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
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, escape",
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=25)
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 parser.add_argument("--learning_rate", type=float, default=1e-4)
61 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
63 ########################################
65 parser.add_argument("--model", type=str, default=None)
67 parser.add_argument("--dim_model", type=int, default=None)
69 parser.add_argument("--dim_keys", type=int, default=None)
71 parser.add_argument("--dim_hidden", type=int, default=None)
73 parser.add_argument("--nb_heads", type=int, default=None)
75 parser.add_argument("--nb_blocks", type=int, default=None)
77 parser.add_argument("--dropout", type=float, default=0.1)
79 ########################################
81 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
83 parser.add_argument("--no_checkpoint", action="store_true", default=False)
85 parser.add_argument("--overwrite_results", action="store_true", default=False)
87 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
89 ##############################
92 parser.add_argument("--filetask_train_file", type=str, default=None)
94 parser.add_argument("--filetask_test_file", type=str, default=None)
96 ##############################
99 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
101 parser.add_argument("--rpl_max_input", type=int, default=9)
103 parser.add_argument("--rpl_prog_len", type=int, default=8)
105 parser.add_argument("--rpl_nb_runs", type=int, default=5)
107 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
109 ##############################
112 parser.add_argument("--grid_size", type=int, default=6)
114 parser.add_argument("--grid_fraction_play", type=float, default=0)
116 ##############################
119 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
121 parser.add_argument("--picoclvr_height", type=int, default=12)
123 parser.add_argument("--picoclvr_width", type=int, default=16)
125 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
127 ##############################
130 parser.add_argument("--maze_height", type=int, default=13)
132 parser.add_argument("--maze_width", type=int, default=21)
134 parser.add_argument("--maze_nb_walls", type=int, default=15)
136 ##############################
139 parser.add_argument("--snake_height", type=int, default=9)
141 parser.add_argument("--snake_width", type=int, default=12)
143 parser.add_argument("--snake_nb_colors", type=int, default=5)
145 parser.add_argument("--snake_length", type=int, default=200)
147 ##############################
150 parser.add_argument("--stack_nb_steps", type=int, default=100)
152 parser.add_argument("--stack_nb_stacks", type=int, default=3)
154 parser.add_argument("--stack_nb_digits", type=int, default=3)
156 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
158 ##############################
161 parser.add_argument("--expr_nb_variables", type=int, default=5)
163 parser.add_argument("--expr_sequence_length", type=int, default=40)
165 parser.add_argument("--expr_operand_max", type=int, default=9)
167 parser.add_argument("--expr_result_max", type=int, default=99)
169 parser.add_argument("--expr_input_file", type=str, default=None)
171 ##############################
174 parser.add_argument("--mixing_hard", action="store_true", default=False)
176 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
178 ##############################
181 parser.add_argument("--escape_height", type=int, default=4)
183 parser.add_argument("--escape_width", type=int, default=6)
185 parser.add_argument("--escape_T", type=int, default=25)
187 ######################################################################
189 args = parser.parse_args()
191 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
193 if args.result_dir is None:
194 args.result_dir = f"results_{args.task}"
196 ######################################################################
198 default_task_args = {
202 "nb_train_samples": 250000,
203 "nb_test_samples": 10000,
208 "nb_train_samples": 250000,
