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)
69 ########################################
71 parser.add_argument("--model", type=str, default=None)
73 parser.add_argument("--attention", type=str, default=None)
75 parser.add_argument("--dim_model", type=int, default=None)
77 parser.add_argument("--dim_keys", type=int, default=None)
79 parser.add_argument("--dim_hidden", type=int, default=None)
81 parser.add_argument("--nb_heads", type=int, default=None)
83 parser.add_argument("--nb_lines", type=int, default=None)
85 parser.add_argument("--caterpillar_height", type=int, default=None)
87 parser.add_argument("--rho", type=float, default=0.0)
89 parser.add_argument("--dim_rec_v", type=int, default=None)
91 parser.add_argument("--nb_blocks", type=int, default=None)
93 parser.add_argument("--dropout", type=float, default=0.1)
95 ########################################
97 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
99 parser.add_argument("--no_checkpoint", action="store_true", default=False)
101 parser.add_argument("--overwrite_results", action="store_true", default=False)
103 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
105 ##############################
108 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
110 parser.add_argument("--rpl_max_input", type=int, default=9)
112 parser.add_argument("--rpl_prog_len", type=int, default=8)
114 parser.add_argument("--rpl_nb_runs", type=int, default=5)
116 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
118 ##############################
121 parser.add_argument("--grid_size", type=int, default=6)
123 ##############################
126 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
128 parser.add_argument("--picoclvr_height", type=int, default=12)
130 parser.add_argument("--picoclvr_width", type=int, default=16)
132 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
134 ##############################
137 parser.add_argument("--maze_height", type=int, default=13)
139 parser.add_argument("--maze_width", type=int, default=21)
141 parser.add_argument("--maze_nb_walls", type=int, default=15)
143 ##############################
146 parser.add_argument("--snake_height", type=int, default=9)
148 parser.add_argument("--snake_width", type=int, default=12)
150 parser.add_argument("--snake_nb_colors", type=int, default=5)
152 parser.add_argument("--snake_length", type=int, default=200)
154 ##############################
157 parser.add_argument("--stack_nb_steps", type=int, default=100)
159 parser.add_argument("--stack_nb_stacks", type=int, default=3)
161 parser.add_argument("--stack_nb_digits", type=int, default=3)
163 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
165 ##############################
168 parser.add_argument("--expr_nb_variables", type=int, default=5)
170 parser.add_argument("--expr_sequence_length", type=int, default=40)
172 parser.add_argument("--expr_operand_max", type=int, default=9)
174 parser.add_argument("--expr_result_max", type=int, default=99)
176 parser.add_argument("--expr_input_file", type=str, default=None)
178 ##############################
181 parser.add_argument("--memory_len_total", type=int, default=32)
183 ##############################
186 parser.add_argument("--mixing_hard", action="store_true", default=False)
188 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
190 ######################################################################
192 args = parser.parse_args()
194 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
196 if args.result_dir is None:
197 args.result_dir = f"results_{args.task}_{args.model}"
199 ######################################################################
201 default_task_args = {
205 "nb_train_samples": 250000,
206 "nb_test_samples": 10000,
211 "nb_train_samples": 50000,
212 "nb_test_samples": 10000,
217 "nb_train_samples": 2500000,
218 "nb_test_samples": 10000,
223 "nb_train_samples": 250000,
224 "nb_test_samples": 10000,
229 "nb_train_samples": 100000,
230 "nb_test_samples": 1000,
235 "nb_train_samples": 1000000,
236 "nb_test_samples": 10000,
241 "nb_train_samples": 50000,
