-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
1311 lines (1105 loc) · 54.6 KB
/
train.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
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# MIT License
#
# Copyright (c) 2023 Christopher Friesen
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import argparse
import logging
import math
import os
import shutil
import subprocess
import atexit
from glob import glob
from typing import Any
from dotenv import load_dotenv
import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import torchaudio
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset, Audio
from tqdm.auto import tqdm
import diffusers
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel
from diffusers.utils import deprecate, is_tensorboard_available
from unet_dual import UNetDualModel
from autoencoder_kl_dual import AutoencoderKLDual
from dual_diffusion_pipeline import DualDiffusionPipeline
from dual_diffusion_utils import init_cuda, compute_snr, load_audio, save_audio, load_raw, save_raw, dict_str, get_mel_density, normalize_lufs
logger = get_logger(__name__, log_level="INFO")
def parse_args():
parser = argparse.ArgumentParser(description="DualDiffusion training script.")
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained / new model",
)
parser.add_argument(
"--module",
type=str,
default="unet",
required=False,
help="Which module in the model to train. Choose between ['unet', 'vae']",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier",
)
parser.add_argument(
"--max_train_samples",
type=int,
default=None,
help=(
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
),
)
parser.add_argument(
"--output_dir",
type=str,
default=None,
help="The output directory where checkpoints will be written. Defaults to pretrained_model_name_or_path.",
)
parser.add_argument(
"--cache_dir",
type=str,
default=None,
help="The directory where the downloaded models and datasets will be stored.",
)
parser.add_argument(
"--seed",
type=int,
default=None,
help="A seed for reproducible training."
)
parser.add_argument(
"--train_batch_size",
type=int,
default=16,
help="Batch size (per device) for the training dataloader."
)
parser.add_argument(
"--vae_encode_batch_size",
type=int,
default=0,
help="If set, use this batch size for VAE encoding when training diffusion UNet module. Defaults to train_batch_size."
)
parser.add_argument("--num_train_epochs", type=int, default=100)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=1000, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--snr_gamma",
type=float,
default=None,
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
"More details here: https://arxiv.org/abs/2303.09556.",
)
parser.add_argument(
"--num_timestep_loss_buckets",
type=int,
default=10,
help=("When training a unet/diffusion module un-weighted loss for groups of timesteps can be logged separately. Set to 0 to disable."),
)
#parser.add_argument(
# "--pitch_augmentation_range",
# type=float,
# default=0, #2/12,
# help="Modulate the pitch of the sample by a random amount within this range (in octaves) - Currently unused",
#)
#parser.add_argument(
# "--tempo_augmentation_range",
# type=float,
# default=0, #0.167,
# help="Modulate the tempo of the sample by a random amount within this range (value of 1 is double/half speed) - Currently unused",
#)
parser.add_argument(
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
)
parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
parser.add_argument("--ema_inv_gamma", type=float, default=1.0, help="The inverse gamma value for the EMA decay.")
parser.add_argument("--ema_power", type=float, default=2 / 3, help="The power value for the EMA decay.")
parser.add_argument("--ema_min_decay", type=float, default=0., help="The minimum decay magnitude for EMA.")
parser.add_argument("--ema_max_decay", type=float, default=0.9999, help="The maximum decay magnitude for EMA.")
parser.add_argument(
"--non_ema_revision",
type=str,
default=None,
required=False,
help=(
"Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or"
" remote repository specified with --pretrained_model_name_or_path."
),
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument(
"--dataset_name",
type=str,
default=None,
required=False,
help=(
"Specify a dataset name to load training data from the huggingface hub. If not specified the training data will be"
" loaded from the path specified in the DATASET_PATH environment variable."
),
)
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
parser.add_argument("--max_grad_norm", default=1., type=float, help="Max gradient norm.")
parser.add_argument(
"--logging_dir",
type=str,
default=None,
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" output_dir/logs_(modulename)/(model_name)"
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--checkpointing_steps",
type=int,
default=None,
help=(
"Save a checkpoint of the training state every X updates. By default, a checkpoint will be saved after every epoch."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=1,
help=("Max number of checkpoints to store."),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default="latest",
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint (default).'
