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 ######################################################################
22 if x in {"1", "true", "yes"}:
24 elif x in {"0", "false", "no"}:
30 parser = argparse.ArgumentParser(
31 description="An implementation of GPT with cache.",
32 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
39 help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
42 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
44 parser.add_argument("--result_dir", type=str, default=None)
46 parser.add_argument("--seed", type=int, default=0)
48 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
50 parser.add_argument("--force_cpu", type=str2bool, default=False)
52 ########################################
54 parser.add_argument("--nb_epochs", type=int, default=50)
56 parser.add_argument("--batch_size", type=int, default=None)
58 parser.add_argument("--nb_train_samples", type=int, default=None)
60 parser.add_argument("--nb_test_samples", type=int, default=None)
62 parser.add_argument("--optim", type=str, default="adam")
64 ########################################
66 parser.add_argument("--nb_warmup_iter", type=int, default=100)
68 parser.add_argument("--nb_decay_iter", type=int, default=5000)
70 parser.add_argument("--learning_rate", type=float, default=6e-4)
72 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
76 parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
78 parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
80 parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
82 parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
84 ########################################
86 parser.add_argument("--model", type=str, default=None)
88 parser.add_argument("--attention", type=str, default=None)
90 parser.add_argument("--dim_model", type=int, default=None)
92 parser.add_argument("--dim_keys", type=int, default=None)
94 parser.add_argument("--dim_hidden", type=int, default=None)
96 parser.add_argument("--nb_heads", type=int, default=None)
98 parser.add_argument("--nb_lines", type=int, default=None)
100 parser.add_argument("--caterpillar_height", type=int, default=None)
102 parser.add_argument("--rho", type=float, default=0.0)
104 parser.add_argument("--nb_blocks", type=int, default=None)
106 parser.add_argument("--dropout", type=float, default=0.1)
108 ########################################
110 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
112 parser.add_argument("--no_checkpoint", action="store_true", default=False)
114 parser.add_argument("--continue_training", action="store_true", default=False)
116 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
118 ##############################
121 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
123 parser.add_argument("--rpl_max_input", type=int, default=9)
125 parser.add_argument("--rpl_prog_len", type=int, default=8)
127 parser.add_argument("--rpl_nb_runs", type=int, default=5)
129 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
131 ##############################
134 parser.add_argument("--grid_size", type=int, default=6)
136 ##############################
139 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
141 parser.add_argument("--picoclvr_height", type=int, default=12)
143 parser.add_argument("--picoclvr_width", type=int, default=16)
145 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
147 ##############################
150 parser.add_argument("--maze_height", type=int, default=13)
152 parser.add_argument("--maze_width", type=int, default=21)
154 parser.add_argument("--maze_nb_walls", type=int, default=15)
156 ##############################
159 parser.add_argument("--snake_height", type=int, default=9)
161 parser.add_argument("--snake_width", type=int, default=12)
163 parser.add_argument("--snake_nb_colors", type=int, default=5)
165 parser.add_argument("--snake_length", type=int, default=200)
167 ##############################
170 parser.add_argument("--stack_nb_steps", type=int, default=100)
172 parser.add_argument("--stack_nb_stacks", type=int, default=3)
174 parser.add_argument("--stack_nb_digits", type=int, default=3)
176 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
178 ##############################
181 parser.add_argument("--expr_nb_variables", type=int, default=5)
183 parser.add_argument("--expr_sequence_length", type=int, default=40)
185 parser.add_argument("--expr_operand_max", type=int, default=9)
187 parser.add_argument("--expr_result_max", type=int, default=99)
189 parser.add_argument("--expr_input_file", type=str, default=None)
191 ##############################
194 parser.add_argument("--memory_len_total", type=int, default=32)
196 ##############################
199 parser.add_argument("--mixing_hard", action="store_true", default=False)
201 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
203 ######################################################################
205 args = parser.parse_args()
