3 # Any copyright is dedicated to the Public Domain.
4 # https://creativecommons.org/publicdomain/zero/1.0/
6 # Written by Francois Fleuret <francois@fleuret.org>
8 import math, sys, argparse, time, tqdm, os, datetime, warnings
10 import torch, torchvision
12 from torch.nn import functional as F
15 import mygpt, tasks, problems
17 ######################################################################
19 if torch.cuda.is_available():
20 device = torch.device("cuda")
21 torch.backends.cuda.matmul.allow_tf32 = True
23 device = torch.device("cpu")
25 ######################################################################
27 parser = argparse.ArgumentParser(
28 description="An implementation of GPT with cache.",
29 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
36 help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
39 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
41 parser.add_argument("--result_dir", type=str, default=None)
43 parser.add_argument("--seed", type=int, default=0)
45 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
47 ########################################
49 parser.add_argument("--nb_epochs", type=int, default=50)
51 parser.add_argument("--batch_size", type=int, default=None)
53 parser.add_argument("--nb_train_samples", type=int, default=None)
55 parser.add_argument("--nb_test_samples", type=int, default=None)
57 parser.add_argument("--optim", type=str, default="adam")
59 ########################################
61 parser.add_argument("--nb_warmup_iter", type=int, default=100)
63 parser.add_argument("--nb_decay_iter", type=int, default=5000)
65 parser.add_argument("--learning_rate", type=float, default=6e-4)
67 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
71 parser.add_argument("--legacy_lr_schedule", action="store_true", default=False)
73 parser.add_argument("--legacy_learning_rate", type=float, default=1e-4)
75 parser.add_argument("--legacy_min_learning_rate", type=float, default=2e-5)
77 parser.add_argument("--nb_large_lr_epochs", type=float, default=10)
79 ########################################
81 parser.add_argument("--model", type=str, default=None)
83 parser.add_argument("--attention", type=str, default=None)
85 parser.add_argument("--dim_model", type=int, default=None)
87 parser.add_argument("--dim_keys", type=int, default=None)
89 parser.add_argument("--dim_hidden", type=int, default=None)
91 parser.add_argument("--nb_heads", type=int, default=None)
93 parser.add_argument("--nb_lines", type=int, default=None)
95 parser.add_argument("--caterpillar_height", type=int, default=None)
97 parser.add_argument("--rho", type=float, default=0.0)
99 parser.add_argument("--dim_rec_v", type=int, default=None)
101 parser.add_argument("--nb_blocks", type=int, default=None)
103 parser.add_argument("--dropout", type=float, default=0.1)
105 ########################################
107 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
109 parser.add_argument("--no_checkpoint", action="store_true", default=False)
111 parser.add_argument("--overwrite_results", action="store_true", default=False)
113 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
115 ##############################
118 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
120 parser.add_argument("--rpl_max_input", type=int, default=9)
122 parser.add_argument("--rpl_prog_len", type=int, default=8)
124 parser.add_argument("--rpl_nb_runs", type=int, default=5)
126 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
128 ##############################
131 parser.add_argument("--grid_size", type=int, default=6)
133 ##############################
136 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
138 parser.add_argument("--picoclvr_height", type=int, default=12)
140 parser.add_argument("--picoclvr_width", type=int, default=16)
142 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
144 ##############################
147 parser.add_argument("--maze_height", type=int, default=13)
149 parser.add_argument("--maze_width", type=int, default=21)
151 parser.add_argument("--maze_nb_walls", type=int, default=15)
153 ##############################
156 parser.add_argument("--snake_height", type=int, default=9)
158 parser.add_argument("--snake_width", type=int, default=12)
160 parser.add_argument("--snake_nb_colors", type=int, default=5)
162 parser.add_argument("--snake_length", type=int, default=200)
164 ##############################
167 parser.add_argument("--stack_nb_steps", type=int, default=100)
169 parser.add_argument("--stack_nb_stacks", type=int, default=3)
171 parser.add_argument("--stack_nb_digits", type=int, default=3)
173 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
175 ##############################
178 parser.add_argument("--expr_nb_variables", type=int, default=5)
180 parser.add_argument("--expr_sequence_length", type=int, default=40)
182 parser.add_argument("--expr_operand_max", type=int, default=9)
184 parser.add_argument("--expr_result_max", type=int, default=99)
