File size: 34,221 Bytes
f291f93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
731244f
f291f93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
731244f
f291f93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
731244f
f291f93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Team All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
text file or a dataset.

Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
https://huggingface.co/models?filter=masked-lm
"""
import logging
import os
import sys
import time
from dataclasses import dataclass, field
from optax import clip_by_global_norm
# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
from pathlib import Path
from typing import Dict, List, Optional, Tuple

import numpy as np
from datasets import load_dataset, DatasetDict
from tqdm import tqdm

import flax
import jax
import jax.numpy as jnp
import optax
from flax import jax_utils, traverse_util
from flax.training import train_state
from flax.training.common_utils import get_metrics, onehot, shard
from transformers import (
    CONFIG_MAPPING,
    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
    AutoConfig,
    AutoTokenizer,
    FlaxAutoModelForMaskedLM,
    HfArgumentParser,
    PreTrainedTokenizerBase,
    TensorType,
    TrainingArguments,
    is_tensorboard_available,
    set_seed,
)
import json
from flax.training import checkpoints
from flax.jax_utils import unreplicate
from flax.training.checkpoints import save_checkpoint, restore_checkpoint
from importlib.util import find_spec


MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)


@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
    """

    model_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": "The model checkpoint for weights initialization."
            "Don't set if you want to train a model from scratch."
        },
    )
    model_type: Optional[str] = field(
        default=None,
        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    dtype: Optional[str] = field(
        default="float32",
        metadata={
            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
        },
    )


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

    dataset_name: Optional[str] = field(
        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    train_ref_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
    )
    validation_ref_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )



    validation_split_percentage: Optional[int] = field(
        default=5,
        metadata={
            "help": "The percentage of the train set used as validation set in case there's no validation split"
        },
    )
    max_seq_length: Optional[int] = field(
        default=None,
        metadata={
            "help": "The maximum total input sequence length after tokenization. Sequences longer "
            "than this will be truncated. Default to the max input length of the model."
        },
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    mlm_probability: float = field(
        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
    )
    pad_to_max_length: bool = field(
        default=False,
        metadata={
            "help": "Whether to pad all samples to `max_seq_length`. "
            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
        },
    )
    line_by_line: bool = field(
        default=False,
        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
            "value if set."
        },
    )

    




@flax.struct.dataclass
class FlaxDataCollatorForLanguageModeling:
    """
    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
    are not all of the same length.

    Args:
        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
            The tokenizer used for encoding the data.
        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
            The probability with which to (randomly) mask tokens in the input.

    .. note::

        For best performance, this data collator should be used with a dataset having items that are dictionaries or
        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
        argument :obj:`return_special_tokens_mask=True`.
    """

    tokenizer: PreTrainedTokenizerBase
    mlm_probability: float = 0.15

    def __post_init__(self):
        if self.tokenizer.mask_token is None:
            raise ValueError(
                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
                "You should pass `mlm=False` to train on causal language modeling instead."
            )

    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
        # Handle dict or lists with proper padding and conversion to tensor.
        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)

        # If special token mask has been preprocessed, pop it from the dict.
        special_tokens_mask = batch.pop("special_tokens_mask", None)

        batch["input_ids"], batch["labels"] = self.mask_tokens(
            batch["input_ids"], special_tokens_mask=special_tokens_mask
        )
        return batch

    def mask_tokens(
        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
    ) -> Tuple[jnp.ndarray, jnp.ndarray]:
        """
        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
        """
        labels = inputs.copy()
        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
        probability_matrix = np.full(labels.shape, self.mlm_probability)
        special_tokens_mask = special_tokens_mask.astype("bool")

        probability_matrix[special_tokens_mask] = 0.0
        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
        labels[~masked_indices] = -100  # We only compute loss on masked tokens

        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)

        # 10% of the time, we replace masked input tokens with random word
        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
        indices_random &= masked_indices & ~indices_replaced

        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
        inputs[indices_random] = random_words[indices_random]