242 "nb_test_samples": 10000,
247 "nb_train_samples": 100000,
248 "nb_test_samples": 10000,
253 "nb_train_samples": 250000,
254 "nb_test_samples": 10000,
259 "nb_train_samples": 2500000,
260 "nb_test_samples": 10000,
265 "nb_train_samples": 250000,
266 "nb_test_samples": 10000,
271 "nb_train_samples": 100000,
272 "nb_test_samples": 1000,
277 "nb_train_samples": 50000,
278 "nb_test_samples": 10000,
283 "nb_train_samples": 25000,
284 "nb_test_samples": 10000,
289 "nb_train_samples": 250000,
290 "nb_test_samples": 10000,
295 "nb_train_samples": 60000,
296 "nb_test_samples": 10000,
300 if args.task in default_task_args:
301 for k, v in default_task_args[args.task].items():
302 if getattr(args, k) is None:
305 ######################################################################
307 default_model_args = {
318 "attention": "caterpillar",
324 "caterpillar_height": 4,
338 "attention": "caterpillar",
344 "caterpillar_height": 4,
345 "dim_rec_v": 64, # dim_model / nb_heads
358 "attention": "caterpillar",
364 "caterpillar_height": 32,
378 "attention": "caterpillar",
397 "attention": "caterpillar",
408 if args.model in default_model_args:
409 for k, v in default_model_args[args.model].items():
410 if getattr(args, k) is None:
413 raise ValueError(f"Unknown model {args.model}")
415 ######################################################################
418 os.mkdir(args.result_dir)
419 except FileExistsError:
420 if not args.overwrite_results:
421 print(f"result directory {args.result_dir} already exists")
424 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
427 # torch.backends.cudnn.deterministic = True
428 # torch.backends.cudnn.benchmark = False
429 # torch.use_deterministic_algorithms(True)
430 torch.manual_seed(args.seed)
431 if torch.cuda.is_available():
432 torch.cuda.manual_seed_all(args.seed)
434 ######################################################################
438 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
440 if log_file is not None:
441 log_file.write(t + s + "\n")
448 with os.popen("sha256sum *.py") as f:
450 log_string(f"sha256sum {l.strip()}")
452 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
453 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
455 log_string(f"argv {' '.join(sys.argv)}")
458 log_string(f"args.{n} {getattr(args, n)}")
461 ######################################################################
467 # 1) linear warmup for warmup_iter steps
468 if it < args.nb_warmup_iter:
469 return args.learning_rate * it / args.nb_warmup_iter
470 # 2) if it > nb_decay_iter, return min learning rate
471 if it > args.nb_decay_iter:
472 return args.min_learning_rate
473 # 3) in between, use cosine decay down to min learning rate
474 decay_ratio = (it - args.nb_warmup_iter) / (
475 args.nb_decay_iter - args.nb_warmup_iter
477 coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
478 return args.min_learning_rate + coeff * (
479 args.learning_rate - args.min_learning_rate
483 ######################################################################
486 def picoclvr_pruner_horizontal_green(p):
487 return not ("green" in p and ("left" in p or "right" in p))
490 picoclvr_pruner_train = (
491 picoclvr_pruner_horizontal_green
492 if args.picocvlr_prune_properties in {"train+eval"}
496 picoclvr_pruner_eval = (
497 (lambda p: not picoclvr_pruner_horizontal_green(p))
498 if args.picocvlr_prune_properties in {"train+eval", "eval"}
502 ######################################################################
506 if args.task == "byheart":
507 task = tasks.SandBox(
508 problem=problems.ProblemByHeart(),
509 nb_train_samples=args.nb_train_samples,
510 nb_test_samples=args.nb_test_samples,
511 batch_size=args.batch_size,
515 args.max_percents_of_test_in_train = -1
517 elif args.task == "learnop":
518 task = tasks.SandBox(
519 problem=problems.ProblemLearnOperator(),
520 nb_train_samples=args.nb_train_samples,
521 nb_test_samples=args.nb_test_samples,
522 batch_size=args.batch_size,
528 elif args.task == "guessop":
529 task = tasks.SandBox(
530 problem=problems.ProblemGuessOperator(),
531 nb_train_samples=args.nb_train_samples,