207 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
209 if args.result_dir is None:
210 args.result_dir = f"results_{args.task}_{args.model}"
212 ######################################################################
214 if not args.force_cpu and torch.cuda.is_available():
215 device = torch.device("cuda")
216 torch.backends.cuda.matmul.allow_tf32 = True
218 device = torch.device("cpu")
220 ######################################################################
222 default_task_args = {
226 "nb_train_samples": 250000,
227 "nb_test_samples": 10000,
232 "nb_train_samples": 50000,
233 "nb_test_samples": 10000,
238 "nb_train_samples": 2500000,
239 "nb_test_samples": 10000,
244 "nb_train_samples": 250000,
245 "nb_test_samples": 10000,
250 "nb_train_samples": 100000,
251 "nb_test_samples": 1000,
256 "nb_train_samples": 1000000,
257 "nb_test_samples": 10000,
262 "nb_train_samples": 50000,
263 "nb_test_samples": 10000,
268 "nb_train_samples": 100000,
269 "nb_test_samples": 10000,
274 "nb_train_samples": 250000,
275 "nb_test_samples": 10000,
280 "nb_train_samples": 2500000,
281 "nb_test_samples": 10000,
286 "nb_train_samples": 250000,
287 "nb_test_samples": 10000,
292 "nb_train_samples": 100000,
293 "nb_test_samples": 1000,
298 "nb_train_samples": 50000,
299 "nb_test_samples": 10000,
304 "nb_train_samples": 25000,
305 "nb_test_samples": 10000,
310 "nb_train_samples": 250000,
311 "nb_test_samples": 10000,
316 "nb_train_samples": 60000,
317 "nb_test_samples": 10000,
321 if args.task in default_task_args:
322 for k, v in default_task_args[args.task].items():
323 if getattr(args, k) is None:
326 ######################################################################
328 default_model_args = {
338 "attention": "caterpillar",
344 "caterpillar_height": 4,
356 "attention": "caterpillar",
362 "caterpillar_height": 4,
374 "attention": "caterpillar",
380 "caterpillar_height": 32,
392 "attention": "caterpillar",
409 "attention": "caterpillar",
419 if args.model in default_model_args:
420 for k, v in default_model_args[args.model].items():
421 if getattr(args, k) is None:
424 raise ValueError(f"Unknown model {args.model}")
426 ######################################################################
429 os.mkdir(args.result_dir)
430 except FileExistsError:
431 if not args.continue_training:
432 print(f"result directory {args.result_dir} already exists")
435 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
438 # torch.backends.cudnn.deterministic = True
439 # torch.backends.cudnn.benchmark = False
440 # torch.use_deterministic_algorithms(True)
441 torch.manual_seed(args.seed)
442 if torch.cuda.is_available():
443 torch.cuda.manual_seed_all(args.seed)
445 ######################################################################
449 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
451 if log_file is not None:
452 log_file.write(t + s + "\n")
459 with os.popen("sha256sum *.py") as f:
461 log_string(f"sha256sum {l.strip()}")
463 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
464 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
466 log_string(f"argv {' '.join(sys.argv)}")
469 log_string(f"args.{n} {getattr(args, n)}")
472 ######################################################################
475 def get_lr(n_epoch, it):
476 if args.legacy_lr_schedule:
477 # my crude scheduling to compare to previous baseline, added
480 if it < args.nb_warmup_iter:
481 return args.legacy_large_lr * it / args.nb_warmup_iter
482 elif n_epoch < args.legacy_nb_epoch_large_lr:
483 return args.legacy_large_lr
485 return args.legacy_small_lr
489 # 1) linear warmup for warmup_iter steps
490 if it < args.nb_warmup_iter:
491 return args.learning_rate * it / args.nb_warmup_iter
492 # 2) if it > nb_decay_iter, return min learning rate
493 if it > args.nb_decay_iter:
494 return args.min_learning_rate
495 # 3) in between, use cosine decay down to min learning rate
496 decay_ratio = (it - args.nb_warmup_iter) / (
497 args.nb_decay_iter - args.nb_warmup_iter
499 coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
500 return args.min_learning_rate + coeff * (
501 args.learning_rate - args.min_learning_rate
505 ######################################################################
508 def picoclvr_pruner_horizontal_green(p):
509 return not ("green" in p and ("left" in p or "right" in p))
512 picoclvr_pruner_train = (
513 picoclvr_pruner_horizontal_green
514 if args.picocvlr_prune_properties in {"train+eval"}
518 picoclvr_pruner_eval = (
519 (lambda p: not picoclvr_pruner_horizontal_green(p))
520 if args.picocvlr_prune_properties in {"train+eval", "eval"}
524 ######################################################################
528 if args.task == "byheart":
529 task = tasks.SandBox(
530 problem=problems.ProblemByHeart(),
531 nb_train_samples=args.nb_train_samples,
532 nb_test_samples=args.nb_test_samples,
533 batch_size=args.batch_size,
537 args.max_percents_of_test_in_train = -1