' Use `"none"` to start a new training run even if checkpoints exist in the output directory.'
),
)
parser.add_argument(
"--num_validation_epochs",
type=int,
default=5,
help="Number of epochs between creating new validation samples.",
)
parser.add_argument(
"--num_validation_samples",
type=int,
default=4,
help="Number of samples to generate for validation.",
)
parser.add_argument(
"--num_validation_steps",
type=int,
default=250,
help="Number of steps to use when creating validation samples.",
)
parser.add_argument(
"--validation_sample_dir",
type=str,
default=None,
help="A folder containing samples used only for validation. By default the training data dir is used.",
)
parser.add_argument(
"--tracker_project_name",
type=str,
default=None,
help=(
"The `project_name` argument passed to Accelerator.init_trackers for"
" more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
),
)
parser.add_argument(
"--enable_anomaly_detection",
action="store_true",
help="Enable pytorch anomaly detection - Kills performance but can be used to find the cause of NaN / inf gradients.",
)
args = parser.parse_args()
# validate args / add automatic args
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
args.train_data_dir = os.environ.get("DATASET_PATH", None)
args.train_data_format = os.environ.get("DATASET_FORMAT", None)
args.train_data_num_channels = os.environ.get("DATASET_NUM_CHANNELS", None)
args.train_data_raw_format = os.environ.get("DATASET_RAW_FORMAT", None)
if args.train_data_dir is None:
raise ValueError("DATASET_PATH environment variable is undefined.")
if args.train_data_format is None:
raise ValueError("DATASET_FORMAT environment variable is undefined.")
if args.train_data_num_channels is None:
raise ValueError("DATASET_NUM_CHANNELS environment variable is undefined.")
else:
args.train_data_num_channels = int(args.train_data_num_channels)
if args.train_data_format == ".raw" and args.train_data_raw_format is None:
raise ValueError("DATASET_FORMAT is '.raw' and DATASET_RAW_SAMPLE_FORMAT environment variable is undefined.")
if not os.path.exists(args.train_data_dir) and args.dataset_name is None:
raise ValueError(f"Training data directory {args.train_data_dir} does not exist.")
args.hf_token = os.environ.get("HF_TOKEN", None)
if args.validation_sample_dir is None:
args.validation_sample_dir = args.train_data_dir
else:
if not os.path.exists(args.validation_sample_dir):
raise ValueError(f"Validation sample directory {args.validation_sample_dir} does not exist.")
args.module = args.module.lower().strip()
if args.module not in ["unet", "vae"]:
raise ValueError(f"Unknown module type {args.module}")
# default to using the same revision for the non-ema model if not specified
if args.non_ema_revision is None:
args.non_ema_revision = args.revision
if args.non_ema_revision is not None:
deprecate(
"non_ema_revision!=None",
"0.15.0",
message=(
"Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to"
" use `--variant=non_ema` instead."
),
)
if args.output_dir is None:
args.output_dir = args.pretrained_model_name_or_path
os.makedirs(args.output_dir, exist_ok=True)
if args.logging_dir is None:
args.logging_dir = os.path.join(args.output_dir, f"logs_{args.module}")
os.makedirs(args.logging_dir, exist_ok=True)
if args.tracker_project_name is None:
args.tracker_project_name = os.path.basename(args.output_dir)
return args
def init_logging(accelerator, logging_dir, module_type, report_to):
log_path = os.path.join(logging_dir, f"train_{module_type}.log")
logging.basicConfig(
handlers=[
logging.FileHandler(log_path),
logging.StreamHandler()
],
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(f"Logging to {log_path}")
datasets.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
if accelerator.is_main_process:
if report_to == "tensorboard":
if not is_tensorboard_available():
raise ImportError("Make sure to install tensorboard if you want to use it for logging during training.")
