fgaim commited on
Commit
7b282b4
1 Parent(s): 46f5a1f

Add scripts

Browse files
Files changed (5) hide show
  1. create_config.py +4 -0
  2. flax_to_torch.py +4 -0
  3. run.sh +22 -0
  4. run_mlm_flax.py +691 -0
  5. train_tokenizer.py +41 -0
create_config.py ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ from transformers import BertConfig
2
+
3
+ config = BertConfig.from_pretrained("bert-base-uncased")
4
+ config.save_pretrained("./")
flax_to_torch.py ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ from transformers import BertForMaskedLM
2
+
3
+ model = BertForMaskedLM.from_pretrained("./", from_flax=True)
4
+ model.save_pretrained("./")
run.sh ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ./run_mlm_flax.py \
2
+ --output_dir="./" \
3
+ --model_type="bert" \
4
+ --config_name="./" \
5
+ --tokenizer_name="./" \
6
+ --dataset_name="flax-community/swahili-safi" \
7
+ --validation_split_percentage="2" \
8
+ --max_seq_length="512" \
9
+ --weight_decay="0.025" \
10
+ --per_device_train_batch_size="64" \
11
+ --per_device_eval_batch_size="64" \
12
+ --learning_rate="3e-4" \
13
+ --warmup_steps="500" \
14
+ --overwrite_output_dir \
15
+ --num_train_epochs="15" \
16
+ --adam_beta1="0.9" \
17
+ --adam_beta2="0.98" \
18
+ --logging_steps="250" \
19
+ --save_steps="1000" \
20
+ --eval_steps="1000" \
21
+ --preprocessing_num_workers="96" \
22
+ --push_to_hub
run_mlm_flax.py ADDED
@@ -0,0 +1,691 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Team All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+ """
17
+ Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
18
+ text file or a dataset.
19
+
20
+ Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
21
+ https://huggingface.co/models?filter=masked-lm
22
+ """
23
+ import logging
24
+ import os
25
+ import sys
26
+ import time
27
+ from dataclasses import dataclass, field
28
+
29
+ # You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
30
+ from pathlib import Path
31
+ from typing import Dict, List, Optional, Tuple
32
+
33
+ import numpy as np
34
+ from datasets import load_dataset
35
+ from tqdm import tqdm
36
+
37
+ import flax
38
+ import jax
39
+ import jax.numpy as jnp
40
+ import optax
41
+ from flax import jax_utils, traverse_util
42
+ from flax.training import train_state
43
+ from flax.training.common_utils import get_metrics, onehot, shard
44
+ from transformers import (
45
+ CONFIG_MAPPING,
46
+ FLAX_MODEL_FOR_MASKED_LM_MAPPING,
47
+ AutoConfig,
48
+ AutoTokenizer,
49
+ FlaxAutoModelForMaskedLM,
50
+ HfArgumentParser,
51
+ PreTrainedTokenizerBase,
52
+ TensorType,
53
+ TrainingArguments,
54
+ is_tensorboard_available,
55
+ set_seed,
56
+ )
57
+
58
+
59
+ MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
60
+ MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
61
+
62
+
63
+ @dataclass
64
+ class ModelArguments:
65
+ """
66
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
67
+ """
68
+
69
+ model_name_or_path: Optional[str] = field(
70
+ default=None,
71
+ metadata={
72
+ "help": "The model checkpoint for weights initialization."
73
+ "Don't set if you want to train a model from scratch."
74
+ },
75
+ )
76
+ model_type: Optional[str] = field(
77
+ default=None,
78
+ metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
79
+ )
80
+ config_name: Optional[str] = field(
81
+ default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
82
+ )
83
+ tokenizer_name: Optional[str] = field(
84
+ default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
85
+ )
86
+ cache_dir: Optional[str] = field(
87
+ default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
88
+ )
89
+ use_fast_tokenizer: bool = field(
90
+ default=True,
91
+ metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
92
+ )
93
+ dtype: Optional[str] = field(
94
+ default="float32",
95
+ metadata={
96
+ "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
97
+ },
98
+ )
99
+
100
+
101
+ @dataclass
102
+ class DataTrainingArguments:
103
+ """
104
+ Arguments pertaining to what data we are going to input our model for training and eval.