532 nb_test_samples=args.nb_test_samples,
533 batch_size=args.batch_size,
539 elif args.task == "twotargets":
540 task = tasks.SandBox(
541 problem=problems.ProblemTwoTargets(),
542 nb_train_samples=args.nb_train_samples,
543 nb_test_samples=args.nb_test_samples,
544 batch_size=args.batch_size,
549 elif args.task == "memory":
550 task = tasks.SandBox(
551 problem=problems.ProblemMemory(len_total=args.memory_len_total),
552 nb_train_samples=args.nb_train_samples,
553 nb_test_samples=args.nb_test_samples,
554 batch_size=args.batch_size,
559 elif args.task == "mixing":
560 task = tasks.SandBox(
561 problem=problems.ProblemMixing(
562 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
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 == "addition":
572 task = tasks.SandBox(
573 problem=problems.ProblemAddition(),
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 == "picoclvr":
582 task = tasks.PicoCLVR(
583 nb_train_samples=args.nb_train_samples,
584 nb_test_samples=args.nb_test_samples,
585 batch_size=args.batch_size,
586 height=args.picoclvr_height,
587 width=args.picoclvr_width,
588 nb_colors=args.picoclvr_nb_colors,
591 pruner_train=picoclvr_pruner_train,
592 pruner_eval=picoclvr_pruner_eval,
595 elif args.task == "mnist":
597 nb_train_samples=args.nb_train_samples,
598 nb_test_samples=args.nb_test_samples,
599 batch_size=args.batch_size,
603 elif args.task == "maze":
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.maze_height,
609 width=args.maze_width,
610 nb_walls=args.maze_nb_walls,
614 elif args.task == "snake":
616 nb_train_samples=args.nb_train_samples,
617 nb_test_samples=args.nb_test_samples,
618 batch_size=args.batch_size,
619 height=args.snake_height,
620 width=args.snake_width,
621 nb_colors=args.snake_nb_colors,
622 length=args.snake_length,
623 prompt_length=args.snake_length // 2,
627 elif args.task == "stack":
629 nb_train_samples=args.nb_train_samples,
630 nb_test_samples=args.nb_test_samples,
631 batch_size=args.batch_size,
633 nb_steps=args.stack_nb_steps,
634 nb_stacks=args.stack_nb_stacks,
635 nb_digits=args.stack_nb_digits,
636 fraction_values_for_train=args.stack_fraction_values_for_train,
640 elif args.task == "expr":
642 nb_train_samples=args.nb_train_samples,
643 nb_test_samples=args.nb_test_samples,
644 nb_variables=args.expr_nb_variables,
645 sequence_length=args.expr_sequence_length,
646 operand_max=args.expr_operand_max,
647 result_max=args.expr_result_max,
648 batch_size=args.batch_size,
652 elif args.task == "rpl":
654 nb_train_samples=args.nb_train_samples,
655 nb_test_samples=args.nb_test_samples,
656 batch_size=args.batch_size,
657 nb_starting_values=args.rpl_nb_starting_values,
658 max_input=args.rpl_max_input,
659 prog_len=args.rpl_prog_len,
660 nb_runs=args.rpl_nb_runs,
661 no_prog=args.rpl_no_prog,
666 elif args.task == "grid":
668 nb_train_samples=args.nb_train_samples,
669 nb_test_samples=args.nb_test_samples,
670 batch_size=args.batch_size,
676 elif args.task == "qmlp":
678 nb_train_samples=args.nb_train_samples,
679 nb_test_samples=args.nb_test_samples,
680 batch_size=args.batch_size,
681 result_dir=args.result_dir,
687 raise ValueError(f"Unknown task {args.task}")
689 ######################################################################
691 log_string(f"device {device}")
693 vocabulary_size = task.vocabulary_size()
695 log_string(f"vocabulary_size {vocabulary_size}")
697 ##############################
700 vocabulary_size=vocabulary_size,
701 dim_model=args.dim_model,
702 dim_keys=args.dim_keys,
703 dim_hidden=args.dim_hidden,
704 nb_heads=args.nb_heads,
705 nb_lines=args.nb_lines,
706 caterpillar_height=args.caterpillar_height,
707 dim_rec_v=args.dim_rec_v,
708 nb_blocks=args.nb_blocks,
710 dropout=args.dropout,
711 attention_layer=args.attention,
716 nb_parameters = sum(p.numel() for p in model.parameters())
717 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
719 ######################################################################
721 nb_epochs_finished = 0
723 if args.no_checkpoint:
724 log_string(f"not trying to load checkpoint.")
