diff options
author | Volpeon <git@volpeon.ink> | 2023-04-03 12:39:17 +0200 |
---|---|---|
committer | Volpeon <git@volpeon.ink> | 2023-04-03 12:39:17 +0200 |
commit | a08cd4b1581ca195f619e8bdb6cb6448287d4d2f (patch) | |
tree | 49c753924b44f8b82403ee52ed21f3ab60c76748 | |
parent | Fix memory leak (diff) | |
download | textual-inversion-diff-a08cd4b1581ca195f619e8bdb6cb6448287d4d2f.tar.gz textual-inversion-diff-a08cd4b1581ca195f619e8bdb6cb6448287d4d2f.tar.bz2 textual-inversion-diff-a08cd4b1581ca195f619e8bdb6cb6448287d4d2f.zip |
Bring back Lion optimizer
-rw-r--r-- | environment.yaml | 1 | ||||
-rw-r--r-- | train_dreambooth.py | 30 | ||||
-rw-r--r-- | train_lora.py | 30 | ||||
-rw-r--r-- | train_ti.py | 30 |
4 files changed, 82 insertions, 9 deletions
diff --git a/environment.yaml b/environment.yaml index 8868532..1de76bd 100644 --- a/environment.yaml +++ b/environment.yaml | |||
@@ -18,6 +18,7 @@ dependencies: | |||
18 | - -e git+https://github.com/facebookresearch/dadaptation#egg=dadaptation | 18 | - -e git+https://github.com/facebookresearch/dadaptation#egg=dadaptation |
19 | - accelerate==0.17.1 | 19 | - accelerate==0.17.1 |
20 | - bitsandbytes==0.37.2 | 20 | - bitsandbytes==0.37.2 |
21 | - lion-pytorch==0.0.7 | ||
21 | - peft==0.2.0 | 22 | - peft==0.2.0 |
22 | - python-slugify>=6.1.2 | 23 | - python-slugify>=6.1.2 |
23 | - safetensors==0.3.0 | 24 | - safetensors==0.3.0 |
diff --git a/train_dreambooth.py b/train_dreambooth.py index 48b7926..be7d6fe 100644 --- a/train_dreambooth.py +++ b/train_dreambooth.py | |||
@@ -306,7 +306,7 @@ def parse_args(): | |||
306 | "--optimizer", | 306 | "--optimizer", |
307 | type=str, | 307 | type=str, |
308 | default="dadan", | 308 | default="dadan", |
309 | help='Optimizer to use ["adam", "adam8bit", "dadam", "dadan", "adafactor"]' | 309 | help='Optimizer to use ["adam", "adam8bit", "lion", "dadam", "dadan", "adafactor"]' |
310 | ) | 310 | ) |
311 | parser.add_argument( | 311 | parser.add_argument( |
312 | "--dadaptation_d0", | 312 | "--dadaptation_d0", |
@@ -317,13 +317,13 @@ def parse_args(): | |||
317 | parser.add_argument( | 317 | parser.add_argument( |
318 | "--adam_beta1", | 318 | "--adam_beta1", |
319 | type=float, | 319 | type=float, |
320 | default=0.9, | 320 | default=None, |
321 | help="The beta1 parameter for the Adam optimizer." | 321 | help="The beta1 parameter for the Adam optimizer." |
322 | ) | 322 | ) |
323 | parser.add_argument( | 323 | parser.add_argument( |
324 | "--adam_beta2", | 324 | "--adam_beta2", |
325 | type=float, | 325 | type=float, |
326 | default=0.999, | 326 | default=None, |
327 | help="The beta2 parameter for the Adam optimizer." | 327 | help="The beta2 parameter for the Adam optimizer." |
328 | ) | 328 | ) |
329 | parser.add_argument( | 329 | parser.add_argument( |
@@ -450,6 +450,18 @@ def parse_args(): | |||
450 | if args.output_dir is None: | 450 | if args.output_dir is None: |
451 | raise ValueError("You must specify --output_dir") | 451 | raise ValueError("You must specify --output_dir") |
452 | 452 | ||
453 | if args.adam_beta1 is None: | ||
454 | if args.optimizer in ('adam', 'adam8bit'): | ||
455 | args.adam_beta1 = 0.9 | ||
456 | elif args.optimizer == 'lion': | ||
457 | args.adam_beta1 = 0.95 | ||
458 | |||
459 | if args.adam_beta2 is None: | ||
460 | if args.optimizer in ('adam', 'adam8bit'): | ||
461 | args.adam_beta2 = 0.999 | ||
462 | elif args.optimizer == 'lion': | ||
463 | args.adam_beta2 = 0.98 | ||
464 | |||