105
+ """
106
+
107
+ dataset_name: Optional[str] = field(
108
+ default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
109
+ )
110
+ dataset_config_name: Optional[str] = field(
111
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
112
+ )
113
+ train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
114
+ validation_file: Optional[str] = field(
115
+ default=None,
116
+ metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
117
+ )
118
+ train_ref_file: Optional[str] = field(
119
+ default=None,
120
+ metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
121
+ )
122
+ validation_ref_file: Optional[str] = field(
123
+ default=None,
124
+ metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
125
+ )
126
+ overwrite_cache: bool = field(
127
+ default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
128
+ )
129
+ validation_split_percentage: Optional[int] = field(
130
+ default=5,
131
+ metadata={
132
+ "help": "The percentage of the train set used as validation set in case there's no validation split"
133
+ },
134
+ )
135
+ max_seq_length: Optional[int] = field(
136
+ default=None,
137
+ metadata={
138
+ "help": "The maximum total input sequence length after tokenization. Sequences longer "
139
+ "than this will be truncated. Default to the max input length of the model."
140
+ },
141
+ )
142
+ preprocessing_num_workers: Optional[int] = field(
143
+ default=None,
144
+ metadata={"help": "The number of processes to use for the preprocessing."},
145
+ )
146
+ mlm_probability: float = field(
147
+ default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
148
+ )
149
+ pad_to_max_length: bool = field(
150
+ default=False,
151
+ metadata={
152
+ "help": "Whether to pad all samples to `max_seq_length`. "
153
+ "If False, will pad the samples dynamically when batching to the maximum length in the batch."
154
+ },
155
+ )
156
+ line_by_line: bool = field(
157
+ default=False,
158
+ metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
159
+ )
160
+
161
+ def __post_init__(self):
162
+ if self.dataset_name is None and self.train_file is None and self.validation_file is None:
163
+ raise ValueError("Need either a dataset name or a training/validation file.")
164
+ else:
165
+ if self.train_file is not None:
166
+ extension = self.train_file.split(".")[-1]
167
+ assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
168
+ if self.validation_file is not None:
169
+ extension = self.validation_file.split(".")[-1]
170
+ assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
171
+
172
+
173
+ @flax.struct.dataclass
174
+ class FlaxDataCollatorForLanguageModeling:
175
+ """
176
+ Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
177
+ are not all of the same length.
178
+
179
+ Args:
180
+ tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
181
+ The tokenizer used for encoding the data.
182
+ mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
183
+ The probability with which to (randomly) mask tokens in the input.
184
+
185
+ .. note::
186
+
187
+ For best performance, this data collator should be used with a dataset having items that are dictionaries or
188
+ BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
189
+ :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
190
+ argument :obj:`return_special_tokens_mask=True`.
191
+ """
192
+
193
+ tokenizer: PreTrainedTokenizerBase
194
+ mlm_probability: float = 0.15
195
+
196
+ def __post_init__(self):
197
+ if self.tokenizer.mask_token is None:
198
+ raise ValueError(
199
+ "This tokenizer does not have a mask token which is necessary for masked language modeling. "
200
+ "You should pass `mlm=False` to train on causal language modeling instead."
201
+ )
202
+
203
+ def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
204
+ # Handle dict or lists with proper padding and conversion to tensor.
205
+ batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
206
+
207
+ # If special token mask has been preprocessed, pop it from the dict.
