-
Notifications
You must be signed in to change notification settings - Fork 0
/
maybe_iter.py
815 lines (717 loc) · 31.1 KB
/
maybe_iter.py
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
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
# if charmap error: copy $env:PYTHONUTF8="1" in terminal
import argparse
import math
import os
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, List, Optional, Union
import mne
import random
import datasets
import torch
from accelerate import Accelerator
from accelerate.logging import get_logger
from datasets import DatasetDict, concatenate_datasets, load_dataset
from huggingface_hub import HfApi
from torch.utils.data.dataloader import DataLoader
from tqdm.auto import tqdm
import transformers
from transformers import (
AdamW,
SchedulerType,
Wav2Vec2Config,
Wav2Vec2FeatureExtractor,
Wav2Vec2ForPreTraining,
Wav2Vec2ForCTC,
get_scheduler,
is_wandb_available,
set_seed,
)
from transformers.models.wav2vec2.modeling_wav2vec2 import _compute_mask_indices, _sample_negative_indices
from transformers.utils import send_example_telemetry
from torch.utils.tensorboard import SummaryWriter, writer
from torch.utils.data import DataLoader, TensorDataset
from datasets import IterableDataset
from tqdm import tqdm
# from loadmp3 import BIDSBrainVisionDataset
#product quantization in code
logger = get_logger(__name__)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Training on {device}")
def parse_args():
parser = argparse.ArgumentParser(description="Finetune a transformers model on a text classification task")
parser.add_argument(
"--dataset_name",
type=str,
default="hf-internal-testing/librispeech_asr_dummy", # MLCommons/peoples_speech
help="The name of the dataset to use (via the datasets library).",
)
parser.add_argument(
"--dataset_config_names",
nargs="+",
type=str,
required=False,
default=["clean"],
help="The configuration names of the dataset to use (via the datasets library).",
)
parser.add_argument(
"--dataset_split_names",
nargs="+",
type=str,
required=False,
default=["validation", "test"], #each about 600h (30k in total)
help="The names of the training data set splits to use (via the datasets library).",
)
parser.add_argument(
"--trust_remote_code",
action="store_true",
help=(
"Whether to trust the execution of code from datasets/models defined on the Hub."
" This option should only be set to `True` for repositories you trust and in which you have read the"
" code, as it will execute code present on the Hub on your local machine."
),
)
parser.add_argument(
"--preprocessing_num_workers",
type=int,
default=None,
help="The number of processes to use for the preprocessing.",
)
parser.add_argument(
"--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
)
parser.add_argument(
"--preprocessing_only",
action="store_true",
help="Only run the preprocessing script to be cached for future use",
)
parser.add_argument(
"--cache_dir",
type=str,
default=None,
help="Where do you want to store the pretrained models downloaded from huggingface.co",
)
parser.add_argument(
"--validation_split_percentage",
type=int,
default=10,
help="Percentage of training data that should be used for validation if no validation is present in dataset.",
)
parser.add_argument(
"--logging_steps",
type=int,
default=2,
help="Number of steps between each logging",
)
parser.add_argument(
"--saving_steps",
type=int,
default=10,
help="Number of steps between each logging",
)
parser.add_argument(
"--audio_column_name",
type=str,
default="audio",
help="Column in the dataset that contains speech file path. Defaults to 'audio'",
)
parser.add_argument(
"--model_name_or_path",
type=str,
default="patrickvonplaten/wav2vec2-base-v2",
help="Path to pretrained model or model identifier from huggingface.co/models.",
required=False,
)
parser.add_argument(
"--config_name",
type=str,
default=None,
help="Pretrained config name or path if not the same as model_name",
)
parser.add_argument(
"--train_cache_file_name",
type=str,
default=None,
help="Path to the train cached file name",
)
parser.add_argument(
"--validation_cache_file_name",
type=str,
default=None,
help="Path to the validation cached file name",
)
parser.add_argument(
"--per_device_train_batch_size",
type=int,
default=4,
help="Batch size (per device) for the training dataloader.",
)
parser.add_argument(
"--per_device_eval_batch_size",
type=int,
default=8,
help="Batch size (per device) for the evaluation dataloader.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=0.001,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument("--weight_decay", type=float, default=0.01, help="Weight decay to use.")
parser.add_argument("--num_train_epochs", type=int, default=1000, help="Total number of training epochs to perform.")
