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="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
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("--physical_batch_size", type=int, default=None)
55 parser.add_argument("--nb_train_samples", type=int, default=None)
57 parser.add_argument("--nb_test_samples", type=int, default=None)
59 parser.add_argument("--optim", type=str, default="adam")
61 parser.add_argument("--learning_rate", type=float, default=1e-4)
63 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
65 ########################################
67 parser.add_argument("--model", type=str, default=None)
69 parser.add_argument("--dim_model", type=int, default=None)
71 parser.add_argument("--dim_keys", type=int, default=None)
73 parser.add_argument("--dim_hidden", type=int, default=None)
75 parser.add_argument("--nb_heads", type=int, default=None)
77 parser.add_argument("--nb_blocks", type=int, default=None)
79 parser.add_argument("--dropout", type=float, default=0.1)
81 ########################################
83 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
85 parser.add_argument("--no_checkpoint", action="store_true", default=False)
87 parser.add_argument("--resume", action="store_true", default=False)
89 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
91 ##############################
94 parser.add_argument("--filetask_train_file", type=str, default=None)
96 parser.add_argument("--filetask_test_file", type=str, default=None)
98 ##############################
101 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
103 parser.add_argument("--rpl_max_input", type=int, default=9)
105 parser.add_argument("--rpl_prog_len", type=int, default=8)
107 parser.add_argument("--rpl_nb_runs", type=int, default=5)
109 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
111 ##############################
114 parser.add_argument("--grid_size", type=int, default=6)
116 parser.add_argument("--grid_fraction_play", type=float, default=0)
118 ##############################
121 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
123 parser.add_argument("--picoclvr_height", type=int, default=12)
125 parser.add_argument("--picoclvr_width", type=int, default=16)
127 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
129 ##############################
132 parser.add_argument("--maze_height", type=int, default=13)
134 parser.add_argument("--maze_width", type=int, default=21)
136 parser.add_argument("--maze_nb_walls", type=int, default=15)
138 ##############################
141 parser.add_argument("--snake_height", type=int, default=9)
143 parser.add_argument("--snake_width", type=int, default=12)
145 parser.add_argument("--snake_nb_colors", type=int, default=5)
147 parser.add_argument("--snake_length", type=int, default=200)
149 ##############################
152 parser.add_argument("--byheart_separation", type=int, default=1)
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=None)
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("--mixing_hard", action="store_true", default=False)
183 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
185 ##############################
188 parser.add_argument("--greed_height", type=int, default=5)
190 parser.add_argument("--greed_width", type=int, default=7)
192 parser.add_argument("--greed_T", type=int, default=25)
194 parser.add_argument("--greed_nb_walls", type=int, default=5)
196 parser.add_argument("--greed_nb_coins", type=int, default=2)
198 ######################################################################
200 args = parser.parse_args()
202 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
204 if args.result_dir is None:
205 args.result_dir = f"results_{args.task}"
207 ######################################################################
209 default_task_args = {
213 "nb_train_samples": 250000,
214 "nb_test_samples": 10000,
219 "nb_train_samples": 250000,
220 "nb_test_samples": 10000,
225 "nb_train_samples": 50000,
226 "nb_test_samples": 10000,
231 "nb_train_samples": 50000,
232 "nb_test_samples": 10000,
237 "nb_train_samples": 2500000,
238 "nb_test_samples": 10000,
243 "nb_train_samples": 250000,
244 "nb_test_samples": 10000,
249 "nb_train_samples": 100000,
250 "nb_test_samples": 1000,
255 "nb_train_samples": 1000000,
256 "nb_test_samples": 10000,
261 "nb_train_samples": 50000,
262 "nb_test_samples": 10000,
267 "nb_train_samples": 100000,
268 "nb_test_samples": 10000,
273 "nb_train_samples": 250000,
274 "nb_test_samples": 10000,
