(\r\n 660 self,\r\n 661 path,\r\n 662 mode,\r\n 663 block_size=block_size,\r\n 664 acl=acl,\r\n 665 version_id=version_id,\r\n 666 fill_cache=fill_cache,\r\n 667 s3_additional_kwargs=kw,\r\n 668 cache_type=cache_type,\r\n 669 autocommit=autocommit,\r\n 670 requester_pays=requester_pays,\r\n 671 cache_options=cache_options,\r\n 672 )\r\n\r\nFile ~/.../lib/python3.8/site-packages/s3fs/core.py:2043, in S3File.__init__(self, s3, path, mode, block_size, acl, version_id, fill_cache, s3_additional_kwargs, autocommit, cache_type, requester_pays, cache_options)\r\n 2041 self.details = s3.info(path)\r\n 2042 self.version_id = self.details.get(\"VersionId\")\r\n-> 2043 super().__init__(\r\n 2044 s3,\r\n 2045 path,\r\n 2046 mode,\r\n 2047 block_size,\r\n 2048 autocommit=autocommit,\r\n 2049 cache_type=cache_type,\r\n 2050 cache_options=cache_options,\r\n 2051 )\r\n 2052 self.s3 = self.fs # compatibility\r\n 2054 # when not using autocommit we want to have transactional state to manage\r\n\r\nFile ~/.../lib/python3.8/site-packages/fsspec/spec.py:1555, in AbstractBufferedFile.__init__(self, fs, path, mode, block_size, autocommit, cache_type, cache_options, size, **kwargs)\r\n 1553 self.size = size\r\n 1554 else:\r\n-> 1555 self.size = self.details[\"size\"]\r\n 1556 self.cache = caches[cache_type](\r\n 1557 self.blocksize, self._fetch_range, self.size, **cache_options\r\n 1558 )\r\n 1559 else:\r\n\r\nFile ~/.../lib/python3.8/site-packages/fsspec/spec.py:1568, in AbstractBufferedFile.details(self)\r\n 1565 @property\r\n 1566 def details(self):\r\n 1567 if self._details is None:\r\n-> 1568 self._details = self.fs.info(self.path)\r\n 1569 return self._details\r\n\r\nFile ~/.../lib/python3.8/site-packages/fsspec/asyn.py:115, in sync_wrapper.<locals>.wrapper(*args, **kwargs)\r\n 112 @functools.wraps(func)\r\n 113 def wrapper(*args, **kwargs):\r\n 114 self = obj or args[0]\r\n--> 115 return sync(self.loop, func, *args, **kwargs)\r\n\r\nFile ~/.../lib/python3.8/site-packages/fsspec/asyn.py:100, in sync(loop, func, timeout, *args, **kwargs)\r\n 98 raise FSTimeoutError from return_result\r\n 99 elif isinstance(return_result, BaseException):\r\n--> 100 raise return_result\r\n 101 else:\r\n 102 return return_result\r\n\r\nFile ~/.../lib/python3.8/site-packages/fsspec/asyn.py:55, in _runner(event, coro, result, timeout)\r\n 53 coro = asyncio.wait_for(coro, timeout=timeout)\r\n 54 try:\r\n---> 55 result[0] = await coro\r\n 56 except Exception as ex:\r\n 57 result[0] = ex\r\n\r\nFile ~/.../lib/python3.8/site-packages/s3fs/core.py:1248, in S3FileSystem._info(self, path, bucket, key, refresh, version_id)\r\n 1246 if key:\r\n 1247 try:\r\n-> 1248 out = await self._call_s3(\r\n 1249 \"head_object\",\r\n 1250 self.kwargs,\r\n 1251 Bucket=bucket,\r\n 1252 Key=key,\r\n 1253 **version_id_kw(version_id),\r\n 1254 **self.req_kw,\r\n 1255 )\r\n 1256 return {\r\n 1257 \"ETag\": out.get(\"ETag\", \"\"),\r\n 1258 \"LastModified\": out[\"LastModified\"],\r\n (...)\r\n 1264 \"ContentType\": out.get(\"ContentType\"),\r\n 1265 }\r\n 1266 except FileNotFoundError:\r\n\r\nFile ~/.../lib/python3.8/site-packages/s3fs/core.py:341, in S3FileSystem._call_s3(self, method, *akwarglist, **kwargs)\r\n 340 async def _call_s3(self, method, *akwarglist, **kwargs):\r\n--> 341 await self.set_session()\r\n 342 s3 = await self.get_s3(kwargs.get(\"Bucket\"))\r\n 343 method = getattr(s3, method)\r\n\r\nFile ~/.../lib/python3.8/site-packages/s3fs/core.py:502, in S3FileSystem.set_session(self, refresh, kwargs)\r\n 500 conf = AioConfig(**config_kwargs)\r\n 501 if self.session is None:\r\n--> 502 self.session = aiobotocore.session.AioSession(**self.kwargs)\r\n 504 for parameters in (config_kwargs, self.kwargs, init_kwargs, client_kwargs):\r\n 505 for option in (\"region_name\", \"endpoint_url\"):\r\n\r\nTypeError: __init__() got an unexpected keyword argument 'headers'\r\n```",
"Is `\"data-parquet\"` a file ? In `data_files` you should pass the paths to the parquet files (not to a directory). Glob patterns are not supported yet for S3 URLs.\r\n\r\nThe bug seems to happen because your provided data file has no extension. Because of that it tries to infer it from the file content, but fails because `_get_extraction_protocol` doesn't support S3 URLs yet.\r\n\r\n",
"@lhoestq \r\nThank you for your answer. Saving the file with `.parquet` extension solved the issue! This is really great! Really appreciate all the help! \r\n\r\nLet me know if I should close the issue or feel free to close it if you want.",
"Cool ! I'm glad it worked out :)\r\n\r\nSure feel free to close the issue, since the original question about streaming with load_from_disk has been answered anyway"
] | ### Feature request
Support for streaming datasets stored in object stores in `load_from_disk`.