parser.add_argument(
"--max_train_steps",
type=int,
default=1000
,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=4,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="If True, use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--lr_scheduler_type",
type=SchedulerType,
default="linear",
help="The scheduler type to use.",
choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
)
parser.add_argument(
"--num_warmup_steps", type=int, default=32000, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument("--output_dir", type=str, default="./wav2vec2-pretrained-demo", help="Where to store the final model.")
parser.add_argument("--seed", type=int, default=0, help="A seed for reproducible training.")
parser.add_argument(
"--max_gumbel_temperature",
type=float,
default=2.0,
help="Maximum temperature for gumbel softmax.",
)
parser.add_argument(
"--min_gumbel_temperature",
type=float,
default=0.5,
help="Minimum temperature for gumbel softmax.",
)
parser.add_argument(
"--gumbel_temperature_decay", type=float, default=0.999995, help="Decay of gumbel temperature during training."
)
parser.add_argument(
"--max_duration_in_seconds",
type=float,
default=20.0,
help="Filter out audio files that are longer than `max_duration_in_seconds` seconds",
)
parser.add_argument(
"--min_duration_in_seconds",
type=float,
default=2.0,
help="Filter out audio files that are shorter than `min_duration_in_seconds` seconds",
)
parser.add_argument(
"--pad_to_multiple_of",
type=int,
default=None,
help=(
"If set will pad the sequence to a multiple of the provided value. This is especially useful to enable the"
" use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta)."
),
)
parser.add_argument(
"--adam_beta1",
type=float,
default=0.9,
help="Beta1 for AdamW optimizer",
)
parser.add_argument(
"--adam_beta2",
type=float,
default=0.98,
help="Beta2 for AdamW optimizer",
)
parser.add_argument(
"--adam_epsilon",
type=float,
default=1e-06,
help="Epsilon for AdamW optimizer",
)
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument(
"--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
)
parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--mask_time_prob",
type=float,
default=0.65,
help=(
"Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked in the"
" contrastive task. If omitted, will pull value from model config."
),
)
parser.add_argument(
"--mask_time_length",
type=int,
default=5,
help=(
"Length of each vector mask span to mask along the time axis in the contrastive task."
" If omitted, will pull value from model config."
),
)
args = parser.parse_args()
if args.push_to_hub:
assert args.output_dir is not None, "Need an `output_dir` to create a repo when `--push_to_hub` is passed."
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
return args
##################
class BIDSBrainVisionDataset(IterableDataset):
def __init__(self, filepaths, feature_extractor, target_sr=16000, window_size=2.0):
self.filepaths=filepaths
self.feature_extractor = feature_extractor
self.target_sr = target_sr
self.window_size = window_size
self._length = self._compute_length()
self.current_epoch = 0
def set_epoch(self, epoch):
self.current_epoch = epoch
def sliding_windows(self, data, sfreq):
step = int(self.window_size * sfreq)
data_length = len(data)
windows = [data[x:x + step] for x in range(0, data_length - step + 1, step)]
return windows
def _compute_length(self):
all_winds = 0
for filepath in self.filepaths:
ecog_channels, sfreq, _ = self._load_brainvision_file(filepath)
for channel_data in ecog_channels.values():
num_winds = len(self.sliding_windows(channel_data, sfreq))
all_winds += num_winds
return all_winds
def __iter__(self):
for filepath in self.filepaths:
ecog_channels, sfreq, available_channels = self._load_brainvision_file(filepath)
for channel_name, channel_data in ecog_channels.items():
windows = self.sliding_windows(channel_data, sfreq)
for window in windows:
inputs = self.feature_extractor(window,
sampling_rate=sfreq,
return_tensors="pt")
yield {
"input_values": inputs.input_values[0],
"sampling_rate": sfreq
}
def _load_brainvision_file(self, filepath):
raw = mne.io.read_raw_brainvision(filepath, preload=True)
available_channels = raw.ch_names
ecogs = {}
for channel in available_channels:
if "ECOG" not in channel and "LFP" not in channel:
continue
data, _ = raw[channel, :]
data = torch.tensor(data, dtype=torch.float32).squeeze()
ecogs[channel] = data
return ecogs, raw.info['sfreq'], list(ecogs.keys())
def __len__(self):
return self._length
##########################################################################
writer = SummaryWriter(log_dir="logging_events_real_data")