        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
        return inputs, labels


def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
    num_samples = len(samples_idx)
    samples_to_remove = num_samples % batch_size

    if samples_to_remove != 0:
        samples_idx = samples_idx[:-samples_to_remove]
    sections_split = num_samples // batch_size
    batch_idx = np.split(samples_idx, sections_split)
    return batch_idx


def write_train_metric(summary_writer, train_metrics, train_time, step):
    summary_writer.scalar("train_time", train_time, step)

    train_metrics = get_metrics(train_metrics)
    for key, vals in train_metrics.items():
        tag = f"train_{key}"
        for i, val in enumerate(vals):
            summary_writer.scalar(tag, val, step - len(vals) + i + 1)


def write_eval_metric(summary_writer, eval_metrics, step):
    for metric_name, value in eval_metrics.items():
        summary_writer.scalar(f"eval_{metric_name}", value, step)

def rotate_checkpoints(ckpt_dir:str, save_total_limit:int):
    "Removes older checkpoints so that `save_total_limit` checkpoints are kept"
    # TODO: what to remove is decided using step number only, we might want to improve that
    ckpts = [str(x) for x in Path(ckpt_dir).glob("ckpt-*")]
    # sort checkpoints by step
    ckpts_sorted = sorted(ckpts, key=lambda x: int(x.split('-')[-1]))
    ckpts_to_delete = ckpts_sorted[:-save_total_limit]
    for ckpt in ckpts_to_delete:
        logger.info(f"Deleting older checkpoint [{ckpt}] due to save_total_limit ({save_total_limit})")
        shutil.rmtree(ckpt)


if __name__ == "__main__":
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        # If we pass only one argument to the script and it's the path to a json file,
        # let's parse it to get our arguments.
        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if (
        os.path.exists(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty."
            "Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        level="NOTSET",
        datefmt="[%X]",
    )

    # Log on each process the small summary:
    logger = logging.getLogger(__name__)

    # Set the verbosity to info of the Transformers logger (on main process only):
    logger.info(f"Training/evaluation parameters {training_args}")

    # Set seed before initializing model.
    set_seed(training_args.seed)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)

        if "validation" not in datasets.keys():
            datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )
    else:
        import glob
        import random
        data_files = []
        def add_jsonlines_dir(path, filespec):
            global data_files
            data_files += glob.glob(f"{path}/{filespec}")
            data_files = list(set(data_files))
            print(f"Number of files {len(data_files)} after adding {path} glob {filespec}")
        add_jsonlines_dir(f"/data/c4_cleaned2", "*.gz")
        add_jsonlines_dir(f"/data/nrc_uniq_cleaned_20210223", "*.gz")
        add_jsonlines_dir(f"/data/nu_uniq_cleaned_20210225", "*.gz")
        random.Random(42).shuffle(data_files)
        total = len(data_files)
        print(total)
        perc = 0.05
        val_size = int(perc * total)
        train_size = total - val_size
        train = data_files[:train_size]
        val = data_files[train_size:]
        print(f"Got {len(train)} training files and {perc * 100} % {len(val)} validation files")
        assert list(set(train) & set(val)) == [], "Train overlaps with test"
        load_grouped = True
        if not load_grouped:
            datasets = load_dataset('json', data_files={'train': train, 'validation': val})
        
    #from datasets import Dataset

    #dataset = Dataset.from_file("/home/dat/.cache/huggingface/datasets/json/default-9add402b38836560/0.0.0/f92a4de297ac644ad9781979b79064b0e222b3af766f8ea3bee32390dca23723/json-train.arrow") 
    #dataset = Dataset.from_file("/home/dat/.cache/huggingface/datasets/json/default-9add402b38836560/0.0.0/f92a4de297ac644ad9781979b79064b0e222b3af766f8ea3bee32390dca23723/json-validation.arrow") 


    def mb_item(x):
        return x.item() if hasattr(x, "item") else x

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer

    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
        )
    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
        )
    else:
        raise ValueError(
            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
        )

    # Preprocessing the datasets.
    # First we tokenize all the texts.

    if load_grouped:
        logger.info("Loading tokenized and grouped dataset")
        tokenized_datasets = DatasetDict.load_from_disk("/data/tokenized_data")
        logger.info("Setting max validation examples to ")
        print(f"Number of validation examples {data_args.max_eval_samples}")
        #tokenized_datasets["train"]= tokenized_datasets["train"].select(range(int(0.3*len(tokenized_datasets["train"]))))
        if data_args.max_eval_samples is not None:
            tokenized_datasets["validation"] = tokenized_datasets["validation"].select(range(data_args.max_eval_samples))
    else:
        if training_args.do_train:
            column_names = datasets["train"].column_names
        else:
            column_names = datasets["validation"].column_names
        text_column_name = "text" if "text" in column_names else column_names[0]

        max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)

        if data_args.line_by_line:
            # When using line_by_line, we just tokenize each nonempty line.
            padding = "max_length" if data_args.pad_to_max_length else False

            def tokenize_function(examples):
                # Remove empty lines
                examples = [line for line in examples if len(line) > 0 and not line.isspace()]
                return tokenizer(
                    examples,
                    return_special_tokens_mask=True,
                    padding=padding,
                    truncation=True,
                    max_length=max_seq_length,
                )

            tokenized_datasets = datasets.map(
                tokenize_function,
                input_columns=[text_column_name],
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
            )

        else:
            # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
            # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
            # efficient when it receives the `special_tokens_mask`.
            def tokenize_function(examples):
                return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

            tokenized_datasets = datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
            )