279 "nb_train_samples": 2500000,
280 "nb_test_samples": 10000,
285 "nb_train_samples": 250000,
286 "nb_test_samples": 10000,
291 "nb_train_samples": 100000,
292 "nb_test_samples": 1000,
297 "nb_train_samples": 50000,
298 "nb_test_samples": 10000,
303 "nb_train_samples": 25000,
304 "nb_test_samples": 1000,
309 "nb_train_samples": 250000,
310 "nb_test_samples": 10000,
315 "nb_train_samples": 60000,
316 "nb_test_samples": 10000,
321 "nb_train_samples": 25000,
322 "nb_test_samples": 10000,
326 if args.task in default_task_args:
327 for k, v in default_task_args[args.task].items():
328 if getattr(args, k) is None:
331 ######################################################################
333 default_model_args = {
371 if args.model in default_model_args:
372 for k, v in default_model_args[args.model].items():
373 if getattr(args, k) is None:
376 raise ValueError(f"Unknown model {args.model}")
378 ######################################################################
381 os.mkdir(args.result_dir)
382 except FileExistsError:
384 print(f"result directory {args.result_dir} already exists")
387 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
390 # torch.backends.cudnn.deterministic = True
391 # torch.backends.cudnn.benchmark = False
392 # torch.use_deterministic_algorithms(True)
393 torch.manual_seed(args.seed)
394 if torch.cuda.is_available():
395 torch.cuda.manual_seed_all(args.seed)
397 ######################################################################
401 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
403 if log_file is not None:
404 log_file.write(t + s + "\n")
411 log_string(f"argv {' '.join(sys.argv)}")
414 log_string(f"args.{n} {getattr(args, n)}")
417 ######################################################################
420 def picoclvr_pruner_horizontal_green(p):
421 return not ("green" in p and ("left" in p or "right" in p))
424 picoclvr_pruner_train = (
425 picoclvr_pruner_horizontal_green
426 if args.picocvlr_prune_properties in {"train+eval"}
430 picoclvr_pruner_eval = (
431 (lambda p: not picoclvr_pruner_horizontal_green(p))
432 if args.picocvlr_prune_properties in {"train+eval", "eval"}
436 ######################################################################
438 if args.physical_batch_size is None:
439 args.physical_batch_size = args.batch_size
441 assert args.batch_size % args.physical_batch_size == 0
443 assert args.nb_train_samples % args.batch_size == 0
444 assert args.nb_test_samples % args.batch_size == 0
446 if args.task == "file":
448 args.filetask_train_file is not None and args.filetask_test_file is not None
449 ), "You have to specify the task train and test files"
450 task = tasks.TaskFromFile(
451 args.filetask_train_file,
452 args.filetask_test_file,
453 nb_train_samples=args.nb_train_samples,
454 nb_test_samples=args.nb_test_samples,
455 batch_size=args.physical_batch_size,
459 args.max_percents_of_test_in_train = 0
461 elif args.task == "byheart":
462 task = tasks.SandBox(
463 problem=problems.ProblemByHeart(separation=args.byheart_separation),
464 nb_train_samples=args.nb_train_samples,
465 nb_test_samples=args.nb_test_samples,
466 batch_size=args.physical_batch_size,
470 args.max_percents_of_test_in_train = -1
472 elif args.task == "world":
474 nb_train_samples=args.nb_train_samples,
475 nb_test_samples=args.nb_test_samples,
476 batch_size=args.physical_batch_size,
477 result_dir=args.result_dir,
481 args.max_percents_of_test_in_train = -1
483 elif args.task == "learnop":
484 task = tasks.SandBox(
485 problem=problems.ProblemLearnOperator(),
486 nb_train_samples=args.nb_train_samples,
487 nb_test_samples=args.nb_test_samples,
488 batch_size=args.physical_batch_size,
494 elif args.task == "guessop":
495 task = tasks.SandBox(
496 problem=problems.ProblemGuessOperator(),
497 nb_train_samples=args.nb_train_samples,
498 nb_test_samples=args.nb_test_samples,
499 batch_size=args.physical_batch_size,
505 elif args.task == "twotargets":
506 task = tasks.SandBox(
507 problem=problems.ProblemTwoTargets(),
508 nb_train_samples=args.nb_train_samples,
509 nb_test_samples=args.nb_test_samples,
510 batch_size=args.physical_batch_size,
515 elif args.task == "memory":
516 task = tasks.SandBox(
517 problem=problems.ProblemMemory(),
518 nb_train_samples=args.nb_train_samples,
519 nb_test_samples=args.nb_test_samples,
520 batch_size=args.physical_batch_size,
525 elif args.task == "mixing":
526 task = tasks.SandBox(
527 problem=problems.ProblemMixing(