@dataclass
class DataCollatorForWav2Vec2Pretraining:
model: Wav2Vec2ForPreTraining
feature_extractor: Wav2Vec2FeatureExtractor
padding: Union[bool, str] = "longest"
pad_to_multiple_of: Optional[int] = None
mask_time_prob: Optional[float] = 0.65
mask_time_length: Optional[int] = 5
window_size_secs: float = 2.0 ########## bei windwos: 2 mal 16000=32000/input_values(320000) = 10 windows per batch
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
wind_features = []
for feature in features:
wind_input = self.feature_extractor(
feature["input_values"],
sampling_rate=feature["sampling_rate"],
return_tensors="pt"
)
wind_features.append({"input_values": wind_input.input_values[0]})
batch = self.feature_extractor.pad(
wind_features,
padding=self.padding,
return_tensors="pt",
)
# batch input_values shape = 10, 32000
device = batch["input_values"].device
batch_size = batch["input_values"].shape[0] # 10
mask_indices_seq_length = self.model._get_feat_extract_output_lengths(batch["input_values"].shape[-1])
# make sure masked sequence length is a Python scalar
mask_indices_seq_length = int(mask_indices_seq_length)
# make sure that no loss is computed on padded inputs
if batch.get("attention_mask") is not None:
# compute real output lengths according to convolution formula
batch["sub_attention_mask"] = self.model._get_feature_vector_attention_mask(
mask_indices_seq_length, batch["attention_mask"]
)
features_shape = (batch_size, mask_indices_seq_length)
# sample randomly masked indices
mask_time_indices = _compute_mask_indices(
features_shape,
self.mask_time_prob,
self.mask_time_length,
attention_mask=batch.get("sub_attention_mask"),
)
# sample negative indices
sampled_negative_indices = _sample_negative_indices(
features_shape,
self.model.config.num_negatives,
mask_time_indices=mask_time_indices,
)
batch["mask_time_indices"] = torch.tensor(mask_time_indices, dtype=torch.long, device=device)
batch["sampled_negative_indices"] = torch.tensor(sampled_negative_indices, dtype=torch.long, device=device)
return batch
# batch["input_values"].shape=[40, 32000]
# attention mask auch 40,32000; sub_attention_mask & mask_time_indices=[40,99] (4 wegen per device training batch param, 10 wegen sr & stop/step rechnung von windows)
# len eines windows: 32000
def multiply_grads(params, c):
"""Multiplies grads by a constant *c*."""
for p in params:
if p.grad is not None:
if torch.is_tensor(c):
c = c.to(p.grad.device)
p.grad.data.mul_(c)
def get_grad_norm(params, scale=1):
"""Compute grad norm given a gradient scale."""
total_norm = 0.0
for p in params:
if p.grad is not None:
param_norm = (p.grad.detach().data / scale).norm(2)
total_norm += param_norm.item() ** 2
total_norm = total_norm**0.5
return total_norm
def main():
args = parse_args()
send_example_telemetry("run_wav2vec2_pretraining_no_trainer", args)
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
# set up weights and biases if available
if is_wandb_available():
import wandb
wandb.init(project=args.output_dir.split("/")[-1])
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub and not args.preprocessing_only:
# Retrieve of infer repo_name
repo_name = args.hub_model_id
if repo_name is None:
repo_name = Path(args.output_dir).absolute().name
# Create repo and retrieve repo_id
api = HfApi()
repo_id = api.create_repo(repo_name, exist_ok=True, token=args.hub_token).repo_id
with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
if "step_*" not in gitignore:
gitignore.write("step_*\n")
if "epoch_*" not in gitignore:
gitignore.write("epoch_*\n")
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# 1. Download and create train, validation dataset
# We load all dataset configuration and datset split pairs passed in
# ``args.dataset_config_names`` and ``args.dataset_split_names``
###################################################################################################################################################
filepaths = list(Path("data").glob("*.vhdr"))
random.shuffle(filepaths)
val_split_ratio = args.validation_split_percentage / 100.0
num_val_samples = int(val_split_ratio * len(filepaths))
train_files = filepaths[num_val_samples:]
val_files = filepaths[:num_val_samples]
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(args.model_name_or_path) #1
train_dataset = BIDSBrainVisionDataset(
filepaths=train_files,
feature_extractor=feature_extractor,
target_sr=16000,
window_size=2.0
)
val_dataset = BIDSBrainVisionDataset(
filepaths=val_files,
feature_extractor=feature_extractor,
target_sr=16000,
window_size=2.0
)
################################################################################################
