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,
480 args.max_percents_of_test_in_train = -1
482 elif args.task == "learnop":
483 task = tasks.SandBox(
484 problem=problems.ProblemLearnOperator(),
485 nb_train_samples=args.nb_train_samples,
486 nb_test_samples=args.nb_test_samples,
487 batch_size=args.physical_batch_size,
493 elif args.task == "guessop":
494 task = tasks.SandBox(
495 problem=problems.ProblemGuessOperator(),
496 nb_train_samples=args.nb_train_samples,
497 nb_test_samples=args.nb_test_samples,
498 batch_size=args.physical_batch_size,
504 elif args.task == "twotargets":
505 task = tasks.SandBox(
506 problem=problems.ProblemTwoTargets(),
507 nb_train_samples=args.nb_train_samples,
508 nb_test_samples=args.nb_test_samples,
509 batch_size=args.physical_batch_size,
514 elif args.task == "memory":
515 task = tasks.SandBox(
516 problem=problems.ProblemMemory(),
517 nb_train_samples=args.nb_train_samples,
518 nb_test_samples=args.nb_test_samples,
519 batch_size=args.physical_batch_size,
524 elif args.task == "mixing":
525 task = tasks.SandBox(
526 problem=problems.ProblemMixing(
527 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
529 nb_train_samples=args.nb_train_samples,
530 nb_test_samples=args.nb_test_samples,
531 batch_size=args.physical_batch_size,
536 elif args.task == "addition":
537 task = tasks.SandBox(
538 problem=problems.ProblemAddition(),
539 nb_train_samples=args.nb_train_samples,
540 nb_test_samples=args.nb_test_samples,
541 batch_size=args.physical_batch_size,
546 elif args.task == "picoclvr":
547 task = tasks.PicoCLVR(
548 nb_train_samples=args.nb_train_samples,
549 nb_test_samples=args.nb_test_samples,
550 batch_size=args.physical_batch_size,
551 height=args.picoclvr_height,
552 width=args.picoclvr_width,
553 nb_colors=args.picoclvr_nb_colors,
556 pruner_train=picoclvr_pruner_train,
557 pruner_eval=picoclvr_pruner_eval,
560 elif args.task == "mnist":
562 nb_train_samples=args.nb_train_samples,
563 nb_test_samples=args.nb_test_samples,
564 batch_size=args.physical_batch_size,
568 elif args.task == "maze":
570 nb_train_samples=args.nb_train_samples,
571 nb_test_samples=args.nb_test_samples,
572 batch_size=args.physical_batch_size,
573 height=args.maze_height,
574 width=args.maze_width,
575 nb_walls=args.maze_nb_walls,
579 elif args.task == "snake":
581 nb_train_samples=args.nb_train_samples,
582 nb_test_samples=args.nb_test_samples,
583 batch_size=args.physical_batch_size,
584 height=args.snake_height,
585 width=args.snake_width,
586 nb_colors=args.snake_nb_colors,
587 length=args.snake_length,
588 prompt_length=args.snake_length // 2,
592 elif args.task == "stack":
594 nb_train_samples=args.nb_train_samples,
595 nb_test_samples=args.nb_test_samples,
596 batch_size=args.physical_batch_size,
598 nb_steps=args.stack_nb_steps,
599 nb_stacks=args.stack_nb_stacks,
600 nb_digits=args.stack_nb_digits,
601 fraction_values_for_train=args.stack_fraction_values_for_train,
605 elif args.task == "expr":
607 nb_train_samples=args.nb_train_samples,
608 nb_test_samples=args.nb_test_samples,
609 nb_variables=args.expr_nb_variables,
610 sequence_length=args.expr_sequence_length,
611 operand_max=args.expr_operand_max,
612 result_max=args.expr_result_max,
613 batch_size=args.physical_batch_size,
617 elif args.task == "rpl":
619 nb_train_samples=args.nb_train_samples,
620 nb_test_samples=args.nb_test_samples,
621 batch_size=args.physical_batch_size,
622 nb_starting_values=args.rpl_nb_starting_values,
623 max_input=args.rpl_max_input,
624 prog_len=args.rpl_prog_len,
625 nb_runs=args.rpl_nb_runs,
626 no_prog=args.rpl_no_prog,
631 elif args.task == "grid":
633 nb_train_samples=args.nb_train_samples,
634 nb_test_samples=args.nb_test_samples,
635 batch_size=args.physical_batch_size,
637 fraction_play=args.grid_fraction_play,
642 elif args.task == "qmlp":
644 nb_train_samples=args.nb_train_samples,
645 nb_test_samples=args.nb_test_samples,
646 batch_size=args.physical_batch_size,
647 result_dir=args.result_dir,
652 elif args.task == "greed":
654 nb_train_samples=args.nb_train_samples,
655 nb_test_samples=args.nb_test_samples,
656 batch_size=args.physical_batch_size,
657 height=args.greed_height,
658 width=args.greed_width,
660 nb_walls=args.greed_nb_walls,
661 nb_coins=args.greed_nb_coins,
667 raise ValueError(f"Unknown task {args.task}")
669 ######################################################################
671 log_string(f"device {device}")
673 vocabulary_size = task.vocabulary_size()
675 log_string(f"vocabulary_size {vocabulary_size}")
677 ##############################
680 vocabulary_size=vocabulary_size,
681 dim_model=args.dim_model,
682 dim_keys=args.dim_keys,
683 dim_hidden=args.dim_hidden,
684 nb_heads=args.nb_heads,
685 nb_blocks=args.nb_blocks,
687 dropout=args.dropout,
692 nb_parameters = sum(p.numel() for p in model.parameters())
693 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
695 ######################################################################
697 nb_epochs_finished = 0
699 if args.no_checkpoint:
700 log_string(f"not trying to load checkpoint.")