209 "nb_test_samples": 10000,
214 "nb_train_samples": 50000,
215 "nb_test_samples": 10000,
220 "nb_train_samples": 2500000,
221 "nb_test_samples": 10000,
226 "nb_train_samples": 250000,
227 "nb_test_samples": 10000,
232 "nb_train_samples": 100000,
233 "nb_test_samples": 1000,
238 "nb_train_samples": 1000000,
239 "nb_test_samples": 10000,
244 "nb_train_samples": 50000,
245 "nb_test_samples": 10000,
250 "nb_train_samples": 100000,
251 "nb_test_samples": 10000,
256 "nb_train_samples": 250000,
257 "nb_test_samples": 10000,
262 "nb_train_samples": 2500000,
263 "nb_test_samples": 10000,
268 "nb_train_samples": 250000,
269 "nb_test_samples": 10000,
274 "nb_train_samples": 100000,
275 "nb_test_samples": 1000,
280 "nb_train_samples": 50000,
281 "nb_test_samples": 10000,
286 "nb_train_samples": 25000,
287 "nb_test_samples": 1000,
292 "nb_train_samples": 250000,
293 "nb_test_samples": 10000,
298 "nb_train_samples": 60000,
299 "nb_test_samples": 10000,
304 "nb_train_samples": 25000,
305 "nb_test_samples": 10000,
309 if args.task in default_task_args:
310 for k, v in default_task_args[args.task].items():
311 if getattr(args, k) is None:
314 ######################################################################
316 default_model_args = {
354 if args.model in default_model_args:
355 for k, v in default_model_args[args.model].items():
356 if getattr(args, k) is None:
359 raise ValueError(f"Unknown model {args.model}")
361 ######################################################################
364 os.mkdir(args.result_dir)
365 except FileExistsError:
366 if not args.overwrite_results:
367 print(f"result directory {args.result_dir} already exists")
370 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
373 # torch.backends.cudnn.deterministic = True
374 # torch.backends.cudnn.benchmark = False
375 # torch.use_deterministic_algorithms(True)
376 torch.manual_seed(args.seed)
377 if torch.cuda.is_available():
378 torch.cuda.manual_seed_all(args.seed)
380 ######################################################################
384 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
386 if log_file is not None:
387 log_file.write(t + s + "\n")
394 log_string(f"argv {' '.join(sys.argv)}")
397 log_string(f"args.{n} {getattr(args, n)}")
400 ######################################################################
403 def picoclvr_pruner_horizontal_green(p):
404 return not ("green" in p and ("left" in p or "right" in p))
407 picoclvr_pruner_train = (
408 picoclvr_pruner_horizontal_green
409 if args.picocvlr_prune_properties in {"train+eval"}
413 picoclvr_pruner_eval = (
414 (lambda p: not picoclvr_pruner_horizontal_green(p))
415 if args.picocvlr_prune_properties in {"train+eval", "eval"}
419 ######################################################################
421 if args.task == "file":
423 args.filetask_train_file is not None and args.filetask_test_file is not None
424 ), "You have to specify the task train and test files"
425 task = tasks.TaskFromFile(
426 args.filetask_train_file,
427 args.filetask_test_file,
428 nb_train_samples=args.nb_train_samples,
429 nb_test_samples=args.nb_test_samples,
430 batch_size=args.batch_size,
434 args.max_percents_of_test_in_train = 0
436 elif args.task == "byheart":
437 task = tasks.SandBox(
438 problem=problems.ProblemByHeart(),
439 nb_train_samples=args.nb_train_samples,
440 nb_test_samples=args.nb_test_samples,
441 batch_size=args.batch_size,
445 args.max_percents_of_test_in_train = -1
447 elif args.task == "learnop":
448 task = tasks.SandBox(
449 problem=problems.ProblemLearnOperator(),
450 nb_train_samples=args.nb_train_samples,
451 nb_test_samples=args.nb_test_samples,
452 batch_size=args.batch_size,
458 elif args.task == "guessop":
459 task = tasks.SandBox(
460 problem=problems.ProblemGuessOperator(),
461 nb_train_samples=args.nb_train_samples,
462 nb_test_samples=args.nb_test_samples,
463 batch_size=args.batch_size,
469 elif args.task == "twotargets":
470 task = tasks.SandBox(
471 problem=problems.ProblemTwoTargets(),
472 nb_train_samples=args.nb_train_samples,
473 nb_test_samples=args.nb_test_samples,
474 batch_size=args.batch_size,