208
+ special_tokens_mask = batch.pop("special_tokens_mask", None)
209
+
210
+ batch["input_ids"], batch["labels"] = self.mask_tokens(
211
+ batch["input_ids"], special_tokens_mask=special_tokens_mask
212
+ )
213
+ return batch
214
+
215
+ def mask_tokens(
216
+ self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
217
+ ) -> Tuple[jnp.ndarray, jnp.ndarray]:
218
+ """
219
+ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
220
+ """
221
+ labels = inputs.copy()
222
+ # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
223
+ probability_matrix = np.full(labels.shape, self.mlm_probability)
224
+ special_tokens_mask = special_tokens_mask.astype("bool")
225
+
226
+ probability_matrix[special_tokens_mask] = 0.0
227
+ masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
228
+ labels[~masked_indices] = -100 # We only compute loss on masked tokens
229
+
230
+ # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
231
+ indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
232
+ inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
233
+
234
+ # 10% of the time, we replace masked input tokens with random word
235
+ indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
236
+ indices_random &= masked_indices & ~indices_replaced
237
+
238
+ random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
239
+ inputs[indices_random] = random_words[indices_random]
240
+
241
+ # The rest of the time (10% of the time) we keep the masked input tokens unchanged
242
+ return inputs, labels
243
+
244
+
245
+ def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
246
+ num_samples = len(samples_idx)
247
+ samples_to_remove = num_samples % batch_size
248
+
249
+ if samples_to_remove != 0:
250
+ samples_idx = samples_idx[:-samples_to_remove]
251
+ sections_split = num_samples // batch_size
252
+ batch_idx = np.split(samples_idx, sections_split)
253
+ return batch_idx
254
+
255
+
256
+ def write_train_metric(summary_writer, train_metrics, train_time, step):
257
+ summary_writer.scalar("train_time", train_time, step)
258
+
259
+ train_metrics = get_metrics(train_metrics)
260
+ for key, vals in train_metrics.items():
261
+ tag = f"train_{key}"
262
+ for i, val in enumerate(vals):
263
+ summary_writer.scalar(tag, val, step - len(vals) + i + 1)
264
+
265
+
266
+ def write_eval_metric(summary_writer, eval_metrics, step):
267
+ for metric_name, value in eval_metrics.items():
268
+ summary_writer.scalar(f"eval_{metric_name}", value, step)
269
+
270
+
271
+ if __name__ == "__main__":
272
+ # See all possible arguments in src/transformers/training_args.py
273
+ # or by passing the --help flag to this script.
274
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
275
+
276
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
277
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
278
+ # If we pass only one argument to the script and it's the path to a json file,
279
+ # let's parse it to get our arguments.
280
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
281
+ else:
282
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
283
+
284
+ if (
285
+ os.path.exists(training_args.output_dir)
286
+ and os.listdir(training_args.output_dir)
287
+ and training_args.do_train
288
+ and not training_args.overwrite_output_dir
289
+ ):
290
+ raise ValueError(
291
+ f"Output directory ({training_args.output_dir}) already exists and is not empty."
292
+ "Use --overwrite_output_dir to overcome."
293
+ )
294
+
295
+ # Setup logging
296
+ logging.basicConfig(
297
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
298
+ level="NOTSET",
299
+ datefmt="[%X]",
300
+ )
301
+
302
+ # Log on each process the small summary:
303
+ logger = logging.getLogger(__name__)
304
+
305
+ # Set the verbosity to info of the Transformers logger (on main process only):
306
+ logger.info(f"Training/evaluation parameters {training_args}")
307
+
308
+ # Set seed before initializing model.
309
+ set_seed(training_args.seed)
310
+
311
+ # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
312
+ # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
313
+ # (the dataset will be downloaded automatically from the datasets Hub).
314
+ #
315
+ # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
316
+ # 'text' is found. You can easily tweak this behavior (see below).
317
+ #
318
+ # In distributed training, the load_dataset function guarantees that only one local process can concurrently
319
+ # download the dataset.
320
+ if data_args.dataset_name is not None:
321
+ # Downloading and loading a dataset from the hub.
322
+ datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
323
+
324
+ if "validation" not in datasets.keys():
325
+ datasets["validation"] = load_dataset(
326
+ data_args.dataset_name,
327
+ data_args.dataset_config_name,
328
+ split=f"train[:{data_args.validation_split_percentage}%]",
329
+ cache_dir=model_args.cache_dir,
330
+ )
331
+ datasets["train"] = load_dataset(
332
+ data_args.dataset_name,
333
+ data_args.dataset_config_name,
334
+ split=f"train[{data_args.validation_split_percentage}%:]",
335
+ cache_dir=model_args.cache_dir,
336
+ )
337
+ else:
338
+ data_files = {}
339
+ if data_args.train_file is not None:
340
+ data_files["train"] = data_args.train_file
341
+ if data_args.validation_file is not None:
342
+ data_files["validation"] = data_args.validation_file
343
+ extension = data_args.train_file.split(".")[-1]
344
+ if extension == "txt":
345
+ extension = "text"
346
+ datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
347
+
348
+ if "validation" not in datasets.keys():
349
+ datasets["validation"] = load_dataset(
350
+ extension,
351
+ data_files=data_files,
352
+ split=f"train[:{data_args.validation_split_percentage}%]",
353
+ cache_dir=model_args.cache_dir,
354
+ )
355
+ datasets["train"] = load_dataset(
356
+ extension,
357
+ data_files=data_files,
358
+ split=f"train[{data_args.validation_split_percentage}%:]",
359
+ cache_dir=model_args.cache_dir,
360
+ )
361
+ # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
362
+ # https://huggingface.co/docs/datasets/loading_datasets.html.