453 | return args | 465 | return args |
454 | 466 | ||
455 | 467 | ||
@@ -536,6 +548,18 @@ def main(): | |||
536 | eps=args.adam_epsilon, | 548 | eps=args.adam_epsilon, |
537 | amsgrad=args.adam_amsgrad, | 549 | amsgrad=args.adam_amsgrad, |
538 | ) | 550 | ) |
551 | elif args.optimizer == 'lion': | ||
552 | try: | ||
553 | import lion_pytorch | ||
554 | except ImportError: | ||
555 | raise ImportError("To use Lion, please install the lion_pytorch library: `pip install lion-pytorch`.") | ||
556 | |||
557 | create_optimizer = partial( | ||
558 | lion_pytorch.Lion, | ||
559 | betas=(args.adam_beta1, args.adam_beta2), | ||
560 | weight_decay=args.adam_weight_decay, | ||
561 | use_triton=True, | ||
562 | ) | ||
539 | elif args.optimizer == 'adafactor': | 563 | elif args.optimizer == 'adafactor': |
540 | create_optimizer = partial( | 564 | create_optimizer = partial( |
541 | transformers.optimization.Adafactor, | 565 | transformers.optimization.Adafactor, |
diff --git a/train_lora.py b/train_lora.py index cf73645..a0cd174 100644 --- a/train_lora.py +++ b/train_lora.py | |||
@@ -318,7 +318,7 @@ def parse_args(): | |||
318 | "--optimizer", | 318 | "--optimizer", |
319 | type=str, | 319 | type=str, |
320 | default="dadan", | 320 | default="dadan", |
321 | help='Optimizer to use ["adam", "adam8bit", "dadam", "dadan", "adafactor"]' | 321 | help='Optimizer to use ["adam", "adam8bit", "lion", "dadam", "dadan", "adafactor"]' |
322 | ) | 322 | ) |
323 | parser.add_argument( | 323 | parser.add_argument( |
324 | "--dadaptation_d0", | 324 | "--dadaptation_d0", |
@@ -329,13 +329,13 @@ def parse_args(): | |||
329 | parser.add_argument( | 329 | parser.add_argument( |
330 | "--adam_beta1", | 330 | "--adam_beta1", |
331 | type=float, | 331 | type=float, |
332 | default=0.9, | 332 | default=None, |
333 | help="The beta1 parameter for the Adam optimizer." | 333 | help="The beta1 parameter for the Adam optimizer." |
334 | ) | 334 | ) |
335 | parser.add_argument( | 335 | parser.add_argument( |
336 | "--adam_beta2", | 336 | "--adam_beta2", |
337 | type=float, | 337 | type=float, |
338 | default=0.999, | 338 | default=None, |
339 | help="The beta2 parameter for the Adam optimizer." | 339 | help="The beta2 parameter for the Adam optimizer." |
340 | ) | 340 | ) |
341 | parser.add_argument( | 341 | parser.add_argument( |
@@ -468,6 +468,18 @@ def parse_args(): | |||
468 | if args.output_dir is None: | 468 | if args.output_dir is None: |
469 | raise ValueError("You must specify --output_dir") | 469 | raise ValueError("You must specify --output_dir") |
470 | 470 | ||
471 | if args.adam_beta1 is None: | ||
472 | if args.optimizer in ('adam', 'adam8bit'): | ||
473 | args.adam_beta1 = 0.9 | ||
474 | elif args.optimizer == 'lion': | ||
475 | args.adam_beta1 = 0.95 | ||
476 | |||
477 | if args.adam_beta2 is None: | ||
478 | if args.optimizer in ('adam', 'adam8bit'): | ||
479 | args.adam_beta2 = 0.999 | ||
480 | elif args.optimizer == 'lion': | ||
481 | args.adam_beta2 = 0.98 | ||
482 | |||
471 | return args | 483 | return args |
472 | 484 | ||
473 | 485 | ||
@@ -568,6 +580,18 @@ def main(): | |||
568 | eps=args.adam_epsilon, | 580 | eps=args.adam_epsilon, |
569 | amsgrad=args.adam_amsgrad, | 581 | amsgrad=args.adam_amsgrad, |
570 | ) | 582 | ) |
583 | elif args.optimizer == 'lion': | ||
584 | try: | ||
585 | import lion_pytorch | ||
586 | except ImportError: | ||
587 | raise ImportError("To use Lion, please install the lion_pytorch library: `pip install lion-pytorch`.") | ||