            # Main data processing function that will concatenate all texts from our dataset and generate chunks of
            # max_seq_length.
            def group_texts(examples):
                # Concatenate all texts.
                concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
                total_length = len(concatenated_examples[list(examples.keys())[0]])
                # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
                # customize this part to your needs.
                if total_length >= max_seq_length:
                    total_length = (total_length // max_seq_length) * max_seq_length
                # Split by chunks of max_len.
                result = {
                    k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                    for k, t in concatenated_examples.items()
                }
                return result

            # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
            # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
            # might be slower to preprocess.
            #
            # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
            # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
            tokenized_datasets = tokenized_datasets.map(
                group_texts,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
            )

            #tokenized_datasets.save_to_disk("/data/tokenized_data")
            #print ("tokenized_datasets saved to disk")

    
    # Enable tensorboard only on the master node
    has_tensorboard = is_tensorboard_available()
    if has_tensorboard and jax.process_index() == 0:
        try:
            from flax.metrics.tensorboard import SummaryWriter

            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
        except ImportError as ie:
            has_tensorboard = False
            logger.warning(
                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
            )
    else:
        logger.warning(
            "Unable to display metrics through TensorBoard because the package is not installed: "
            "Please run pip install tensorboard to enable."
        )
    has_wandb = find_spec("wandb") is not None 
    if jax.process_index() == 0 and has_wandb and ("wandb" in training_args.report_to):
        try:
            import wandb
            wandb.init(
                entity="wandb", 
                project="hf-flax-pino-roberta",
                sync_tensorboard=True
            )
            wandb.config.update(training_args)
            wandb.config.update(model_args)
            wandb.config.update(data_args)
        except ImportError as e:
            print(e)
            has_wandb = False
    # Data collator
    # This one will take care of randomly masking the tokens.
    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)

    # Initialize our training
    rng = jax.random.PRNGKey(training_args.seed)
    dropout_rngs = jax.random.split(rng, jax.local_device_count())

    if model_args.model_name_or_path:
        model = FlaxAutoModelForMaskedLM.from_pretrained(
            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
        )
    else:
        model = FlaxAutoModelForMaskedLM.from_config(
            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
        )

    # Store some constant
    num_epochs = int(training_args.num_train_epochs)
    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()

    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
    
    # Create learning rate schedule
    warmup_fn = optax.linear_schedule(
        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
    )
    decay_fn = optax.linear_schedule(
        init_value=training_args.learning_rate,
        end_value=0,
        transition_steps=num_train_steps - training_args.warmup_steps,
    )
    linear_decay_lr_schedule_fn = optax.join_schedules(
        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
    )

    # We use Optax's "masking" functionality to not apply weight decay
    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
    # mask boolean with the same structure as the parameters.
    # The mask is True for parameters that should be decayed.
    # Note that this mask is specifically adapted for FlaxBERT-like models.
    # For other models, one should correct the layer norm parameter naming
    # accordingly.
    def decay_mask_fn(params):
        flat_params = traverse_util.flatten_dict(params)
        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
        return traverse_util.unflatten_dict(flat_mask)

    # create adam optimizer
    if training_args.adafactor:
        # We use the default parameters here to initialize adafactor,
        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
        optimizer = optax.adafactor(
            learning_rate=linear_decay_lr_schedule_fn,
        )
    else:
        optimizer = optax.adamw(
            learning_rate=linear_decay_lr_schedule_fn,
            b1=training_args.adam_beta1,
            b2=training_args.adam_beta2,
            eps=training_args.adam_epsilon,
            weight_decay=training_args.weight_decay,
            mask=decay_mask_fn,
        )
        optimizer = optax.chain(
            optax.clip_by_global_norm(1.),
            optimizer
        )

    # Setup train state
    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)

    if training_args.resume_from_checkpoint:
        state = restore_checkpoint(training_args.resume_from_checkpoint, state)
        resume_step = mb_item(state.step.item())
    else:
        resume_step = 0


    # Define gradient update step fn
    def train_step(state, batch, dropout_rng):
        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)

        def loss_fn(params):
            labels = batch.pop("labels")