186 parser.add_argument("--expr_input_file", type=str, default=None)
188 ##############################
191 parser.add_argument("--memory_len_total", type=int, default=32)
193 ##############################
196 parser.add_argument("--mixing_hard", action="store_true", default=False)
198 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
200 ######################################################################
202 args = parser.parse_args()
204 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
206 if args.result_dir is None:
207 args.result_dir = f"results_{args.task}_{args.model}"
209 ######################################################################
211 default_task_args = {
215 "nb_train_samples": 250000,
216 "nb_test_samples": 10000,
221 "nb_train_samples": 50000,
222 "nb_test_samples": 10000,
227 "nb_train_samples": 2500000,
228 "nb_test_samples": 10000,
233 "nb_train_samples": 250000,
234 "nb_test_samples": 10000,
239 "nb_train_samples": 100000,
240 "nb_test_samples": 1000,
245 "nb_train_samples": 1000000,
246 "nb_test_samples": 10000,
251 "nb_train_samples": 50000,
252 "nb_test_samples": 10000,
257 "nb_train_samples": 100000,
258 "nb_test_samples": 10000,
263 "nb_train_samples": 250000,
264 "nb_test_samples": 10000,
269 "nb_train_samples": 2500000,
270 "nb_test_samples": 10000,
275 "nb_train_samples": 250000,
276 "nb_test_samples": 10000,
281 "nb_train_samples": 100000,
282 "nb_test_samples": 1000,
287 "nb_train_samples": 50000,
288 "nb_test_samples": 10000,
293 "nb_train_samples": 25000,
294 "nb_test_samples": 10000,
299 "nb_train_samples": 250000,
300 "nb_test_samples": 10000,
305 "nb_train_samples": 60000,
306 "nb_test_samples": 10000,
310 if args.task in default_task_args:
311 for k, v in default_task_args[args.task].items():
312 if getattr(args, k) is None:
315 ######################################################################
317 default_model_args = {
328 "attention": "caterpillar",
334 "caterpillar_height": 4,
348 "attention": "caterpillar",
354 "caterpillar_height": 4,
355 "dim_rec_v": 64, # dim_model / nb_heads
368 "attention": "caterpillar",
374 "caterpillar_height": 32,
388 "attention": "caterpillar",
407 "attention": "caterpillar",
418 if args.model in default_model_args:
419 for k, v in default_model_args[args.model].items():
420 if getattr(args, k) is None:
423 raise ValueError(f"Unknown model {args.model}")
425 ######################################################################
428 os.mkdir(args.result_dir)
429 except FileExistsError:
430 if not args.overwrite_results:
431 print(f"result directory {args.result_dir} already exists")
434 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
437 # torch.backends.cudnn.deterministic = True
438 # torch.backends.cudnn.benchmark = False
439 # torch.use_deterministic_algorithms(True)
440 torch.manual_seed(args.seed)
441 if torch.cuda.is_available():
442 torch.cuda.manual_seed_all(args.seed)
444 ######################################################################
448 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
450 if log_file is not None:
451 log_file.write(t + s + "\n")
458 with os.popen("sha256sum *.py") as f:
460 log_string(f"sha256sum {l.strip()}")
462 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
463 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
465 log_string(f"argv {' '.join(sys.argv)}")
468 log_string(f"args.{n} {getattr(args, n)}")
471 ######################################################################
474 def get_lr(n_epoch, it):
475 if args.legacy_lr_schedule:
476 # my crude scheduling to compare to previous baseline, added
479 if it < args.nb_warmup_iter:
480 return args.legacy_learning_rate * it / args.nb_warmup_iter
481 elif it < args.nb_large_lr_epochs:
482 return args.legacy_learning_rate
484 return args.legacy_min_learning_rate
488 # 1) linear warmup for warmup_iter steps
489 if it < args.nb_warmup_iter:
490 return args.learning_rate * it / args.nb_warmup_iter
491 # 2) if it > nb_decay_iter, return min learning rate
492 if it > args.nb_decay_iter:
493 return args.min_learning_rate
494 # 3) in between, use cosine decay down to min learning rate
495 decay_ratio = (it - args.nb_warmup_iter) / (
496 args.nb_decay_iter - args.nb_warmup_iter
498 coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
499 return args.min_learning_rate + coeff * (
500 args.learning_rate - args.min_learning_rate
504 ######################################################################
507 def picoclvr_pruner_horizontal_green(p):
508 return not ("green" in p and ("left" in p or "right" in p))
511 picoclvr_pruner_train = (
512 picoclvr_pruner_horizontal_green
513 if args.picocvlr_prune_properties in {"train+eval"}
517 picoclvr_pruner_eval = (
518 (lambda p: not picoclvr_pruner_horizontal_green(p))
519 if args.picocvlr_prune_properties in {"train+eval", "eval"}
523 ######################################################################
