Datasets:
Tasks:
Text Classification
Modalities:
Text
Sub-tasks:
sentiment-classification
Size:
10K - 100K
License:
Add files
Browse files- src/translate_imdb_flax.py +309 -0
- test.jsonl.gz +3 -0
- train.jsonl.gz +3 -0
- unsupervised.jsonl.gz +3 -0
src/translate_imdb_flax.py
ADDED
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import functools
|
2 |
+
import json
|
3 |
+
import logging
|
4 |
+
import os
|
5 |
+
import pprint
|
6 |
+
from typing import Tuple
|
7 |
+
|
8 |
+
from datasets import get_dataset_config_names, load_dataset, get_dataset_split_names
|
9 |
+
import jax
|
10 |
+
import numpy as np
|
11 |
+
from flax import jax_utils
|
12 |
+
from flax.jax_utils import pad_shard_unpad
|
13 |
+
from transformers import AutoTokenizer, FlaxAutoModelForSeq2SeqLM
|
14 |
+
import pandas as pd
|
15 |
+
|
16 |
+
logger = logging.getLogger(__name__)
|
17 |
+
|
18 |
+
DATASET_NAME = "imdb"
|
19 |
+
OUTPUT_DIR = "./imdb_dutch"
|
20 |
+
MODEL_370 = "yhavinga/ul2-large-en-nl"
|
21 |
+
# BATCH_SIZE = 64
|
22 |
+
BATCH_SIZE = 32
|
23 |
+
# BATCH_SIZE = 2
|
24 |
+
MODEL_MAX_LENGTH = 370
|
25 |
+
MAX_WORDS = int(MODEL_MAX_LENGTH / 3)
|
26 |
+
END_MARKS = (".", "?", "!", '"', "'", "\n")
|
27 |
+
|
28 |
+
|
29 |
+
class FlaxModel:
|
30 |
+
def __init__(self, model_name: str, tokenizer_name: str, tokenizer_args={}):
|
31 |
+
"""
|
32 |
+
Initializes the FlaxModel with the specified model and tokenizer names, as well as tokenizer arguments.
|
33 |
+
"""
|
34 |
+
self.model = FlaxAutoModelForSeq2SeqLM.from_pretrained(
|
35 |
+
model_name, use_auth_token=True
|
36 |
+
)
|
37 |
+
self.model.params = self.model.to_fp32(self.model.params, mask=None)
|
38 |
+
self.tokenizer_args = {
|
39 |
+
# "model_max_length": self.model.config.max_length,
|
40 |
+
**tokenizer_args,
|
41 |
+
}
|
42 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
43 |
+
tokenizer_name, use_auth_token=True, **self.tokenizer_args
|
44 |
+
)
|
45 |
+
# if not (
|
46 |
+
# self.model.config.max_length
|
47 |
+
# == self.tokenizer.model_max_length
|
48 |
+
# == self.tokenizer_args.get("model_max_length")
|
49 |
+
# ):
|
50 |
+
# print(
|
51 |
+
# f"Warning: model max length {self.model.config.max_length} != tokenizer max length {self.tokenizer.model_max_length} != tokenizer_args max length {tokenizer_args.get('model_max_length')}"
|
52 |
+
# )
|
53 |
+
# raise ValueError("Model and tokenizer max_length should be equal")
|
54 |
+
|
55 |
+
self.params = jax_utils.replicate(self.model.params)
|
56 |
+
|
57 |
+
kwargs = {
|
58 |
+
"max_length": self.tokenizer.model_max_length,
|
59 |
+
"length_penalty": 1.0,
|
60 |
+
"num_beams": 4,
|
61 |
+
"early_stopping": True,
|
62 |
+
}
|
63 |
+
|
64 |
+
def shard(xs):
|
65 |
+
local_device_count = jax.local_device_count()
|
66 |
+
return jax.tree_map(
|
67 |
+
lambda x: x.reshape((local_device_count, -1) + x.shape[1:]), xs