528 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
530 nb_train_samples=args.nb_train_samples,
531 nb_test_samples=args.nb_test_samples,
532 batch_size=args.physical_batch_size,
537 elif args.task == "addition":
538 task = tasks.SandBox(
539 problem=problems.ProblemAddition(),
540 nb_train_samples=args.nb_train_samples,
541 nb_test_samples=args.nb_test_samples,
542 batch_size=args.physical_batch_size,
547 elif args.task == "picoclvr":
548 task = tasks.PicoCLVR(
549 nb_train_samples=args.nb_train_samples,
550 nb_test_samples=args.nb_test_samples,
551 batch_size=args.physical_batch_size,
552 height=args.picoclvr_height,
553 width=args.picoclvr_width,
554 nb_colors=args.picoclvr_nb_colors,
557 pruner_train=picoclvr_pruner_train,
558 pruner_eval=picoclvr_pruner_eval,
561 elif args.task == "mnist":
563 nb_train_samples=args.nb_train_samples,
564 nb_test_samples=args.nb_test_samples,
565 batch_size=args.physical_batch_size,
569 elif args.task == "maze":
571 nb_train_samples=args.nb_train_samples,
572 nb_test_samples=args.nb_test_samples,
573 batch_size=args.physical_batch_size,
574 height=args.maze_height,
575 width=args.maze_width,
576 nb_walls=args.maze_nb_walls,
580 elif args.task == "snake":
582 nb_train_samples=args.nb_train_samples,
583 nb_test_samples=args.nb_test_samples,
584 batch_size=args.physical_batch_size,
585 height=args.snake_height,
586 width=args.snake_width,
587 nb_colors=args.snake_nb_colors,
588 length=args.snake_length,
589 prompt_length=args.snake_length // 2,
593 elif args.task == "stack":
595 nb_train_samples=args.nb_train_samples,
596 nb_test_samples=args.nb_test_samples,
597 batch_size=args.physical_batch_size,
599 nb_steps=args.stack_nb_steps,
600 nb_stacks=args.stack_nb_stacks,
601 nb_digits=args.stack_nb_digits,
602 fraction_values_for_train=args.stack_fraction_values_for_train,
606 elif args.task == "expr":
608 nb_train_samples=args.nb_train_samples,
609 nb_test_samples=args.nb_test_samples,
610 nb_variables=args.expr_nb_variables,
611 sequence_length=args.expr_sequence_length,
612 operand_max=args.expr_operand_max,
613 result_max=args.expr_result_max,
614 batch_size=args.physical_batch_size,
618 elif args.task == "rpl":
620 nb_train_samples=args.nb_train_samples,
621 nb_test_samples=args.nb_test_samples,
622 batch_size=args.physical_batch_size,
623 nb_starting_values=args.rpl_nb_starting_values,
624 max_input=args.rpl_max_input,
625 prog_len=args.rpl_prog_len,
626 nb_runs=args.rpl_nb_runs,
627 no_prog=args.rpl_no_prog,
632 elif args.task == "grid":
634 nb_train_samples=args.nb_train_samples,
635 nb_test_samples=args.nb_test_samples,
636 batch_size=args.physical_batch_size,
638 fraction_play=args.grid_fraction_play,
643 elif args.task == "qmlp":
645 nb_train_samples=args.nb_train_samples,
646 nb_test_samples=args.nb_test_samples,
647 batch_size=args.physical_batch_size,
648 result_dir=args.result_dir,
653 elif args.task == "greed":
655 nb_train_samples=args.nb_train_samples,
656 nb_test_samples=args.nb_test_samples,
657 batch_size=args.physical_batch_size,
658 height=args.greed_height,
659 width=args.greed_width,
661 nb_walls=args.greed_nb_walls,
662 nb_coins=args.greed_nb_coins,
668 raise ValueError(f"Unknown task {args.task}")
670 ######################################################################
672 log_string(f"device {device}")
674 vocabulary_size = task.vocabulary_size()
676 log_string(f"vocabulary_size {vocabulary_size}")
678 ##############################
681 vocabulary_size=vocabulary_size,
682 dim_model=args.dim_model,
683 dim_keys=args.dim_keys,
684 dim_hidden=args.dim_hidden,
685 nb_heads=args.nb_heads,
686 nb_blocks=args.nb_blocks,
688 dropout=args.dropout,
693 nb_parameters = sum(p.numel() for p in model.parameters())
694 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
696 ######################################################################
698 nb_epochs_finished = 0
700 if args.no_checkpoint:
701 log_string(f"not trying to load checkpoint.")
705 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
706 checkpoint = torch.load(checkpoint_name)
707 nb_epochs_finished = checkpoint["nb_epochs_finished"]
708 model.load_state_dict(checkpoint["model_state"])
709 torch.set_rng_state(checkpoint["rng_state"])
710 if torch.cuda.is_available():
711 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
713 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
715 except FileNotFoundError:
716 log_string("starting from scratch.")
719 log_string("error when loading the checkpoint.")