# 2. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
# raw_datasets = raw_datasets.cast_column(
# args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
# )
# only normalized-inputs-training is supported
if not feature_extractor.do_normalize:
raise ValueError(
"Training is only supported for normalized inputs. Make sure ``feature_extractor.do_normalize == True``"
)
# set max & min audio length in number of samples
max_length = int(args.max_duration_in_seconds * feature_extractor.sampling_rate) #320000
min_length = int(args.min_duration_in_seconds * feature_extractor.sampling_rate) #32000
cache_file_names = None
if args.train_cache_file_name is not None:
cache_file_names = {"train": args.train_cache_file_name, "validation": args.validation_cache_file_name}
if args.preprocessing_only:
return
# 3. Load model
config = Wav2Vec2Config.from_pretrained(args.model_name_or_path)
# pretraining is only supported for "newer" stable layer norm architecture
# apply_spec_augment has to be True, mask_feature_prob has to be 0.0
if not config.do_stable_layer_norm or config.feat_extract_norm != "layer":
raise ValueError(
"PreTraining is only supported for ``config.do_stable_layer_norm=True`` and"
" ``config.feat_extract_norm='layer'"
)
try:
if os.path.isdir(args.output_dir):
print(f"Loading model from {args.output_dir}...")
model = Wav2Vec2ForPreTraining.from_pretrained(args.output_dir).to(device)
processor = feature_extractor
else:
raise ValueError(f"No model found at {args.output_dir}")
except Exception as e:
model = Wav2Vec2ForPreTraining(config).to(device)
print(f"{e}. Starting new training from scratch.")
# Activate gradient checkpointing if needed
if args.gradient_checkpointing:
model.gradient_checkpointing_enable()
# 4. Define data collator, optimizer and scheduler
mask_time_prob = config.mask_time_prob if args.mask_time_prob is None else args.mask_time_prob
mask_time_length = config.mask_time_length if args.mask_time_length is None else args.mask_time_length
data_collator = DataCollatorForWav2Vec2Pretraining(
model=model,
feature_extractor=feature_extractor,
pad_to_multiple_of=args.pad_to_multiple_of,
mask_time_prob=mask_time_prob,
mask_time_length=mask_time_length,
)
train_dataloader = DataLoader(
train_dataset,
shuffle=False,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size,
pin_memory=True
)
eval_dataloader = DataLoader(
val_dataset,
shuffle=False,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size,
pin_memory=True
)
# Optimizer
optimizer = AdamW(
list(model.parameters()),
lr=args.learning_rate,
betas=[args.adam_beta1, args.adam_beta2],
eps=args.adam_epsilon,
)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader
)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(len(train_dataset) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# 5. Train
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
completed_steps = 0
starting_epoch = 0
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
completed_steps = 0
starting_epoch = 0
model.train()
for epoch in range(starting_epoch, args.num_train_epochs):
# model.train() # here batch["input_values"].shape=40, 32000
for step, batch in enumerate(train_dataloader): # TODO: Check shape of batch
# compute num of losses
num_losses = batch["mask_time_indices"].sum()
sub_attention_mask = batch.pop("sub_attention_mask", None)
sub_attention_mask = (
sub_attention_mask if sub_attention_mask is not None else torch.ones_like(batch["mask_time_indices"])
)
percent_masked = num_losses / sub_attention_mask.sum()
# forward
outputs = model(**batch) # .to(device)
loss = outputs.loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if accelerator.state.num_processes > 1:
num_losses = accelerator.gather_for_metrics(num_losses).sum()
gradient_multiplier = accelerator.state.num_processes / num_losses
multiply_grads(model.module.parameters(), gradient_multiplier)
else:
multiply_grads(model.parameters(), 1 / num_losses)
# update step
if (step + 1) % args.gradient_accumulation_steps == 0:
# compute grad norm for monitoring
scale = (
accelerator.scaler._scale.item()
if hasattr(accelerator, "scaler") and accelerator.scaler is not None
else 1
)
if accelerator.state.num_processes > 1:
grad_norm = get_grad_norm(model.module.parameters(), scale)
else:
grad_norm = get_grad_norm(model.parameters(), scale)
# update parameters
optimizer.step()
optimizer.zero_grad()
if not accelerator.optimizer_step_was_skipped:
lr_scheduler.step()
elif accelerator.is_local_main_process:
progress_bar.write(
f"Gradients have overflown - skipping update step... Updating gradient scale to {scale}..."