port = int(os.environ.get("TENSORBOARD_HTTP_PORT", 6006))
tensorboard_args = [
"tensorboard",
"--logdir",
logging_dir,
"--bind_all",
"--port",
str(port),
"--samples_per_plugin",
"scalars=2000",
]
tensorboard_monitor_process = subprocess.Popen(tensorboard_args)
def cleanup_process():
try:
tensorboard_monitor_process.terminate()
except Exception:
logger.warn("Failed to terminate tensorboard process")
pass
atexit.register(cleanup_process)
def init_accelerator(project_dir,
grad_accumulation_steps,
mixed_precision,
logging_dir,
log_with,
tracker_project_name):
accelerator_project_config = ProjectConfiguration(project_dir=project_dir, logging_dir=logging_dir)
accelerator = Accelerator(
gradient_accumulation_steps=grad_accumulation_steps,
mixed_precision=mixed_precision,
log_with=log_with,
project_config=accelerator_project_config,
)
if accelerator.is_main_process:
accelerator.init_trackers(tracker_project_name)
return accelerator
def init_accelerator_loadsave_hooks(accelerator, module_type, module_class, ema_module):
def save_model_hook(models, weights, output_dir):
if accelerator.is_main_process:
if ema_module is not None:
ema_module.save_pretrained(os.path.join(output_dir, f"{module_type}_ema"))
for model in models:
model.save_pretrained(os.path.join(output_dir, module_type))
weights.pop() # make sure to pop weight so that corresponding model is not saved again
def load_model_hook(models, input_dir):
if ema_module is not None:
if not os.path.exists(os.path.join(input_dir, f"{module_type}_ema")):
logger.info("EMA model in checkpoint not found, using new ema model")
else:
load_model = EMAModel.from_pretrained(os.path.join(input_dir, f"{module_type}_ema"), module_class)
ema_module.load_state_dict(load_model.state_dict())
ema_module.to(accelerator.device)
del load_model
for _ in range(len(models)):
model = models.pop() # pop models so that they are not loaded again
# load diffusers style into model
load_model = module_class.from_pretrained(input_dir, subfolder=module_type)
model.register_to_config(**load_model.config)
model.load_state_dict(load_model.state_dict())
del load_model
accelerator.register_save_state_pre_hook(save_model_hook)
accelerator.register_load_state_pre_hook(load_model_hook)
logger.info("Registered accelerator hooks for model load/save")
def save_checkpoint(module, module_type, output_dir, global_step, accelerator, checkpoints_total_limit):
module.config["last_global_step"] = global_step
save_path = os.path.join(output_dir, f"{module_type}_checkpoint-{global_step}")
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
# copy all source code / scripts to model folder for posterity
source_src_path = os.path.dirname(__file__)
target_src_path = os.path.join(save_path, "src")
logger.info(f"Copying source code at '{source_src_path}' to checkpoint folder '{target_src_path}'")
try:
os.makedirs(target_src_path, exist_ok=True)
src_file_types = ["py", "cmd", "yml", "sh", "env"]
src_files = []
for file_type in src_file_types:
src_files += glob(f"*.{file_type}") # todo: update this to be case insensitive when upgrading to python 3.12
for src_file in src_files:
shutil.copy(src_file, os.path.join(target_src_path, os.path.basename(src_file)))
except Exception as e:
logger.warning(f"Failed to copy source code to model folder: {e}")
# delete old checkpoints AFTER saving new checkpoint
if checkpoints_total_limit is not None:
try:
checkpoints = os.listdir(output_dir)
checkpoints = [d for d in checkpoints if d.startswith(f"{module_type}_checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
if len(checkpoints) > checkpoints_total_limit:
num_to_remove = len(checkpoints) - checkpoints_total_limit
if num_to_remove > 0:
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
except Exception as e:
logger.error(f"Error removing old checkpoints: {e}")
def load_checkpoint(checkpoint,
output_dir,
module_type,
accelerator,
optimizer,
lr_scheduler,
learning_rate,
adam_weight_decay,
gradient_accumulation_steps,
num_update_steps_per_epoch):
global_step = 0
resume_step = 0
first_epoch = 0
if checkpoint == "none":
return global_step, resume_step, first_epoch
if checkpoint != "latest": # load specific checkpoint
path = os.path.basename(checkpoint)
else: # get most recent checkpoint
dirs = os.listdir(output_dir)
dirs = [d for d in dirs if d.startswith(f"{module_type}_checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
logger.warning(f"Checkpoint '{checkpoint}' does not exist. Starting a new training run.")