363
+
364
+ # Load pretrained model and tokenizer
365
+
366
+ # Distributed training:
367
+ # The .from_pretrained methods guarantee that only one local process can concurrently
368
+ # download model & vocab.
369
+ if model_args.config_name:
370
+ config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
371
+ elif model_args.model_name_or_path:
372
+ config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
373
+ else:
374
+ config = CONFIG_MAPPING[model_args.model_type]()
375
+ logger.warning("You are instantiating a new config instance from scratch.")
376
+
377
+ if model_args.tokenizer_name:
378
+ tokenizer = AutoTokenizer.from_pretrained(
379
+ model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
380
+ )
381
+ elif model_args.model_name_or_path:
382
+ tokenizer = AutoTokenizer.from_pretrained(
383
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
384
+ )
385
+ else:
386
+ raise ValueError(
387
+ "You are instantiating a new tokenizer from scratch. This is not supported by this script."
388
+ "You can do it from another script, save it, and load it from here, using --tokenizer_name."
389
+ )
390
+
391
+ # Preprocessing the datasets.
392
+ # First we tokenize all the texts.
393
+ if training_args.do_train:
394
+ column_names = datasets["train"].column_names
395
+ else:
396
+ column_names = datasets["validation"].column_names
397
+ text_column_name = "text" if "text" in column_names else column_names[0]
398
+
399
+ max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
400
+
401
+ if data_args.line_by_line:
402
+ # When using line_by_line, we just tokenize each nonempty line.
403
+ padding = "max_length" if data_args.pad_to_max_length else False
404
+
405
+ def tokenize_function(examples):
406
+ # Remove empty lines
407
+ examples = [line for line in examples if len(line) > 0 and not line.isspace()]
408
+ return tokenizer(
409
+ examples,
410
+ return_special_tokens_mask=True,
411
+ padding=padding,
412
+ truncation=True,
413
+ max_length=max_seq_length,
414
+ )
415
+
416
+ tokenized_datasets = datasets.map(
417
+ tokenize_function,
418
+ input_columns=[text_column_name],
419
+ batched=True,
420
+ num_proc=data_args.preprocessing_num_workers,
421
+ remove_columns=column_names,
422
+ load_from_cache_file=not data_args.overwrite_cache,
423
+ )
424
+
425
+ else:
426
+ # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
427
+ # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
428
+ # efficient when it receives the `special_tokens_mask`.
429
+ def tokenize_function(examples):
430
+ return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
431
+
432
+ tokenized_datasets = datasets.map(
433
+ tokenize_function,
434
+ batched=True,
435
+ num_proc=data_args.preprocessing_num_workers,
436
+ remove_columns=column_names,
437
+ load_from_cache_file=not data_args.overwrite_cache,
438
+ )
439
+
440
+ # Main data processing function that will concatenate all texts from our dataset and generate chunks of
441
+ # max_seq_length.
442
+ def group_texts(examples):
443
+ # Concatenate all texts.
444
+ concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
445
+ total_length = len(concatenated_examples[list(examples.keys())[0]])
446
+ # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
447
+ # customize this part to your needs.
448
+ if total_length >= max_seq_length:
449
+ total_length = (total_length // max_seq_length) * max_seq_length
450
+ # Split by chunks of max_len.
451
+ result = {
452
+ k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
453
+ for k, t in concatenated_examples.items()
454
+ }
455
+ return result
456
+
457
+ # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
458
+ # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
459
+ # might be slower to preprocess.
460
+ #
461
+ # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
462
+ # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
463
+ tokenized_datasets = tokenized_datasets.map(
464
+ group_texts,
465
+ batched=True,
466
+ num_proc=data_args.preprocessing_num_workers,
467
+ load_from_cache_file=not data_args.overwrite_cache,
468
+ )
469
+
470
+ # Enable tensorboard only on the master node
471
+ has_tensorboard = is_tensorboard_available()
472
+ if has_tensorboard and jax.process_index() == 0:
473
+ try:
474
+ from flax.metrics.tensorboard import SummaryWriter
475
+
476
+ summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
477
+ except ImportError as ie:
478
+ has_tensorboard = False
479
+ logger.warning(
480
+ f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
481
+ )
482
+ else:
483
+ logger.warning(
484
+ "Unable to display metrics through TensorBoard because the package is not installed: "
485
+ "Please run pip install tensorboard to enable."