|
68 |
+
)
|
69 |
+
|
70 |
+
def generate_step(params, batch):
|
71 |
+
self.model.params = params
|
72 |
+
output_ids = self.model.generate(
|
73 |
+
batch["input_ids"], attention_mask=batch["attention_mask"], **kwargs
|
74 |
+
)
|
75 |
+
return output_ids.sequences
|
76 |
+
|
77 |
+
self.p_generate_step = jax.pmap(generate_step, "batch")
|
78 |
+
|
79 |
+
@functools.lru_cache()
|
80 |
+
def translate_batch(self, texts: Tuple[str]):
|
81 |
+
overflowed = False
|
82 |
+
texts = list(texts)
|
83 |
+
if self.model.config.prefix:
|
84 |
+
texts = [self.model.config.prefix + x for x in texts]
|
85 |
+
texts = [x.replace("\n", "<n>").replace("<br />", "<n>") for x in texts]
|
86 |
+
inputs = self.tokenizer(
|
87 |
+
texts,
|
88 |
+
max_length=self.tokenizer_args.get("model_max_length"),
|
89 |
+
truncation=True,
|
90 |
+
padding="max_length",
|
91 |
+
return_tensors="np",
|
92 |
+
)
|
93 |
+
if not np.array_equal(
|
94 |
+
inputs.data["input_ids"][:, self.tokenizer.model_max_length - 1],
|
95 |
+
np.zeros(BATCH_SIZE),
|
96 |
+
):
|
97 |
+
overflowed = True
|
98 |
+
return BATCH_SIZE * [""], overflowed
|
99 |
+
|
100 |
+
batch = inputs.data
|
101 |
+
print(f"Batch inputs shape is {batch['input_ids'].shape}")
|
102 |
+
translated = pad_shard_unpad(self.p_generate_step)(self.params, batch)
|
103 |
+
predictions = jax.device_get(
|
104 |
+
translated.reshape(-1, self.tokenizer.model_max_length)
|
105 |
+
)
|
106 |
+
if not np.array_equal(
|
107 |
+
predictions[:, self.tokenizer.model_max_length - 1],
|
108 |
+
np.zeros(BATCH_SIZE),
|
109 |
+
):
|
110 |
+
overflowed = True
|
111 |
+
|
112 |
+
output = [
|
113 |
+
self.tokenizer.decode(t, skip_special_tokens=False) for t in predictions
|
114 |
+
]
|
115 |
+
# If there is <extra_id in the output, remove it and everything after it
|
116 |
+
output = [
|
117 |
+
x.replace("<pad>", "").replace("</s>", "").split("<extra_id")[0]
|
118 |
+
for x in output
|
119 |
+
]
|
120 |
+
output = [x.replace("<n>", "<br />").strip() for x in output]
|
121 |
+
return output, overflowed
|
122 |
+
|
123 |
+
|
124 |
+
def split_text(text):
|
125 |
+
text_parts = []
|
126 |
+
current_part = ""
|
127 |
+
|
128 |
+
def split_on_end_marks(text):
|
129 |
+
sentences = []
|
130 |
+
current_sentence = ""
|
131 |
+
for char in text:
|
132 |
+
if char in END_MARKS:
|
133 |
+
sentences.append(current_sentence + char)
|
134 |
+
current_sentence = ""
|
135 |
+
else:
|
136 |
+
current_sentence += char
|
137 |
+
|
138 |
+
# Add the final sentence if it wasn't ended by an end of line mark
|
139 |
+
if current_sentence:
|
140 |
+
sentences.append(current_sentence)
|
141 |
+
return sentences
|
142 |
+
|
143 |
+
text_lines = split_on_end_marks(text)
|
144 |
+
|
145 |
+
for line in text_lines:
|
146 |
+
# If adding the line to the current part would not exceed MAX_WORDS words, add it to the current part
|
147 |
+