else:
global_step = int(path.split("-")[1])
logger.info(f"Resuming from checkpoint {path} (global step: {global_step})")
accelerator.load_state(os.path.join(output_dir, path))
# update learning rate in case we've changed it
updated_learn_rate = False
for g in optimizer.param_groups:
if g["lr"] != learning_rate:
g["lr"] = learning_rate
updated_learn_rate = True
if updated_learn_rate:
lr_scheduler.scheduler.base_lrs = [learning_rate]
logger.info(f"Using updated learning rate: {learning_rate}")
# update weight decay in case we've changed it
updated_weight_decay = False
for g in optimizer.param_groups:
if g["weight_decay"] != adam_weight_decay:
g["weight_decay"] = adam_weight_decay
updated_weight_decay = True
if updated_weight_decay:
logger.info(f"Using updated adam weight decay: {adam_weight_decay}")
if global_step > 0:
resume_global_step = global_step * gradient_accumulation_steps
first_epoch = global_step // num_update_steps_per_epoch
resume_step = resume_global_step % (num_update_steps_per_epoch * gradient_accumulation_steps)
return global_step, resume_step, first_epoch
def init_module_pipeline(pretrained_model_name_or_path, module_type, vae_encode_batch_size, device):
pipeline = DualDiffusionPipeline.from_pretrained(pretrained_model_name_or_path)
module = getattr(pipeline, module_type)
if module_type == "unet":
module_class = UNetDualModel
vae = getattr(pipeline, "vae", None)
if vae is not None:
vae = vae.to(device).half()
if vae_encode_batch_size > 0:
vae.enable_slicing(vae_encode_batch_size)
logger.info(f"Training diffusion model with VAE")
else:
logger.info(f"Training diffusion model without VAE")
elif module_type == "vae":
module_class = AutoencoderKLDual
vae = None
if getattr(pipeline, "unet", None) is not None:
pipeline.unet = pipeline.unet.to("cpu")
else:
raise ValueError(f"Unknown module {module_type}")
pipeline.format = pipeline.format.to(device)
logger.info(f"Training module class: {module_class.__name__}")
return pipeline, module, module_class, vae
def init_ema_module(pretrained_model_name_or_path,
module_type,
module_class,
revision,
ema_min_decay,
ema_max_decay,
ema_inv_gamma,
ema_power,
device):
ema_module = module_class.from_pretrained(
pretrained_model_name_or_path, subfolder=module_type, revision=revision
)
ema_module = EMAModel(ema_module.parameters(),
model_cls=module_class,
model_config=ema_module.config,
min_decay=ema_min_decay,
decay=ema_max_decay,
inv_gamma=ema_inv_gamma,
power=ema_power).to(device)
logger.info(f"Using EMA model with max decay: {ema_max_decay} min decay: {ema_min_decay} inv gamma: {ema_inv_gamma} power: {ema_power}")
return ema_module
def init_optimizer(use_8bit_adam,
learning_rate,
adam_beta1,
adam_beta2,
adam_weight_decay,
adam_epsilon,
module):
if use_8bit_adam:
try:
import bitsandbytes as bnb # type: ignore
except ImportError:
raise ImportError(
"Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
)
optimizer_cls = bnb.optim.AdamW8bit
else:
optimizer_cls = torch.optim.AdamW
optimizer = optimizer_cls(
module.parameters(),
lr=learning_rate,
betas=(adam_beta1, adam_beta2),
weight_decay=adam_weight_decay,
eps=adam_epsilon,
)
logger.info(f"Using optimiser {optimizer_cls.__name__} with learning rate {learning_rate}")
logger.info(f"AdamW beta1: {adam_beta1} beta2: {adam_beta2} eps: {adam_epsilon} weight_decay: {adam_weight_decay}")
return optimizer
def init_lr_scheduler(lr_schedule, optimizer, lr_warmup_steps, max_train_steps, num_processes):
lr_scheduler = get_scheduler(
lr_schedule,
optimizer=optimizer,
num_warmup_steps=lr_warmup_steps * num_processes,
num_training_steps=max_train_steps * num_processes,