            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]

            # compute loss, ignore padded input tokens
            label_mask = jnp.where(labels > 0, 1.0, 0.0)
            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask

            # take average
            loss = loss.sum() / label_mask.sum()

            return loss

        grad_fn = jax.value_and_grad(loss_fn)
        loss, grad = grad_fn(state.params)
        grad = jax.lax.pmean(grad, "batch")
        new_state = state.apply_gradients(grads=grad)

        metrics = jax.lax.pmean(
            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
        )

        return new_state, metrics, new_dropout_rng

    # Create parallel version of the train step
    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))

    # Define eval fn
    def eval_step(params, batch):
        labels = batch.pop("labels")

        logits = model(**batch, params=params, train=False)[0]

        # compute loss, ignore padded input tokens
        label_mask = jnp.where(labels > 0, 1.0, 0.0)
        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask

        # compute accuracy
        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask

        # summarize metrics
        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
        metrics = jax.lax.psum(metrics, axis_name="batch")

        return metrics

    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))

    # Replicate the train state on each device
    state = jax_utils.replicate(state)

    train_time = 0
    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
    for epoch in epochs:
        # ======================== Training ================================
        train_start = time.time()
        train_metrics = []

        # Create sampling rng
        rng, input_rng = jax.random.split(rng)

        # Generate an epoch by shuffling sampling indices from the train dataset
        num_train_samples = len(tokenized_datasets["train"])
        train_samples_idx = jax.random.permutation(input_rng, jnp.arange(num_train_samples))
        train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)

        # Gather the indexes for creating the batch and do a training step
        for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1,initial=resume_step)):
            samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
            model_inputs = data_collator(samples, pad_to_multiple_of=16)

            # Model forward
            model_inputs = shard(model_inputs.data)
            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
            train_metrics.append(train_metric)

            cur_step = epoch * (num_train_samples // train_batch_size) + step
            if cur_step < resume_step:
                continue

            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
                # Save metrics
                train_metric = jax_utils.unreplicate(train_metric)
                train_time += time.time() - train_start
                if has_tensorboard and jax.process_index() == 0:
                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)

                if has_wandb and jax.process_index() == 0 and ("wandb" in training_args.report_to):
                    # TODO: add accumulation of metrics
                    _metrics = {k if k=="learning_rate" else f"train_{k}":mb_item(v.mean()) for k, v in train_metric.items()}
                    wandb.log({"training_step":cur_step, **_metrics}, commit=True)

                epochs.write(
                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
                )

                train_metrics = []

            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
                # ======================== Evaluating ==============================
                num_eval_samples = len(tokenized_datasets["validation"])
                eval_samples_idx = jnp.arange(num_eval_samples)
                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)

                eval_metrics = []
                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
                    model_inputs = data_collator(samples, pad_to_multiple_of=16)

                    # Model forward
                    model_inputs = shard(model_inputs.data)
                    metrics = p_eval_step(state.params, model_inputs)
                    eval_metrics.append(metrics)

                # normalize eval metrics
                eval_metrics = get_metrics(eval_metrics)
                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
                eval_normalizer = eval_metrics.pop("normalizer")
                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)

                # Update progress bar
                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"

                # Save metrics
                if has_tensorboard and jax.process_index() == 0:
                    write_eval_metric(summary_writer, eval_metrics, cur_step)
                if has_wandb and jax.process_index() == 0 and ("wandb" in training_args.report_to):
                    _metrics = {f"eval_{k}":mb_item(v) for k, v in eval_metrics.items()}
                    wandb.log({"eval_step":cur_step, **_metrics})

            if cur_step % training_args.save_steps == 0 and cur_step > 0:
                # save checkpoint after each epoch and push checkpoint to the hub
                if jax.process_index() == 0:
                    save_checkpoint(training_args.output_dir, jax_utils.unreplicate(state), cur_step, keep=training_args.save_total_limit, overwrite=True)
                    params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
                    model.save_pretrained(
                        training_args.output_dir,
                        params=params,
                        push_to_hub=training_args.push_to_hub,
                        commit_message=f"Saving weights and logs of step {cur_step}",
                    )
                    if training_args.save_total_limit is not None:
                        rotate_checkpoints(training_args.output_dir, training_args.save_total_limit)
    if jax.process_index() == 0:
        params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
        model.save_pretrained(
            training_args.output_dir,
            params=params,
            push_to_hub=training_args.push_to_hub,
            commit_message=f"Saving weights and logs of step {cur_step}",
            )