486
+ )
487
+
488
+ # Data collator
489
+ # This one will take care of randomly masking the tokens.
490
+ data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
491
+
492
+ # Initialize our training
493
+ rng = jax.random.PRNGKey(training_args.seed)
494
+ dropout_rngs = jax.random.split(rng, jax.local_device_count())
495
+
496
+ if model_args.model_name_or_path:
497
+ model = FlaxAutoModelForMaskedLM.from_pretrained(
498
+ model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
499
+ )
500
+ else:
501
+ model = FlaxAutoModelForMaskedLM.from_config(
502
+ config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
503
+ )
504
+
505
+ # Store some constant
506
+ num_epochs = int(training_args.num_train_epochs)
507
+ train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
508
+ eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
509
+
510
+ num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
511
+
512
+ # Create learning rate schedule
513
+ warmup_fn = optax.linear_schedule(
514
+ init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
515
+ )
516
+ decay_fn = optax.linear_schedule(
517
+ init_value=training_args.learning_rate,
518
+ end_value=0,
519
+ transition_steps=num_train_steps - training_args.warmup_steps,
520
+ )
521
+ linear_decay_lr_schedule_fn = optax.join_schedules(
522
+ schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
523
+ )
524
+
525
+ # We use Optax's "masking" functionality to not apply weight decay
526
+ # to bias and LayerNorm scale parameters. decay_mask_fn returns a
527
+ # mask boolean with the same structure as the parameters.
528
+ # The mask is True for parameters that should be decayed.
529
+ # Note that this mask is specifically adapted for FlaxBERT-like models.
530
+ # For other models, one should correct the layer norm parameter naming
531
+ # accordingly.
532
+ def decay_mask_fn(params):
533
+ flat_params = traverse_util.flatten_dict(params)
534
+ flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
535
+ return traverse_util.unflatten_dict(flat_mask)
536
+
537
+ # create adam optimizer
538
+ if training_args.adafactor:
539
+ # We use the default parameters here to initialize adafactor,
540
+ # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
541
+ optimizer = optax.adafactor(
542
+ learning_rate=linear_decay_lr_schedule_fn,
543
+ )
544
+ else:
545
+ optimizer = optax.adamw(
546
+ learning_rate=linear_decay_lr_schedule_fn,
547
+ b1=training_args.adam_beta1,
548
+ b2=training_args.adam_beta2,
549
+ eps=training_args.adam_epsilon,
550
+ weight_decay=training_args.weight_decay,
551
+ mask=decay_mask_fn,
552
+ )
553
+
554
+ # Setup train state
555
+ state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
556
+
557
+ # Define gradient update step fn
558
+ def train_step(state, batch, dropout_rng):
559
+ dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
560
+
561
+ def loss_fn(params):
562
+ labels = batch.pop("labels")
563
+
564
+ logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
565
+
566
+ # compute loss, ignore padded input tokens
567
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
568
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
569
+
570
+ # take average
571
+ loss = loss.sum() / label_mask.sum()
572
+
573
+ return loss
574
+
575
+ grad_fn = jax.value_and_grad(loss_fn)
576
+ loss, grad = grad_fn(state.params)
577
+ grad = jax.lax.pmean(grad, "batch")
578
+ new_state = state.apply_gradients(grads=grad)
579
+
580
+ metrics = jax.lax.pmean(
581
+ {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
582
+ )
583
+
584
+ return new_state, metrics, new_dropout_rng
585
+
586
+ # Create parallel version of the train step
587
+ p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
588
+
589
+ # Define eval fn
590
+ def eval_step(params, batch):
591
+ labels = batch.pop("labels")
592
+
593
+ logits = model(**batch, params=params, train=False)[0]
594
+
595
+ # compute loss, ignore padded input tokens
596
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
597
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
598
+
599
+ # compute accuracy
600
+ accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
601
+
602
+ # summarize metrics
603
+ metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
604
+ metrics = jax.lax.psum(metrics, axis_name="batch")
605
+
606
+ return metrics
607
+
608
+ p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
609
+
610
+ # Replicate the train state on each device
611
+ state = jax_utils.replicate(state)