)
return lr_scheduler
class DatasetTransformer(torch.nn.Module):
def __init__(self, dataset_transform_params):
self.dataset_format = dataset_transform_params["format"]
self.dataset_raw_format = dataset_transform_params["raw_format"]
self.dataset_num_channels = dataset_transform_params["num_channels"]
self.sample_crop_width = dataset_transform_params["crop_width"]
def __call__(self, examples):
samples = []
paths = []
for audio in examples["audio"]:
input_audio = audio["bytes"] or audio["path"]
if self.dataset_format == ".raw":
sample = load_raw(input_audio,
dtype=self.dataset_raw_format,
num_channels=self.dataset_num_channels,
start=-1,
count=self.sample_crop_width)
else:
sample = load_audio(input_audio,
start=-1,
count=self.sample_crop_width)
samples.append(sample)
paths.append(audio["path"])
return {"input": samples, "sample_paths": paths}
def init_dataloader(accelerator,
dataset_name,
hf_token,
train_data_dir,
cache_dir,
train_batch_size,
dataloader_num_workers,
max_train_samples,
dataset_format,
dataset_raw_format,
dataset_num_channels,
sample_rate,
sample_crop_width,
prefetch_factor=4):
dataset_transform_params = {
"format": dataset_format,
"raw_format": dataset_raw_format,
"num_channels": dataset_num_channels,
"sample_rate": sample_rate,
"crop_width": sample_crop_width,
}
dataset_transform = DatasetTransformer(dataset_transform_params)
if dataset_name is None:
dataset_name = "audiofolder"
data_files = {"train": os.path.join(train_data_dir, "**")}
else:
data_files = f"**{dataset_format}"
dataset = load_dataset(
dataset_name,
data_files=data_files,
cache_dir=cache_dir,
num_proc=dataloader_num_workers if dataloader_num_workers > 0 else None,
token=hf_token,
).cast_column("audio", Audio(decode=False))
with accelerator.main_process_first():
if max_train_samples is not None:
dataset["train"] = dataset["train"].select(range(max_train_samples))
train_dataset = dataset["train"].with_transform(dataset_transform)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
shuffle=True,
batch_size=train_batch_size,
num_workers=dataloader_num_workers,
pin_memory=True,
persistent_workers=True if dataloader_num_workers > 0 else False,
prefetch_factor=prefetch_factor if dataloader_num_workers > 0 else None,
drop_last=True,
)
logger.info(f"Using training data from {train_data_dir} with {len(train_dataset)} samples and batch size {train_batch_size}")
if dataloader_num_workers > 0:
logger.info(f"Using dataloader with {dataloader_num_workers} workers - prefetch factor: {prefetch_factor}")
logger.info(f"Dataset transform params: {dict_str(dataset_transform_params)}")
return train_dataset, train_dataloader
def do_training_loop(args,
accelerator,
module,
ema_module,
pipeline,
vae,
lr_scheduler,
optimizer,
first_epoch,
global_step,
resume_step,
num_update_steps_per_epoch,
max_train_steps,
train_dataloader):
model_params = pipeline.config["model_params"]
sample_shape = pipeline.format.get_sample_shape(bsz=args.train_batch_size)
latent_shape = None
if args.module == "vae":
latent_shape = module.get_latent_shape(sample_shape)
kl_loss_weight = model_params["kl_loss_weight"]
channel_kl_loss_weight = model_params["channel_kl_loss_weight"]
recon_loss_weight = model_params["recon_loss_weight"]
point_loss_weight = model_params["point_loss_weight"]
logger.info("Training VAE model:")
logger.info(f"Loss params: {dict_str(model_params['loss_params'])}")
logger.info(f"Using KL loss weight: {kl_loss_weight} - Channel KL loss weight: {channel_kl_loss_weight}")
logger.info(f"Recon loss weight: {recon_loss_weight} - Point loss weight: {point_loss_weight}")