479 elif args.task == "memory":
480 task = tasks.SandBox(
481 problem=problems.ProblemMemory(),
482 nb_train_samples=args.nb_train_samples,
483 nb_test_samples=args.nb_test_samples,
484 batch_size=args.batch_size,
489 elif args.task == "mixing":
490 task = tasks.SandBox(
491 problem=problems.ProblemMixing(
492 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
494 nb_train_samples=args.nb_train_samples,
495 nb_test_samples=args.nb_test_samples,
496 batch_size=args.batch_size,
501 elif args.task == "addition":
502 task = tasks.SandBox(
503 problem=problems.ProblemAddition(),
504 nb_train_samples=args.nb_train_samples,
505 nb_test_samples=args.nb_test_samples,
506 batch_size=args.batch_size,
511 elif args.task == "picoclvr":
512 task = tasks.PicoCLVR(
513 nb_train_samples=args.nb_train_samples,
514 nb_test_samples=args.nb_test_samples,
515 batch_size=args.batch_size,
516 height=args.picoclvr_height,
517 width=args.picoclvr_width,
518 nb_colors=args.picoclvr_nb_colors,
521 pruner_train=picoclvr_pruner_train,
522 pruner_eval=picoclvr_pruner_eval,
525 elif args.task == "mnist":
527 nb_train_samples=args.nb_train_samples,
528 nb_test_samples=args.nb_test_samples,
529 batch_size=args.batch_size,
533 elif args.task == "maze":
535 nb_train_samples=args.nb_train_samples,
536 nb_test_samples=args.nb_test_samples,
537 batch_size=args.batch_size,
538 height=args.maze_height,
539 width=args.maze_width,
540 nb_walls=args.maze_nb_walls,
544 elif args.task == "snake":
546 nb_train_samples=args.nb_train_samples,
547 nb_test_samples=args.nb_test_samples,
548 batch_size=args.batch_size,
549 height=args.snake_height,
550 width=args.snake_width,
551 nb_colors=args.snake_nb_colors,
552 length=args.snake_length,
553 prompt_length=args.snake_length // 2,
557 elif args.task == "stack":
559 nb_train_samples=args.nb_train_samples,
560 nb_test_samples=args.nb_test_samples,
561 batch_size=args.batch_size,
563 nb_steps=args.stack_nb_steps,
564 nb_stacks=args.stack_nb_stacks,
565 nb_digits=args.stack_nb_digits,
566 fraction_values_for_train=args.stack_fraction_values_for_train,
570 elif args.task == "expr":
572 nb_train_samples=args.nb_train_samples,
573 nb_test_samples=args.nb_test_samples,
574 nb_variables=args.expr_nb_variables,
575 sequence_length=args.expr_sequence_length,
576 operand_max=args.expr_operand_max,
577 result_max=args.expr_result_max,
578 batch_size=args.batch_size,
582 elif args.task == "rpl":
584 nb_train_samples=args.nb_train_samples,
585 nb_test_samples=args.nb_test_samples,
586 batch_size=args.batch_size,
587 nb_starting_values=args.rpl_nb_starting_values,
588 max_input=args.rpl_max_input,
589 prog_len=args.rpl_prog_len,
590 nb_runs=args.rpl_nb_runs,
591 no_prog=args.rpl_no_prog,
596 elif args.task == "grid":
598 nb_train_samples=args.nb_train_samples,
599 nb_test_samples=args.nb_test_samples,
600 batch_size=args.batch_size,
602 fraction_play=args.grid_fraction_play,
607 elif args.task == "qmlp":
609 nb_train_samples=args.nb_train_samples,
610 nb_test_samples=args.nb_test_samples,
611 batch_size=args.batch_size,
612 result_dir=args.result_dir,
617 elif args.task == "escape":
619 nb_train_samples=args.nb_train_samples,
620 nb_test_samples=args.nb_test_samples,
621 batch_size=args.batch_size,
622 height=args.escape_height,
623 width=args.escape_width,
630 raise ValueError(f"Unknown task {args.task}")
632 ######################################################################
634 log_string(f"device {device}")
636 vocabulary_size = task.vocabulary_size()
638 log_string(f"vocabulary_size {vocabulary_size}")
640 ##############################
643 vocabulary_size=vocabulary_size,
644 dim_model=args.dim_model,
645 dim_keys=args.dim_keys,
646 dim_hidden=args.dim_hidden,
647 nb_heads=args.nb_heads,
648 nb_blocks=args.nb_blocks,
650 dropout=args.dropout,
655 nb_parameters = sum(p.numel() for p in model.parameters())
656 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
658 ######################################################################
660 nb_epochs_finished = 0
662 if args.no_checkpoint:
663 log_string(f"not trying to load checkpoint.")