if len((current_part + line).split()) <= MAX_WORDS:
|
148 |
+
current_part += line
|
149 |
+
# If adding the line to the current part would exceed 200 characters, add the current part to the list and reset the current part
|
150 |
+
else:
|
151 |
+
if len(current_part) > 0:
|
152 |
+
text_parts.append(current_part)
|
153 |
+
while len(line.split()) > MAX_WORDS:
|
154 |
+
# print(f"Line {line} is longer than MAX_WORDS words")
|
155 |
+
current_part = " ".join(line.split()[:MAX_WORDS])
|
156 |
+
text_parts.append(current_part + " ")
|
157 |
+
line = " ".join(line.split()[MAX_WORDS:])
|
158 |
+
current_part = line
|
159 |
+
# Add the final part to the list
|
160 |
+
text_parts.append(current_part)
|
161 |
+
text_parts[-1] = text_parts[-1].rstrip()
|
162 |
+
return text_parts
|
163 |
+
|
164 |
+
|
165 |
+
def test_split_text():
|
166 |
+
# Test with single line that is less than MAX_WORDS words
|
167 |
+
text = " ".join([f"n{i}" for i in range(MAX_WORDS - 20)])
|
168 |
+
a = list(text)
|
169 |
+
a[150] = END_MARKS[0]
|
170 |
+
text = "".join(a)
|
171 |
+
text_parts = split_text(text)
|
172 |
+
assert text_parts == [text]
|
173 |
+
|
174 |
+
# Test with single line that is exactly MAX_WORDS words
|
175 |
+
text = " ".join([f"n{i}" for i in range(MAX_WORDS)])
|
176 |
+
a = list(text)
|
177 |
+
a[10] = END_MARKS[0]
|
178 |
+
text = "".join(a)
|
179 |
+
text_parts = split_text(text)
|
180 |
+
assert text_parts == [text]
|
181 |
+
|
182 |
+
# Test with single line that is more than MAX_WORDS words
|
183 |
+
text = " ".join([f"n{i}" for i in range(MAX_WORDS + 1)])
|
184 |
+
a = list(text)
|
185 |
+
a[150] = END_MARKS[0]
|
186 |
+
text = "".join(a)
|
187 |
+
text_parts = split_text(text)
|
188 |
+
assert text_parts == [text[:151], text[151:]]
|
189 |
+
|
190 |
+
# Test with multiple lines, none of which are more than 200 characters
|
191 |
+
text = "\n".join([f"n{i}" for i in range(10)])
|
192 |
+
text_parts = split_text(text)
|
193 |
+
assert text_parts == [text]
|
194 |
+
|
195 |
+
# Test with 500 words
|
196 |
+
text = " ".join([f"n{i}" for i in range(500)])
|
197 |
+
a = list(text)
|
198 |
+
a[150] = END_MARKS[0]
|
199 |
+
a[300] = END_MARKS[0]
|
200 |
+
a[550] = END_MARKS[0]
|
201 |
+
a[600] = END_MARKS[0]
|
202 |
+
a[750] = END_MARKS[0]
|
203 |
+
a[900] = END_MARKS[0]
|
204 |
+
a[950] = END_MARKS[0]
|
205 |
+
a[1000] = END_MARKS[0]
|
206 |
+
text = "".join(a)
|
207 |
+
text_parts = split_text(text)
|
208 |
+
assert all(
|
209 |
+
[len(x.split()) <= MAX_WORDS for x in text_parts]
|
210 |
+
), "Not all text parts are less than MAX_WORDS words"
|
211 |
+
assert "".join(text_parts) == text, "Text parts concatenated != original text"
|
212 |
+
|
213 |
+
|
214 |
+
test_split_text()
|
215 |
+
|
216 |
+
|
217 |
+
def get_file_lines(filename):
|
218 |
+
"""
|
219 |
+
Get the number of lines in a file, 0 if the file does not exist.