728 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
729 checkpoint = torch.load(checkpoint_name)
730 nb_epochs_finished = checkpoint["nb_epochs_finished"]
731 model.load_state_dict(checkpoint["model_state"])
732 torch.set_rng_state(checkpoint["rng_state"])
733 if torch.cuda.is_available():
734 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
736 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
738 except FileNotFoundError:
739 log_string("starting from scratch.")
742 log_string("error when loading the checkpoint.")
745 ######################################################################
747 if args.task == "expr" and args.expr_input_file is not None:
748 task.produce_results(
749 n_epoch=nb_epochs_finished,
751 result_dir=args.result_dir,
753 deterministic_synthesis=args.deterministic_synthesis,
754 input_file=args.expr_input_file,
759 ######################################################################
761 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
763 # Compute the entropy of the training tokens
766 for input in task.batches(split="train"):
767 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
768 token_probas = token_count / token_count.sum()
769 entropy = -torch.xlogy(token_probas, token_probas).sum()
770 train_set_perplexity = math.exp(entropy)
772 ######################################################################
773 # A bit of paranoia never hurts
775 if args.max_percents_of_test_in_train >= 0:
777 def subsets_as_tuples(batches, cs):
779 for batch in batches:
781 s.add(tuple([v.item() for v in x]))
787 nb_test, nb_in_train = 0, 0
788 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
790 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
791 in_train.update(test_subset.intersection(train_subset))
792 nb_in_train += len(in_train)
793 nb_test += len(test_subset)
796 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
800 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
801 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
803 ##############################
807 if nb_epochs_finished >= nb_epochs:
808 task.produce_results(
809 n_epoch=nb_epochs_finished,
811 result_dir=args.result_dir,
813 deterministic_synthesis=args.deterministic_synthesis,
816 time_pred_result = None
820 for n_epoch in range(nb_epochs_finished, nb_epochs):
821 if args.optim == "sgd":
822 optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
823 elif args.optim == "adam":
824 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
825 elif args.optim == "adamw":
826 optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
828 raise ValueError(f"Unknown optimizer {args.optim}.")
832 nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
834 for input in task.batches(split="train"):
835 model.reset_inner_loss()
836 input = input.to(device)
838 output = model(mygpt.BracketedSequence(input)).x
839 loss = F.cross_entropy(output.transpose(1, 2), input)
840 inner_loss = model.get_inner_loss()
842 acc_train_loss += loss.item() * input.size(0)
843 acc_train_inner_loss += inner_loss.item() * input.size(0)
845 nb_train_samples += input.size(0)
846 nb_samples_seen += input.size(0)
848 total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
852 for param_group in optimizer.param_groups:
853 param_group["lr"] = lr
855 # log_string(f"learning_rate {lr}")
857 optimizer.zero_grad()
858 total_loss.backward()
861 with torch.autograd.no_grad():
864 nb_test_samples, acc_test_loss = 0, 0.0
866 for input in task.batches(split="test"):
867 input = input.to(device)
869 output = model(mygpt.BracketedSequence(input)).x
870 loss = F.cross_entropy(output.transpose(1, 2), input)
871 acc_test_loss += loss.item() * input.size(0)
872 nb_test_samples += input.size(0)
875 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}"
878 task.produce_results(
881 result_dir=args.result_dir,
883 deterministic_synthesis=args.deterministic_synthesis,
886 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
887 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
890 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
893 time_current_result = datetime.datetime.now()
894 if time_pred_result is not None:
896 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
898 time_pred_result = time_current_result
901 "nb_epochs_finished": n_epoch + 1,
902 "model_state": model.state_dict(),
903 "rng_state": torch.get_rng_state(),
906 if torch.cuda.is_available():
907 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
909 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
910 torch.save(checkpoint, checkpoint_name)
911 log_string(f"saved checkpoint {checkpoint_name}")
913 ######################################################################