kl_loss_weight = torch.tensor(kl_loss_weight, device=accelerator.device, dtype=torch.float32)
channel_kl_loss_weight = torch.tensor(channel_kl_loss_weight, device=accelerator.device, dtype=torch.float32)
recon_loss_weight = torch.tensor(recon_loss_weight, device=accelerator.device, dtype=torch.float32)
point_loss_weight = torch.tensor(point_loss_weight, device=accelerator.device, dtype=torch.float32)
module_log_channels = [
"kl_loss_weight",
"channel_kl_loss_weight",
"recon_loss_weight",
"real_loss",
"imag_loss",
"kl_loss",
"channel_kl_loss",
"latents_mean",
"latents_std",
"point_similarity",
"point_loss_weight",
]
elif args.module == "unet":
logger.info("Training UNet model:")
if args.num_timestep_loss_buckets > 0:
logger.info(f"Using {args.num_timestep_loss_buckets} timestep loss buckets")
timestep_loss_buckets = torch.zeros(args.num_timestep_loss_buckets,
device="cpu", dtype=torch.float32)
timestep_loss_bucket_counts = torch.zeros(args.num_timestep_loss_buckets,
device="cpu", dtype=torch.float32)
else:
logger.info("Timestep loss buckets are disabled")
module_log_channels = []
if vae is not None:
if vae.config.last_global_step == 0:
logger.error("VAE model has not been trained, aborting...")
exit(1)
latent_shape = vae.get_latent_shape(sample_shape)
latent_mean = model_params["latent_mean"]
latent_std = model_params["latent_std"]
timestep_ln_center = model_params["timestep_ln_center"]
timestep_ln_scale = model_params["timestep_ln_scale"]
logger.info(f"Sampling timesteps using logistic normal distribution - center: {timestep_ln_center} - scale: {timestep_ln_scale}")
input_perturbation = model_params["input_perturbation"]
if input_perturbation > 0:
logger.info(f"Using input perturbation of {input_perturbation}")
else:
logger.info("Input perturbation is disabled")
logger.info(f"Sample shape: {sample_shape}")
if latent_shape is not None:
logger.info(f"Latent shape: {latent_shape}")
if args.module == "unet":
logger.info(f"Latent mean: {latent_mean} - Latent std: {latent_std}")
for epoch in range(first_epoch, args.num_train_epochs):
torch.cuda.empty_cache()
module.train().requires_grad_(True)
train_loss = 0.
grad_accum_steps = 0
module_logs = {}
for channel in module_log_channels:
module_logs[channel] = 0.
progress_bar = tqdm(total=num_update_steps_per_epoch, disable=not accelerator.is_local_main_process)
progress_bar.set_description(f"Epoch {epoch}")
for step, batch in enumerate(train_dataloader):
# skip steps until we reach the resumed step if resuming from checkpoint - todo: this is inefficient
if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
if step % args.gradient_accumulation_steps == 0:
progress_bar.update(1)
continue
# this breaks reproducible training but is worth it to get more value out of the min-snr weighting with smaller batches
if args.module == "unet" and grad_accum_steps == 0:
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
batch_timesteps = torch.randn(total_batch_size, device=accelerator.device) * timestep_ln_scale + timestep_ln_center
batch_timesteps = torch.sigmoid(batch_timesteps) * 999.
if args.num_timestep_loss_buckets > 0:
timestep_loss_buckets.zero_()
timestep_loss_bucket_counts.zero_()
with accelerator.accumulate(module):
raw_samples = batch["input"]
raw_sample_paths = batch["sample_paths"]
if args.module == "unet":
samples = pipeline.format.raw_to_sample(raw_samples)
if vae is not None:
vae_encoded = vae.encode(samples.half(), return_dict=False)[0]
#loss_weight = (0.01 / vae_encoded.var).clip(min=1e-5).detach().requires_grad_(False)
samples = vae_encoded.mode().float()
samples = (samples - latent_mean) / latent_std
#else:
# loss_weight = 1.
noise = torch.randn_like(samples)
if input_perturbation > 0:
new_noise = noise + input_perturbation * torch.randn_like(noise)
process_batch_timesteps = batch_timesteps[accelerator.local_process_index::accelerator.num_processes]
timesteps = process_batch_timesteps[grad_accum_steps * args.train_batch_size:(grad_accum_steps+1) * args.train_batch_size]
if input_perturbation > 0:
model_input = torch.lerp(samples, new_noise, (timesteps / 999.).view(-1, 1, 1, 1))
else:
model_input = torch.lerp(samples, noise, (timesteps / 999.).view(-1, 1, 1, 1))
model_output = module(model_input, timesteps).sample
target = samples - noise
loss = F.mse_loss(model_output.float(), target.float(), reduction="none")# * loss_weight
timestep_loss = loss.mean(dim=list(range(1, len(loss.shape))))
loss = timestep_loss.mean()
if args.num_timestep_loss_buckets > 0:
all_timesteps = accelerator.gather(timesteps.detach()).cpu()
all_timestep_loss = accelerator.gather(timestep_loss.detach()).cpu()
target_buckets = (all_timesteps / 1000. * timestep_loss_buckets.shape[0]).long()
timestep_loss_buckets.index_add_(0, target_buckets, all_timestep_loss)
timestep_loss_bucket_counts.index_add_(0, target_buckets, torch.ones_like(all_timestep_loss))
elif args.module == "vae":
samples_dict = pipeline.format.raw_to_sample(raw_samples, return_dict=True)
posterior = module.encode(samples_dict["samples"], return_dict=False)[0]
latents = posterior.sample()
latents_mean = latents.mean()
latents_std = latents.std()
model_output = module.decode(latents, return_dict=False)[0]
recon_samples_dict = pipeline.format.sample_to_raw(model_output, return_dict=True, decode=False)
point_similarity = (samples_dict["samples"] - recon_samples_dict["samples"]).abs().mean()
real_loss, imag_loss = pipeline.format.get_loss(recon_samples_dict, samples_dict)
real_nll_loss = (real_loss / module.recon_loss_logvar.exp() + module.recon_loss_logvar) * recon_loss_weight
imag_nll_loss = (imag_loss / module.recon_loss_logvar.exp() + module.recon_loss_logvar) * recon_loss_weight
kl_loss = posterior.kl()
latents_channel_mean = latents.mean(dim=(2,3))
latents_channel_std = latents.std(dim=(2,3)).clip(min=1e-5)
channel_kl_loss = (latents_channel_mean.square() + latents_channel_std - 1 - latents_channel_std.log()).mean()
loss = real_nll_loss + imag_nll_loss + kl_loss * kl_loss_weight + channel_kl_loss * channel_kl_loss_weight + point_similarity * point_loss_weight
else:
raise ValueError(f"Unknown module {args.module}")
# Gather the losses across all processes for logging (if we use distributed training).
grad_accum_steps += 1
avg_loss = accelerator.gather(loss).mean()
train_loss += avg_loss.item()
for channel in module_log_channels:
avg = accelerator.gather(locals()[channel]).mean()
module_logs[channel] += avg.item()
# Backpropagate
accelerator.backward(loss)
if accelerator.sync_gradients: # clip and check for nan/inf grad
grad_norm = accelerator.clip_grad_norm_(module.parameters(), args.max_grad_norm).item()
if math.isinf(grad_norm) or math.isnan(grad_norm):
logger.warning(f"Warning: grad norm is {grad_norm} - step={global_step} loss={loss.item()} debug_last_sample_paths={raw_sample_paths}")
logger.info(f"Batch timesteps: {batch_timesteps}")
if math.isnan(grad_norm):
logger.error(f"Error: grad norm is {grad_norm}, aborting...")
import pdb; pdb.set_trace(); exit(1)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
global_step += 1
if grad_accum_steps >= args.gradient_accumulation_steps: # don't log incomplete batches
logs = {"loss": train_loss / grad_accum_steps,
"lr": lr_scheduler.get_last_lr()[0],
"step": global_step,
"grad_norm": grad_norm}