704 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
705 checkpoint = torch.load(checkpoint_name)
706 nb_epochs_finished = checkpoint["nb_epochs_finished"]
707 model.load_state_dict(checkpoint["model_state"])
708 torch.set_rng_state(checkpoint["rng_state"])
709 if torch.cuda.is_available():
710 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
712 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
714 except FileNotFoundError:
715 log_string("starting from scratch.")
718 log_string("error when loading the checkpoint.")
721 ######################################################################
723 if args.task == "expr" and args.expr_input_file is not None:
724 task.produce_results(
725 n_epoch=nb_epochs_finished,
727 result_dir=args.result_dir,
729 deterministic_synthesis=args.deterministic_synthesis,
730 input_file=args.expr_input_file,
735 ######################################################################
737 # Compute the entropy of the training tokens
740 for input in task.batches(split="train", desc="train-entropy"):
741 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
742 token_probas = token_count / token_count.sum()
743 entropy = -torch.xlogy(token_probas, token_probas).sum()
744 train_set_perplexity = math.exp(entropy)
746 ######################################################################
747 # A bit of paranoia never hurts
749 if args.max_percents_of_test_in_train >= 0:
751 def subsets_as_tuples(batches, cs):
753 for batch in batches:
755 s.add(tuple([v.item() for v in x]))
761 nb_test, nb_in_train = 0, 0
762 for test_subset in subsets_as_tuples(
763 task.batches(split="test", desc="test-check"), 25000
766 for train_subset in subsets_as_tuples(
767 task.batches(split="train", desc="train-check"), 25000
769 in_train.update(test_subset.intersection(train_subset))
770 nb_in_train += len(in_train)
771 nb_test += len(test_subset)
774 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
778 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
779 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
781 ##############################
783 if args.learning_rate_schedule == "cos":
784 learning_rate_schedule = {}
785 for n_epoch in range(args.nb_epochs):
786 u = n_epoch / args.nb_epochs * math.pi
787 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
792 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
796 learning_rate_schedule = {}
797 learning_rate = args.learning_rate
798 for n_epoch in range(args.nb_epochs):
800 learning_rate = u[n_epoch]
801 learning_rate_schedule[n_epoch] = learning_rate
803 log_string(f"learning_rate_schedule {learning_rate_schedule}")
805 ##############################
807 if nb_epochs_finished >= args.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
818 ######################################################################
821 def one_epoch(model, task, learning_rate):
822 log_string(f"learning_rate {learning_rate}")
824 if args.optim == "sgd":
825 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
826 elif args.optim == "adam":
827 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
828 elif args.optim == "adamw":
829 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
831 raise ValueError(f"Unknown optimizer {args.optim}.")
835 nb_train_samples, acc_train_loss = 0, 0.0
837 for input in task.batches(split="train"):
838 input = input.to(device)
840 if nb_train_samples % args.batch_size == 0:
841 optimizer.zero_grad()
843 output = model(mygpt.BracketedSequence(input)).x
844 loss = F.cross_entropy(output.transpose(1, 2), input)
845 acc_train_loss += loss.item() * input.size(0)
847 nb_train_samples += input.size(0)
851 if nb_train_samples % args.batch_size == 0:
854 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
856 log_string(f"train)perplexity {n_epoch} {train_perplexity}")
859 ######################################################################
862 def run_tests(model, task, deterministic_synthesis):
863 with torch.autograd.no_grad():
866 nb_test_samples, acc_test_loss = 0, 0.0
867 nb_samples_accumulated = 0
869 for input in task.batches(split="test"):
870 input = input.to(device)
872 bs = model(mygpt.BracketedSequence(input))
875 loss = F.cross_entropy(output.transpose(1, 2), input)
877 acc_test_loss += loss.item() * input.size(0)
879 nb_test_samples += input.size(0)
881 task.produce_results(
884 result_dir=args.result_dir,
886 deterministic_synthesis=deterministic_synthesis,
889 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
890 log_string(f"test)perplexity {n_epoch} {test_perplexity}")
893 ######################################################################
895 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
896 learning_rate = learning_rate_schedule[n_epoch]
898 one_epoch(model, task, learning_rate)
900 run_tests(model, task, deterministic_synthesis=True)
902 # --------------------------------------------
904 time_current_result = datetime.datetime.now()
905 if time_pred_result is not None:
907 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
909 time_pred_result = time_current_result
911 # --------------------------------------------
914 "nb_epochs_finished": n_epoch + 1,
915 "model_state": model.state_dict(),
916 "rng_state": torch.get_rng_state(),
919 if torch.cuda.is_available():
920 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
922 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
923 torch.save(checkpoint, checkpoint_name)
924 log_string(f"saved checkpoint {checkpoint_name}")
926 ######################################################################