527 if args.task == "byheart":
528 task = tasks.SandBox(
529 problem=problems.ProblemByHeart(),
530 nb_train_samples=args.nb_train_samples,
531 nb_test_samples=args.nb_test_samples,
532 batch_size=args.batch_size,
536 args.max_percents_of_test_in_train = -1
538 elif args.task == "learnop":
539 task = tasks.SandBox(
540 problem=problems.ProblemLearnOperator(),
541 nb_train_samples=args.nb_train_samples,
542 nb_test_samples=args.nb_test_samples,
543 batch_size=args.batch_size,
549 elif args.task == "guessop":
550 task = tasks.SandBox(
551 problem=problems.ProblemGuessOperator(),
552 nb_train_samples=args.nb_train_samples,
553 nb_test_samples=args.nb_test_samples,
554 batch_size=args.batch_size,
560 elif args.task == "twotargets":
561 task = tasks.SandBox(
562 problem=problems.ProblemTwoTargets(),
563 nb_train_samples=args.nb_train_samples,
564 nb_test_samples=args.nb_test_samples,
565 batch_size=args.batch_size,
570 elif args.task == "memory":
571 task = tasks.SandBox(
572 problem=problems.ProblemMemory(len_total=args.memory_len_total),
573 nb_train_samples=args.nb_train_samples,
574 nb_test_samples=args.nb_test_samples,
575 batch_size=args.batch_size,
580 elif args.task == "mixing":
581 task = tasks.SandBox(
582 problem=problems.ProblemMixing(
583 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
585 nb_train_samples=args.nb_train_samples,
586 nb_test_samples=args.nb_test_samples,
587 batch_size=args.batch_size,
592 elif args.task == "addition":
593 task = tasks.SandBox(
594 problem=problems.ProblemAddition(),
595 nb_train_samples=args.nb_train_samples,
596 nb_test_samples=args.nb_test_samples,
597 batch_size=args.batch_size,
602 elif args.task == "picoclvr":
603 task = tasks.PicoCLVR(
604 nb_train_samples=args.nb_train_samples,
605 nb_test_samples=args.nb_test_samples,
606 batch_size=args.batch_size,
607 height=args.picoclvr_height,
608 width=args.picoclvr_width,
609 nb_colors=args.picoclvr_nb_colors,
612 pruner_train=picoclvr_pruner_train,
613 pruner_eval=picoclvr_pruner_eval,
616 elif args.task == "mnist":
618 nb_train_samples=args.nb_train_samples,
619 nb_test_samples=args.nb_test_samples,
620 batch_size=args.batch_size,
624 elif args.task == "maze":
626 nb_train_samples=args.nb_train_samples,
627 nb_test_samples=args.nb_test_samples,
628 batch_size=args.batch_size,
629 height=args.maze_height,
630 width=args.maze_width,
631 nb_walls=args.maze_nb_walls,
635 elif args.task == "snake":
637 nb_train_samples=args.nb_train_samples,
638 nb_test_samples=args.nb_test_samples,
639 batch_size=args.batch_size,
640 height=args.snake_height,
641 width=args.snake_width,
642 nb_colors=args.snake_nb_colors,
643 length=args.snake_length,
644 prompt_length=args.snake_length // 2,
648 elif args.task == "stack":
650 nb_train_samples=args.nb_train_samples,
651 nb_test_samples=args.nb_test_samples,
652 batch_size=args.batch_size,
654 nb_steps=args.stack_nb_steps,
655 nb_stacks=args.stack_nb_stacks,
656 nb_digits=args.stack_nb_digits,
657 fraction_values_for_train=args.stack_fraction_values_for_train,
661 elif args.task == "expr":
663 nb_train_samples=args.nb_train_samples,
664 nb_test_samples=args.nb_test_samples,
665 nb_variables=args.expr_nb_variables,
666 sequence_length=args.expr_sequence_length,
667 operand_max=args.expr_operand_max,
668 result_max=args.expr_result_max,
669 batch_size=args.batch_size,
673 elif args.task == "rpl":
675 nb_train_samples=args.nb_train_samples,
676 nb_test_samples=args.nb_test_samples,
677 batch_size=args.batch_size,
678 nb_starting_values=args.rpl_nb_starting_values,
679 max_input=args.rpl_max_input,
680 prog_len=args.rpl_prog_len,
681 nb_runs=args.rpl_nb_runs,
682 no_prog=args.rpl_no_prog,
687 elif args.task == "grid":
689 nb_train_samples=args.nb_train_samples,
690 nb_test_samples=args.nb_test_samples,
691 batch_size=args.batch_size,
697 elif args.task == "qmlp":
699 nb_train_samples=args.nb_train_samples,
700 nb_test_samples=args.nb_test_samples,
701 batch_size=args.batch_size,
702 result_dir=args.result_dir,
708 raise ValueError(f"Unknown task {args.task}")
710 ######################################################################
712 log_string(f"device {device}")
714 vocabulary_size = task.vocabulary_size()
716 log_string(f"vocabulary_size {vocabulary_size}")
718 ##############################
721 vocabulary_size=vocabulary_size,
722 dim_model=args.dim_model,
723 dim_keys=args.dim_keys,
724 dim_hidden=args.dim_hidden,
725 nb_heads=args.nb_heads,
726 nb_lines=args.nb_lines,
727 caterpillar_height=args.caterpillar_height,
728 dim_rec_v=args.dim_rec_v,
729 nb_blocks=args.nb_blocks,
731 dropout=args.dropout,
732 attention_layer=args.attention,
737 nb_parameters = sum(p.numel() for p in model.parameters())
738 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
740 ######################################################################
742 nb_epochs_finished = 0
744 if args.no_checkpoint:
745 log_string(f"not trying to load checkpoint.")