667 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
668 checkpoint = torch.load(checkpoint_name)
669 nb_epochs_finished = checkpoint["nb_epochs_finished"]
670 model.load_state_dict(checkpoint["model_state"])
671 torch.set_rng_state(checkpoint["rng_state"])
672 if torch.cuda.is_available():
673 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
675 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
677 except FileNotFoundError:
678 log_string("starting from scratch.")
681 log_string("error when loading the checkpoint.")
684 ######################################################################
686 if args.task == "expr" and args.expr_input_file is not None:
687 task.produce_results(
688 n_epoch=nb_epochs_finished,
690 result_dir=args.result_dir,
692 deterministic_synthesis=args.deterministic_synthesis,
693 input_file=args.expr_input_file,
698 ######################################################################
700 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
702 # Compute the entropy of the training tokens
705 for input in task.batches(split="train"):
706 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
707 token_probas = token_count / token_count.sum()
708 entropy = -torch.xlogy(token_probas, token_probas).sum()
709 train_set_perplexity = math.exp(entropy)
711 ######################################################################
712 # A bit of paranoia never hurts
714 if args.max_percents_of_test_in_train >= 0:
716 def subsets_as_tuples(batches, cs):
718 for batch in batches:
720 s.add(tuple([v.item() for v in x]))
726 nb_test, nb_in_train = 0, 0
727 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
729 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
730 in_train.update(test_subset.intersection(train_subset))
731 nb_in_train += len(in_train)
732 nb_test += len(test_subset)
735 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
739 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
740 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
742 ##############################
744 if args.learning_rate_schedule == "cos":
745 learning_rate_schedule = {}
746 for n_epoch in range(args.nb_epochs):
747 u = n_epoch / args.nb_epochs * math.pi
748 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
753 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
757 learning_rate_schedule = {}
758 learning_rate = args.learning_rate
759 for n_epoch in range(args.nb_epochs):
761 learning_rate = u[n_epoch]
762 learning_rate_schedule[n_epoch] = learning_rate
764 log_string(f"learning_rate_schedule {learning_rate_schedule}")
766 ##############################
770 if nb_epochs_finished >= nb_epochs:
771 task.produce_results(
772 n_epoch=nb_epochs_finished,
774 result_dir=args.result_dir,
776 deterministic_synthesis=args.deterministic_synthesis,
779 time_pred_result = None
781 for n_epoch in range(nb_epochs_finished, nb_epochs):
782 learning_rate = learning_rate_schedule[n_epoch]
784 log_string(f"learning_rate {learning_rate}")
786 if args.optim == "sgd":
787 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
788 elif args.optim == "adam":
789 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
790 elif args.optim == "adamw":
791 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
793 raise ValueError(f"Unknown optimizer {args.optim}.")
797 nb_train_samples, acc_train_loss = 0, 0.0
799 for input in task.batches(split="train"):
800 input = input.to(device)
801 output = model(mygpt.BracketedSequence(input)).x
802 loss = F.cross_entropy(output.transpose(1, 2), input)
803 acc_train_loss += loss.item() * input.size(0)
804 nb_train_samples += input.size(0)
805 nb_samples_seen += input.size(0)
807 optimizer.zero_grad()
811 with torch.autograd.no_grad():
814 nb_test_samples, acc_test_loss = 0, 0.0
816 for input in task.batches(split="test"):
817 input = input.to(device)
819 output = model(mygpt.BracketedSequence(input)).x
820 loss = F.cross_entropy(output.transpose(1, 2), input)
821 acc_test_loss += loss.item() * input.size(0)
822 nb_test_samples += input.size(0)
824 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
825 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
828 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
831 task.produce_results(
834 result_dir=args.result_dir,
836 deterministic_synthesis=args.deterministic_synthesis,
839 time_current_result = datetime.datetime.now()
840 if time_pred_result is not None:
842 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
844 time_pred_result = time_current_result
847 "nb_epochs_finished": n_epoch + 1,
848 "model_state": model.state_dict(),
849 "rng_state": torch.get_rng_state(),
852 if torch.cuda.is_available():
853 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
855 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
856 torch.save(checkpoint, checkpoint_name)
857 log_string(f"saved checkpoint {checkpoint_name}")
859 ######################################################################