darshanmakwana
commited on
Update README.md
Browse files
README.md
CHANGED
@@ -26,3 +26,56 @@ configs:
|
|
26 |
- split: test
|
27 |
path: data/test-*
|
28 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
- split: test
|
27 |
path: data/test-*
|
28 |
---
|
29 |
+
|
30 |
+
This dataset contains tokenized audio from [lewtun/music_genres](https://huggingface.co/datasets/lewtun/music_genres) using [SemantiCodec](https://arxiv.org/abs/2405.00233) for performing experiments on AR music generation
|
31 |
+
|
32 |
+
The following script is used for tokenization
|
33 |
+
```python
|
34 |
+
from datasets import load_dataset, Dataset, DatasetDict
|
35 |
+
|
36 |
+
model_id = ""
|
37 |
+
repo_name = ""
|
38 |
+
user_name = ""
|
39 |
+
token = ""
|
40 |
+
cache_dir = "cache"
|
41 |
+
vocab_size = 4096
|
42 |
+
|
43 |
+
dataset = load_dataset(model_id, cache_dir=cache_dir, trust_remote_code=True)
|
44 |
+
|
45 |
+
from semanticodec import SemantiCodec
|
46 |
+
|
47 |
+
semanticodec = SemantiCodec(token_rate=100, semantic_vocab_size=vocab_size)
|
48 |
+
|
49 |
+
import soundfile as sf
|
50 |
+
from tqdm import tqdm
|
51 |
+
import math
|
52 |
+
|
53 |
+
dd = {
|
54 |
+
"train": 0,
|
55 |
+
"test": 0
|
56 |
+
}
|
57 |
+
|
58 |
+
for split in ["train", "test"]:
|
59 |
+
tkns = []
|
60 |
+
for idx in tqdm(range(len(dataset[split]))):
|
61 |
+
sample = dataset[split][idx]["audio"]
|
62 |
+
array = sample["array"]
|
63 |
+
sr = sample["sampling_rate"]
|
64 |
+
|
65 |
+
sf.write("output.wav", array, sr)
|
66 |
+
|
67 |
+
tokens = semanticodec.encode("output.wav").detach().cpu().numpy().flatten()
|
68 |
+
tkns.append(tokens)
|
69 |
+
|
70 |
+
ds = Dataset.from_dict({
|
71 |
+
"audio_tokens": tkns,
|
72 |
+
"genre_id": list(dataset[split]["genre_id"]),
|
73 |
+
"genre": list(dataset[split]["genre"]),
|
74 |
+
"song_id": list(dataset[split]["song_id"])
|
75 |
+
})
|
76 |
+
dd[split] = ds
|
77 |
+
|
78 |
+
dd = DatasetDict(dd)
|
79 |
+
dd.save_to_disk(repo_name)
|
80 |
+
dd.push_to_hub(f"{user_name}/{repo_name}", token=token)
|
81 |
+
```
|