539 elif args.task == "learnop":
540 task = tasks.SandBox(
541 problem=problems.ProblemLearnOperator(),
542 nb_train_samples=args.nb_train_samples,
543 nb_test_samples=args.nb_test_samples,
544 batch_size=args.batch_size,
550 elif args.task == "guessop":
551 task = tasks.SandBox(
552 problem=problems.ProblemGuessOperator(),
553 nb_train_samples=args.nb_train_samples,
554 nb_test_samples=args.nb_test_samples,
555 batch_size=args.batch_size,
561 elif args.task == "twotargets":
562 task = tasks.SandBox(
563 problem=problems.ProblemTwoTargets(),
564 nb_train_samples=args.nb_train_samples,
565 nb_test_samples=args.nb_test_samples,
566 batch_size=args.batch_size,
571 elif args.task == "memory":
572 task = tasks.SandBox(
573 problem=problems.ProblemMemory(len_total=args.memory_len_total),
574 nb_train_samples=args.nb_train_samples,
575 nb_test_samples=args.nb_test_samples,
576 batch_size=args.batch_size,
581 elif args.task == "mixing":
582 task = tasks.SandBox(
583 problem=problems.ProblemMixing(
584 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
586 nb_train_samples=args.nb_train_samples,
587 nb_test_samples=args.nb_test_samples,
588 batch_size=args.batch_size,
593 elif args.task == "addition":
594 task = tasks.SandBox(
595 problem=problems.ProblemAddition(),
596 nb_train_samples=args.nb_train_samples,
597 nb_test_samples=args.nb_test_samples,
598 batch_size=args.batch_size,
603 elif args.task == "picoclvr":
604 task = tasks.PicoCLVR(
605 nb_train_samples=args.nb_train_samples,
606 nb_test_samples=args.nb_test_samples,
607 batch_size=args.batch_size,
608 height=args.picoclvr_height,
609 width=args.picoclvr_width,
610 nb_colors=args.picoclvr_nb_colors,
613 pruner_train=picoclvr_pruner_train,
614 pruner_eval=picoclvr_pruner_eval,
617 elif args.task == "mnist":
619 nb_train_samples=args.nb_train_samples,
620 nb_test_samples=args.nb_test_samples,
621 batch_size=args.batch_size,
625 elif args.task == "maze":
627 nb_train_samples=args.nb_train_samples,
628 nb_test_samples=args.nb_test_samples,
629 batch_size=args.batch_size,
630 height=args.maze_height,
631 width=args.maze_width,
632 nb_walls=args.maze_nb_walls,
636 elif args.task == "snake":
638 nb_train_samples=args.nb_train_samples,
639 nb_test_samples=args.nb_test_samples,
640 batch_size=args.batch_size,
641 height=args.snake_height,
642 width=args.snake_width,
643 nb_colors=args.snake_nb_colors,
644 length=args.snake_length,
645 prompt_length=args.snake_length // 2,
649 elif args.task == "stack":
651 nb_train_samples=args.nb_train_samples,
652 nb_test_samples=args.nb_test_samples,
653 batch_size=args.batch_size,
655 nb_steps=args.stack_nb_steps,
656 nb_stacks=args.stack_nb_stacks,
657 nb_digits=args.stack_nb_digits,
658 fraction_values_for_train=args.stack_fraction_values_for_train,
662 elif args.task == "expr":
664 nb_train_samples=args.nb_train_samples,
665 nb_test_samples=args.nb_test_samples,
666 nb_variables=args.expr_nb_variables,
667 sequence_length=args.expr_sequence_length,
668 operand_max=args.expr_operand_max,
669 result_max=args.expr_result_max,
670 batch_size=args.batch_size,
674 elif args.task == "rpl":
676 nb_train_samples=args.nb_train_samples,
677 nb_test_samples=args.nb_test_samples,
678 batch_size=args.batch_size,
679 nb_starting_values=args.rpl_nb_starting_values,
680 max_input=args.rpl_max_input,
681 prog_len=args.rpl_prog_len,
682 nb_runs=args.rpl_nb_runs,
683 no_prog=args.rpl_no_prog,
688 elif args.task == "grid":
690 nb_train_samples=args.nb_train_samples,
691 nb_test_samples=args.nb_test_samples,
692 batch_size=args.batch_size,
698 elif args.task == "qmlp":
700 nb_train_samples=args.nb_train_samples,
701 nb_test_samples=args.nb_test_samples,
702 batch_size=args.batch_size,
703 result_dir=args.result_dir,
709 raise ValueError(f"Unknown task {args.task}")
711 ######################################################################
713 log_string(f"device {device}")
715 vocabulary_size = task.vocabulary_size()
717 log_string(f"vocabulary_size {vocabulary_size}")
719 ##############################
722 vocabulary_size=vocabulary_size,
723 dim_model=args.dim_model,
724 dim_keys=args.dim_keys,
725 dim_hidden=args.dim_hidden,
726 nb_heads=args.nb_heads,
727 nb_lines=args.nb_lines,
728 caterpillar_height=args.caterpillar_height,
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 = datetime.datetime.now()
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()
916 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
918 time_pred_result = time_current_result
921 "nb_epochs_finished": n_epoch + 1,
922 "model_state": model.state_dict(),
923 "rng_state": torch.get_rng_state(),
926 if torch.cuda.is_available():
927 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
929 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
930 torch.save(checkpoint, checkpoint_name)
931 log_string(f"saved checkpoint {checkpoint_name}")
933 ######################################################################