|
220 |
+
"""
|
221 |
+
lines = 0
|
222 |
+
if os.path.exists(filename):
|
223 |
+
with open(filename) as f:
|
224 |
+
with open(filename, "r") as f:
|
225 |
+
lines = len(f.readlines())
|
226 |
+
print(f"{filename} already has {lines} lines")
|
227 |
+
return lines
|
228 |
+
|
229 |
+
|
230 |
+
SEP = "\n"
|
231 |
+
# SEP="<unk>"
|
232 |
+
|
233 |
+
|
234 |
+
def main():
|
235 |
+
os.makedirs(OUTPUT_DIR, exist_ok=True)
|
236 |
+
|
237 |
+
model_370 = FlaxModel(
|
238 |
+
MODEL_370, MODEL_370, tokenizer_args={"model_max_length": MODEL_MAX_LENGTH}
|
239 |
+
)
|
240 |
+
|
241 |
+
for config in get_dataset_config_names(DATASET_NAME):
|
242 |
+
print(f"Processing config {config}")
|
243 |
+
ds = load_dataset(DATASET_NAME, config)
|
244 |
+
# for split in ["validation"]:
|
245 |
+
for split in get_dataset_split_names(DATASET_NAME, config):
|
246 |
+
output_file = f"{OUTPUT_DIR}/{DATASET_NAME}_dutch_{config}-{split}.json"
|
247 |
+
num_examples = len(ds[split])
|
248 |
+
# fn = partial(encode_in_single_text, validation=(split == "validation"))
|
249 |
+
# single_text_ds = ds[split].map(fn, num_proc=6).sort("length", reverse=True)
|
250 |
+
# # fn = partial(batch_single_text_decode, validation=(split == "validation"))
|
251 |
+
# # decoded_ds = single_text_ds.map(fn, num_proc=6)
|
252 |
+
#
|
253 |
+
lines = get_file_lines(output_file)
|
254 |
+
start_batch_index = lines // BATCH_SIZE
|
255 |
+
with open(output_file, mode="ab" if lines else "wb") as writer:
|
256 |
+
for batch_index in range(start_batch_index, num_examples // BATCH_SIZE):
|
257 |
+
ds_split = ds[split]
|
258 |
+
batch = ds_split[
|
259 |
+
batch_index * BATCH_SIZE : (batch_index + 1) * BATCH_SIZE
|
260 |
+
]
|
261 |
+
print(
|
262 |
+
f"Translating batch {batch_index} of {num_examples // BATCH_SIZE}"
|
263 |
+
)
|
264 |
+
|
265 |
+
translated, overflow = model_370.translate_batch(
|
266 |
+
tuple(batch["text"])
|
267 |
+
)
|
268 |
+
translated_batch = [{"text": x} for x in translated]
|
269 |
+
if overflow:
|
270 |
+
batch_text_splitted = [
|
271 |
+
split_text(text) for text in batch["text"]
|
272 |
+
]
|
273 |
+
max_parts = max(
|
274 |
+
[len(text) for text in batch_text_splitted]
|
275 |
+
)
|
276 |
+
text_translated = [""] * BATCH_SIZE
|
277 |
+
for part_index in range(max_parts):
|
278 |
+
text_parts_i = [
|
279 |
+
text[part_index] if part_index < len(text) else ""
|
280 |
+
for text in batch_text_splitted
|
281 |
+
]
|
282 |
+
(
|
283 |
+
text_part_translated,
|
284 |
+
overflow,
|
285 |
+
) = model_370.translate_batch(tuple(text_parts_i))
|
286 |
+
if overflow:
|
287 |
+
print(
|
288 |
+
f"This shouldn't happen, overflow on a splitted text: {text_parts_i}"
|
289 |
+
)
|
290 |
+
for bi in range(BATCH_SIZE):
|
291 |
+
text_translated[bi] += " " + text_part_translated[bi] if text_parts_i[bi] != "" else ""
|
292 |
+
for bi in range(BATCH_SIZE):
|
293 |
+
translated_batch[bi]["text"] = text_translated[bi].strip()
|
294 |
+
|
295 |
+
# write each object in the batch as a separate line
|
296 |
+
for bi in range(BATCH_SIZE):
|
297 |
+
example = {
|
298 |
+
"text": translated_batch[bi]["text"],
|
299 |
+
"text_en": batch["text"][bi],
|
300 |
+
"label": batch["label"][bi],
|
301 |
+
}
|
302 |
+
|
303 |
+
pprint.pprint(example)
|
304 |
+
writer.write(json.dumps(example).encode("utf-8"))
|
305 |
+
writer.write("\n".encode("utf-8"))
|
306 |
+
|
307 |
+
|
308 |
+
if __name__ == "__main__":
|
309 |
+
main()
|
test.jsonl.gz
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b1f7f0224277f89fe229c95c6437a5ee6e95db630734cbde6e549ece93d32205
|
3 |
+
size 26536475
|
train.jsonl.gz
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:432b57ce1ce49554be6dcd79da08900c0330fb094bd43182314d8baefe189852
|
3 |
+
size 27182787
|
unsupervised.jsonl.gz
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b14d7977619b67d07f4e4b7fdd7f2d3ea79717d865bec60983e9ff678c2565cf
|
3 |
+
size 54451678
|