### Motivation
The `load_from_disk` function supports fetching datasets stored in object stores such as `s3`. In many cases, the datasets that are stored in object stores are very large and being able to stream the data from the buckets becomes essential.
### Your contribution
I'd be happy to contribute this feature if I could get the guidance on how to do so. | 5,838 |
https://github.com/huggingface/datasets/issues/5837 | Use DeepSpeed load myself " .csv " dataset. | [
"Hi ! Doing `load_dataset(\"path/to/data.csv\")` is not supported yet, but you can do\r\n\r\n```python\r\nds = load_dataset(\"csv\", data_files=[\"path/to/data.csv\"])\r\n```",
"@lhoestq thank you.",
"The other question: \r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 1767, in load_dataset\r\n builder_instance = load_dataset_builder(\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 1498, in load_dataset_builder\r\n dataset_module = dataset_module_factory(\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 1127, in dataset_module_factory\r\n return PackagedDatasetModuleFactory(\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 708, in get_module\r\n data_files = DataFilesDict.from_local_or_remote(\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/data_files.py\", line 796, in from_local_or_remote\r\n DataFilesList.from_local_or_remote(\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/data_files.py\", line 764, in from_local_or_remote\r\n data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/data_files.py\", line 362, in resolve_patterns_locally_or_by_urls\r\n for path in _resolve_single_pattern_locally(base_path, pattern, allowed_extensions):\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/data_files.py\", line 306, in _resolve_single_pattern_locally\r\n raise FileNotFoundError(error_msg)\r\nFileNotFoundError: Unable to find '/home/fm001/hzl/Data/qa/' at /\r\n>>> mydata = load_dataset(\"/home/fm001/hzl/Data/qa/\")\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 1767, in load_dataset\r\n builder_instance = load_dataset_builder(\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 1508, in load_dataset_builder\r\n builder_cls = import_main_class(dataset_module.module_path)\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 115, in import_main_class\r\n module = importlib.import_module(module_path)\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/importlib/__init__.py\", line 127, in import_module\r\n return _bootstrap._gcd_import(name[level:], package, level)\r\n File \"<frozen importlib._bootstrap>\", line 1014, in _gcd_import\r\n File \"<frozen importlib._bootstrap>\", line 991, in _find_and_load\r\n File \"<frozen importlib._bootstrap>\", line 975, in _find_and_load_unlocked\r\n File \"<frozen importlib._bootstrap>\", line 671, in _load_unlocked\r\n File \"<frozen importlib._bootstrap_external>\", line 783, in exec_module\r\n File \"<frozen importlib._bootstrap>\", line 219, in _call_with_frames_removed\r\n File \"/home/fm001/.cache/huggingface/modules/datasets_modules/datasets/qa/b8b9f481eff9d17b769b4b50f30a51da32b47c94d1af4d2bdffb9fc2c589513a/qa.py\", line 2, in <module>\r\n mydata = load_dataset(\"/home/fm001/hzl/Data/qa/\")\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 1767, in load_dataset\r\n builder_instance = load_dataset_builder(\r\n File \"/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py\", line 1524, in load_dataset_builder\r\n builder_instance: DatasetBuilder = builder_cls(\r\nTypeError: 'NoneType' object is not callable\r\n\r\nAnd I follow the setting with https://huggingface.co/docs/datasets/dataset_script"
] | ### Describe the bug
When I use DeepSpeed train a model with my own " XXX.csv" dataset I got the follow question:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py", line 1767, in load_dataset
builder_instance = load_dataset_builder(
File "/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py", line 1498, in load_dataset_builder
dataset_module = dataset_module_factory(
File "/home/fm001/.conda/envs/hzl/lib/python3.8/site-packages/datasets/load.py", line 1217, in dataset_module_factory
raise FileNotFoundError(
FileNotFoundError: Couldn't find a dataset script at /home/fm001/hzl/Data/qa.csv/qa.csv.py or any data file in the same directory.
### Steps to reproduce the bug
my code is :
from datasets import load_dataset
mydata = load_dataset("/home/fm001/hzl/Data/qa.csv")
### Expected behavior
。。。
### Environment info
。。。 | 5,837 |