588 | |||
589 | create_optimizer = partial( | ||
590 | lion_pytorch.Lion, | ||
591 | betas=(args.adam_beta1, args.adam_beta2), | ||
592 | weight_decay=args.adam_weight_decay, | ||
593 | use_triton=True, | ||
594 | ) | ||
571 | elif args.optimizer == 'adafactor': | 595 | elif args.optimizer == 'adafactor': |
572 | create_optimizer = partial( | 596 | create_optimizer = partial( |
573 | transformers.optimization.Adafactor, | 597 | transformers.optimization.Adafactor, |
diff --git a/train_ti.py b/train_ti.py index 651dfbe..c242625 100644 --- a/train_ti.py +++ b/train_ti.py | |||
@@ -330,7 +330,7 @@ def parse_args(): | |||
330 | "--optimizer", | 330 | "--optimizer", |
331 | type=str, | 331 | type=str, |
332 | default="dadan", | 332 | default="dadan", |
333 | help='Optimizer to use ["adam", "adam8bit", "dadam", "dadan", "adafactor"]' | 333 | help='Optimizer to use ["adam", "adam8bit", "lion", "dadam", "dadan", "adafactor"]' |
334 | ) | 334 | ) |
335 | parser.add_argument( | 335 | parser.add_argument( |
336 | "--dadaptation_d0", | 336 | "--dadaptation_d0", |
@@ -341,13 +341,13 @@ def parse_args(): | |||
341 | parser.add_argument( | 341 | parser.add_argument( |
342 | "--adam_beta1", | 342 | "--adam_beta1", |
343 | type=float, | 343 | type=float, |
344 | default=0.9, | 344 | default=None, |
345 | help="The beta1 parameter for the Adam optimizer." | 345 | help="The beta1 parameter for the Adam optimizer." |
346 | ) | 346 | ) |
347 | parser.add_argument( | 347 | parser.add_argument( |
348 | "--adam_beta2", | 348 | "--adam_beta2", |
349 | type=float, | 349 | type=float, |
350 | default=0.999, | 350 | default=None, |
351 | help="The beta2 parameter for the Adam optimizer." | 351 | help="The beta2 parameter for the Adam optimizer." |
352 | ) | 352 | ) |
353 | parser.add_argument( | 353 | parser.add_argument( |
@@ -566,6 +566,18 @@ def parse_args(): | |||
566 | if args.output_dir is None: | 566 | if args.output_dir is None: |
567 | raise ValueError("You must specify --output_dir") | 567 | raise ValueError("You must specify --output_dir") |
568 | 568 | ||
569 | if args.adam_beta1 is None: | ||
570 | if args.optimizer in ('adam', 'adam8bit'): | ||
571 | args.adam_beta1 = 0.9 | ||
572 | elif args.optimizer == 'lion': | ||
573 | args.adam_beta1 = 0.95 | ||
574 | |||
575 | if args.adam_beta2 is None: | ||
576 | if args.optimizer in ('adam', 'adam8bit'): | ||
577 | args.adam_beta2 = 0.999 | ||
578 | elif args.optimizer == 'lion': | ||
579 | args.adam_beta2 = 0.98 | ||
580 | |||
569 | return args | 581 | return args |
570 | 582 | ||
571 | 583 | ||
@@ -666,6 +678,18 @@ def main(): | |||
666 | eps=args.adam_epsilon, | 678 | eps=args.adam_epsilon, |
667 | amsgrad=args.adam_amsgrad, | 679 | amsgrad=args.adam_amsgrad, |
668 | ) | 680 | ) |
681 | elif args.optimizer == 'lion': | ||
682 | try: | ||
683 | import lion_pytorch | ||
684 | except ImportError: | ||
685 | raise ImportError("To use Lion, please install the lion_pytorch library: `pip install lion-pytorch`.") | ||
686 | |||
687 | create_optimizer = partial( | ||
688 | lion_pytorch.Lion, | ||
689 | betas=(args.adam_beta1, args.adam_beta2), | ||
690 | weight_decay=args.adam_weight_decay, | ||
691 | use_triton=True, | ||
692 | ) | ||
669 | elif args.optimizer == 'adafactor': | 693 | elif args.optimizer == 'adafactor': |
670 | create_optimizer = partial( | 694 | create_optimizer = partial( |
671 | transformers.optimization.Adafactor, | 695 | transformers.optimization.Adafactor, |