612
+
613
+ train_time = 0
614
+ epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
615
+ for epoch in epochs:
616
+ # ======================== Training ================================
617
+ train_start = time.time()
618
+ train_metrics = []
619
+
620
+ # Create sampling rng
621
+ rng, input_rng = jax.random.split(rng)
622
+
623
+ # Generate an epoch by shuffling sampling indices from the train dataset
624
+ num_train_samples = len(tokenized_datasets["train"])
625
+ train_samples_idx = jax.random.permutation(input_rng, jnp.arange(num_train_samples))
626
+ train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)
627
+
628
+ # Gather the indexes for creating the batch and do a training step
629
+ for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
630
+ samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
631
+ model_inputs = data_collator(samples, pad_to_multiple_of=16)
632
+
633
+ # Model forward
634
+ model_inputs = shard(model_inputs.data)
635
+ state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
636
+ train_metrics.append(train_metric)
637
+
638
+ cur_step = epoch * (num_train_samples // train_batch_size) + step
639
+
640
+ if cur_step % training_args.logging_steps == 0 and cur_step > 0:
641
+ # Save metrics
642
+ train_metric = jax_utils.unreplicate(train_metric)
643
+ train_time += time.time() - train_start
644
+ if has_tensorboard and jax.process_index() == 0:
645
+ write_train_metric(summary_writer, train_metrics, train_time, cur_step)
646
+
647
+ epochs.write(
648
+ f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
649
+ )
650
+
651
+ train_metrics = []
652
+
653
+ if cur_step % training_args.eval_steps == 0 and cur_step > 0:
654
+ # ======================== Evaluating ==============================
655
+ num_eval_samples = len(tokenized_datasets["validation"])
656
+ eval_samples_idx = jnp.arange(num_eval_samples)
657
+ eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
658
+
659
+ eval_metrics = []
660
+ for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
661
+ samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
662
+ model_inputs = data_collator(samples, pad_to_multiple_of=16)
663
+
664
+ # Model forward
665
+ model_inputs = shard(model_inputs.data)
666
+ metrics = p_eval_step(state.params, model_inputs)
667
+ eval_metrics.append(metrics)
668
+
669
+ # normalize eval metrics
670
+ eval_metrics = get_metrics(eval_metrics)
671
+ eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
672
+ eval_normalizer = eval_metrics.pop("normalizer")
673
+ eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
674
+
675
+ # Update progress bar
676
+ epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
677
+
678
+ # Save metrics
679
+ if has_tensorboard and jax.process_index() == 0:
680
+ write_eval_metric(summary_writer, eval_metrics, cur_step)
681
+
682
+ if cur_step % training_args.save_steps == 0 and cur_step > 0:
683
+ # save checkpoint after each epoch and push checkpoint to the hub
684
+ if jax.process_index() == 0:
685
+ params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
686
+ model.save_pretrained(
687
+ training_args.output_dir,
688
+ params=params,
689
+ push_to_hub=training_args.push_to_hub,
690
+ commit_message=f"Saving weights and logs of step {cur_step}",
691
+ )
train_tokenizer.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from datasets import load_dataset
2
+ from transformers import AutoConfig, AutoTokenizer
3
+ from tokenizers import BertWordPieceTokenizer
4
+
5
+
6
+ config = AutoConfig.from_pretrained("./")
7
+
8
+ # load dataset
9
+ dataset = load_dataset("flax-community/swahili-safi", split="train")
10
+
11
+
12
+ def batch_iterator(batch_size=1000):
13
+ for i in range(0, len(dataset), batch_size):
14
+ yield dataset[i: i + batch_size]["text"]
15
+
16
+
17
+ # Instantiate tokenizer
18
+ tokenizer = BertWordPieceTokenizer(
19
+ clean_text=False,
20
+ handle_chinese_chars=False,
21
+ strip_accents=False,
22
+ lowercase=True,
23
+ )
24
+
25
+ # Customized training
26
+ tokenizer.train_from_iterator(
27
+ batch_iterator(),
28
+ vocab_size=config.vocab_size,
29
+ min_frequency=2,
30
+ special_tokens=['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[MASK]'],
31
+ limit_alphabet=1000,
32
+ wordpieces_prefix="##"
33
+ )
34
+
35
+ # Save files to disk
36
+ tokenizer.save("tokenizer.json")
37
+ tokenizer.save_model("./")
38
+
39
+ # Resave in HF Format
40
+ tokenizer = AutoTokenizer.from_pretrained("./")
41
+ tokenizer.save_pretrained("./")