749 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
750 checkpoint = torch.load(checkpoint_name)
751 nb_epochs_finished = checkpoint["nb_epochs_finished"]
752 model.load_state_dict(checkpoint["model_state"])
753 torch.set_rng_state(checkpoint["rng_state"])
754 if torch.cuda.is_available():
755 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
757 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
759 except FileNotFoundError:
760 log_string("starting from scratch.")
763 log_string("error when loading the checkpoint.")
766 ######################################################################
768 if args.task == "expr" and args.expr_input_file is not None:
769 task.produce_results(
770 n_epoch=nb_epochs_finished,
772 result_dir=args.result_dir,
774 deterministic_synthesis=args.deterministic_synthesis,
775 input_file=args.expr_input_file,
780 ######################################################################
782 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
784 # Compute the entropy of the training tokens
787 for input in task.batches(split="train"):
788 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
789 token_probas = token_count / token_count.sum()
790 entropy = -torch.xlogy(token_probas, token_probas).sum()
791 train_set_perplexity = math.exp(entropy)
793 ######################################################################
794 # A bit of paranoia never hurts
796 if args.max_percents_of_test_in_train >= 0:
798 def subsets_as_tuples(batches, cs):
800 for batch in batches:
802 s.add(tuple([v.item() for v in x]))
808 nb_test, nb_in_train = 0, 0
809 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
811 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
812 in_train.update(test_subset.intersection(train_subset))
813 nb_in_train += len(in_train)
814 nb_test += len(test_subset)
817 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
821 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
822 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
824 ##############################
828 if nb_epochs_finished >= nb_epochs:
829 task.produce_results(
830 n_epoch=nb_epochs_finished,
832 result_dir=args.result_dir,
834 deterministic_synthesis=args.deterministic_synthesis,
837 time_pred_result = None
841 for n_epoch in range(nb_epochs_finished, nb_epochs):
842 if args.optim == "sgd":
843 optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
844 elif args.optim == "adam":
845 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
846 elif args.optim == "adamw":
847 optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
849 raise ValueError(f"Unknown optimizer {args.optim}.")
853 nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
855 for input in task.batches(split="train"):
856 model.reset_inner_loss()
857 input = input.to(device)
859 output = model(mygpt.BracketedSequence(input)).x
860 loss = F.cross_entropy(output.transpose(1, 2), input)
861 inner_loss = model.get_inner_loss()
863 acc_train_loss += loss.item() * input.size(0)
864 acc_train_inner_loss += inner_loss.item() * input.size(0)
866 nb_train_samples += input.size(0)
867 nb_samples_seen += input.size(0)
869 total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
872 lr = get_lr(n_epoch, it)
873 for param_group in optimizer.param_groups:
874 param_group["lr"] = lr
876 # log_string(f"learning_rate {lr}")
878 optimizer.zero_grad()
879 total_loss.backward()
882 with torch.autograd.no_grad():
885 nb_test_samples, acc_test_loss = 0, 0.0
887 for input in task.batches(split="test"):
888 input = input.to(device)
890 output = model(mygpt.BracketedSequence(input)).x
891 loss = F.cross_entropy(output.transpose(1, 2), input)
892 acc_test_loss += loss.item() * input.size(0)
893 nb_test_samples += input.size(0)
896 f"loss {n_epoch} train_loss {acc_train_loss/nb_train_samples} train_inner_loss {acc_train_inner_loss/nb_train_samples} test_prediction {acc_test_loss/nb_test_samples}"
899 task.produce_results(
902 result_dir=args.result_dir,
904 deterministic_synthesis=args.deterministic_synthesis,
907 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
908 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
911 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
914 time_current_result = datetime.datetime.now()
915 if time_pred_result is not None:
917 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
919 time_pred_result = time_current_result
922 "nb_epochs_finished": n_epoch + 1,
923 "model_state": model.state_dict(),
924 "rng_state": torch.get_rng_state(),
927 if torch.cuda.is_available():
928 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
930 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
931 torch.save(checkpoint, checkpoint_name)
932 log_string(f"saved checkpoint {checkpoint_name}")
934 ######################################################################