)
# update gumbel temperature
gumbel_temperature = max(
args.max_gumbel_temperature * args.gumbel_temperature_decay**completed_steps,
args.min_gumbel_temperature,
)
if hasattr(model, "module"):
model.module.set_gumbel_temperature(gumbel_temperature)
else:
model.set_gumbel_temperature(gumbel_temperature)
progress_bar.update(1)
completed_steps += 1
# 6. Log all results
if (step + 1) % (args.gradient_accumulation_steps * args.logging_steps) == 0:
print("all good")
loss.detach()
outputs.contrastive_loss.detach()
outputs.diversity_loss.detach()
if accelerator.state.num_processes > 1:
loss = accelerator.gather_for_metrics(loss).sum()
outputs.contrastive_loss = accelerator.gather_for_metrics(outputs.contrastive_loss).sum()
outputs.diversity_loss = accelerator.gather_for_metrics(outputs.diversity_loss).sum()
percent_masked = accelerator.gather_for_metrics(percent_masked).sum()
train_logs = {
"loss": (loss * args.gradient_accumulation_steps) / num_losses,
"constrast_loss": outputs.contrastive_loss / num_losses,
"div_loss": outputs.diversity_loss / num_losses,
"%_mask_idx": percent_masked / accelerator.num_processes,
"ppl": outputs.codevector_perplexity,
"lr": torch.tensor(optimizer.param_groups[0]["lr"]),
"temp": torch.tensor(gumbel_temperature),
"grad_norm": torch.tensor(grad_norm),
}
log_str = ""
for k, v in train_logs.items():
log_str += "| {}: {:.3e}".format(k, v.item())
if accelerator.is_local_main_process:
print("still all good")
progress_bar.write(log_str)
if is_wandb_available():
wandb.log(train_logs)
writer.add_scalar("loss/train", float(train_logs["loss"]), step)
writer.add_scalar("div_loss/train", float(train_logs["div_loss"]), step)
writer.add_scalar("learning_rate/train", float(train_logs["lr"].item()), step)
writer.add_scalar("grad_norm/train", float(train_logs["grad_norm"].item()), step)
writer.add_scalar("test_value", 1.0, 0)
writer.flush()
# save model every `args.saving_steps` steps
if (step + 1) % (args.gradient_accumulation_steps * args.saving_steps) == 0:
if (args.push_to_hub and epoch < args.num_train_epochs - 1) or args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
)
if (args.push_to_hub and epoch < args.num_train_epochs - 1) and accelerator.is_main_process:
api.upload_folder(
commit_message=f"Training in progress epoch {epoch}",
folder_path=args.output_dir,
repo_id=repo_id,
repo_type="model",
token=args.hub_token,
)
# if completed steps > `args.max_train_steps` stop
if completed_steps >= args.max_train_steps:
break
# 7. Validate!
model.eval()
# init logs
val_logs = {
"val_loss": 0,
"val_contrastive_loss": 0,
"val_diversity_loss": 0,
"val_num_losses": 0,
}
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
batch.pop("sub_attention_mask", None)
outputs = model(**batch)
val_logs["val_loss"] += outputs.loss
val_logs["val_contrastive_loss"] += outputs.contrastive_loss
val_logs["val_diversity_loss"] += outputs.diversity_loss
val_logs["val_num_losses"] += batch["mask_time_indices"].sum()
# sum over devices in multi-processing
if accelerator.num_processes > 1:
val_logs = {k: accelerator.gather_for_metrics(v).sum() for k, v in val_logs.items()}
val_logs = {k: v / val_logs["val_num_losses"] for k, v in val_logs.items()}
log_str = ""
for k, v in val_logs.items():
log_str += "| {}: {:.3e}".format(k, v.item())
if accelerator.is_local_main_process:
progress_bar.write(log_str)
if is_wandb_available():
wandb.log(val_logs)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
)
if accelerator.is_main_process:
if args.push_to_hub:
api.upload_folder(
commit_message="End of training",
folder_path=args.output_dir,
repo_id=repo_id,
repo_type="model",
token=args.hub_token,
)
if __name__ == "__main__":
main()
#tensorboard --logdir=logs