722 ######################################################################
724 if args.task == "expr" and args.expr_input_file is not None:
725 task.produce_results(
726 n_epoch=nb_epochs_finished,
728 result_dir=args.result_dir,
730 deterministic_synthesis=args.deterministic_synthesis,
731 input_file=args.expr_input_file,
736 ######################################################################
738 # Compute the entropy of the training tokens
741 for input in task.batches(split="train", desc="train-entropy"):
742 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
743 token_probas = token_count / token_count.sum()
744 entropy = -torch.xlogy(token_probas, token_probas).sum()
745 train_set_perplexity = math.exp(entropy)
747 ######################################################################
748 # A bit of paranoia never hurts
750 if args.max_percents_of_test_in_train >= 0:
752 def subsets_as_tuples(batches, cs):
754 for batch in batches:
756 s.add(tuple([v.item() for v in x]))
762 nb_test, nb_in_train = 0, 0
763 for test_subset in subsets_as_tuples(
764 task.batches(split="test", desc="test-check"), 25000
767 for train_subset in subsets_as_tuples(
768 task.batches(split="train", desc="train-check"), 25000
770 in_train.update(test_subset.intersection(train_subset))
771 nb_in_train += len(in_train)
772 nb_test += len(test_subset)
775 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
779 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
780 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
782 ##############################
784 if args.learning_rate_schedule == "cos":
785 learning_rate_schedule = {}
786 for n_epoch in range(args.nb_epochs):
787 u = n_epoch / args.nb_epochs * math.pi
788 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
793 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
797 learning_rate_schedule = {}
798 learning_rate = args.learning_rate
799 for n_epoch in range(args.nb_epochs):
801 learning_rate = u[n_epoch]
802 learning_rate_schedule[n_epoch] = learning_rate
804 log_string(f"learning_rate_schedule {learning_rate_schedule}")
806 ##############################
808 if nb_epochs_finished >= args.nb_epochs:
809 task.produce_results(
810 n_epoch=nb_epochs_finished,
812 result_dir=args.result_dir,
814 deterministic_synthesis=args.deterministic_synthesis,
817 time_pred_result = None
819 ######################################################################
822 def one_epoch(model, task, learning_rate):
823 log_string(f"learning_rate {learning_rate}")
825 if args.optim == "sgd":
826 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
827 elif args.optim == "adam":
828 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
829 elif args.optim == "adamw":
830 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
832 raise ValueError(f"Unknown optimizer {args.optim}.")
836 nb_train_samples, acc_train_loss = 0, 0.0
838 for input in task.batches(split="train"):
839 input = input.to(device)
841 if nb_train_samples % args.batch_size == 0:
842 optimizer.zero_grad()
844 output = model(mygpt.BracketedSequence(input)).x
845 loss = F.cross_entropy(output.transpose(1, 2), input)
846 acc_train_loss += loss.item() * input.size(0)
848 nb_train_samples += input.size(0)
852 if nb_train_samples % args.batch_size == 0:
855 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
857 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
860 ######################################################################
863 def run_tests(model, task, deterministic_synthesis):
864 with torch.autograd.no_grad():
867 nb_test_samples, acc_test_loss = 0, 0.0
868 nb_samples_accumulated = 0
870 for input in task.batches(split="test"):
871 input = input.to(device)
873 bs = model(mygpt.BracketedSequence(input))
876 loss = F.cross_entropy(output.transpose(1, 2), input)
878 acc_test_loss += loss.item() * input.size(0)
880 nb_test_samples += input.size(0)
882 main_test_accuracy = task.produce_results(
885 result_dir=args.result_dir,
887 deterministic_synthesis=deterministic_synthesis,
890 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
892 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
894 return main_test_accuracy
897 ######################################################################
899 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
900 learning_rate = learning_rate_schedule[n_epoch]
902 one_epoch(model, task, learning_rate)
904 test_accuracy = run_tests(model, task, deterministic_synthesis=False)
906 # --------------------------------------------
908 if test_accuracy >= 0.8:
909 nb_for_train, nb_for_test = 1000, 100
912 while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
913 new_quizzes, nb_correct = task.create_new_quizzes(
915 result_dir=args.result_dir,
922 to_keep = new_quizzes[torch.logical_and(nb_correct >= 8, nb_correct < 10)]
923 log_string(f"keep {to_keep.size(0)} quizzes")
926 new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
928 task.store_new_quizzes(new_quizzes[:nb_for_train], train=True)
929 task.store_new_quizzes(new_quizzes[nb_for_train:], train=False)
934 f"world_new_{n_epoch:04d}.png",
938 # --------------------------------------------
940 time_current_result = datetime.datetime.now()
941 if time_pred_result is not None:
943 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
945 time_pred_result = time_current_result
947 # --------------------------------------------
950 "nb_epochs_finished": n_epoch + 1,
951 "model_state": model.state_dict(),
952 "rng_state": torch.get_rng_state(),
955 if torch.cuda.is_available():
956 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
958 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
959 torch.save(checkpoint, checkpoint_name)
960 log_string(f"saved checkpoint {checkpoint_name}")
962 ######################################################################