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=5)
183 parser.add_argument("--escape_width", type=int, default=7)
185 parser.add_argument("--escape_T", type=int, default=25)
187 parser.add_argument("--escape_nb_walls", type=int, default=5)
189 ######################################################################
191 args = parser.parse_args()
193 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
195 if args.result_dir is None:
196 args.result_dir = f"results_{args.task}"
198 ######################################################################
200 default_task_args = {
204 "nb_train_samples": 250000,
205 "nb_test_samples": 10000,
210 "nb_train_samples": 250000,
211 "nb_test_samples": 10000,
216 "nb_train_samples": 50000,
217 "nb_test_samples": 10000,
222 "nb_train_samples": 2500000,
223 "nb_test_samples": 10000,
228 "nb_train_samples": 250000,
229 "nb_test_samples": 10000,
234 "nb_train_samples": 100000,
235 "nb_test_samples": 1000,
240 "nb_train_samples": 1000000,
241 "nb_test_samples": 10000,
246 "nb_train_samples": 50000,
247 "nb_test_samples": 10000,
252 "nb_train_samples": 100000,
253 "nb_test_samples": 10000,
258 "nb_train_samples": 250000,
259 "nb_test_samples": 10000,
264 "nb_train_samples": 2500000,
265 "nb_test_samples": 10000,
270 "nb_train_samples": 250000,
271 "nb_test_samples": 10000,
276 "nb_train_samples": 100000,
277 "nb_test_samples": 1000,
282 "nb_train_samples": 50000,
283 "nb_test_samples": 10000,
288 "nb_train_samples": 25000,
289 "nb_test_samples": 1000,
294 "nb_train_samples": 250000,
295 "nb_test_samples": 10000,
300 "nb_train_samples": 60000,
301 "nb_test_samples": 10000,
306 "nb_train_samples": 25000,
307 "nb_test_samples": 10000,
311 if args.task in default_task_args:
312 for k, v in default_task_args[args.task].items():
313 if getattr(args, k) is None:
316 ######################################################################
318 default_model_args = {
356 if args.model in default_model_args:
357 for k, v in default_model_args[args.model].items():
358 if getattr(args, k) is None:
361 raise ValueError(f"Unknown model {args.model}")
363 ######################################################################
366 os.mkdir(args.result_dir)
367 except FileExistsError:
368 if not args.overwrite_results:
369 print(f"result directory {args.result_dir} already exists")
372 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
375 # torch.backends.cudnn.deterministic = True
376 # torch.backends.cudnn.benchmark = False
377 # torch.use_deterministic_algorithms(True)
378 torch.manual_seed(args.seed)
379 if torch.cuda.is_available():
380 torch.cuda.manual_seed_all(args.seed)
382 ######################################################################
386 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
388 if log_file is not None:
389 log_file.write(t + s + "\n")
396 log_string(f"argv {' '.join(sys.argv)}")
399 log_string(f"args.{n} {getattr(args, n)}")
402 ######################################################################
405 def picoclvr_pruner_horizontal_green(p):
406 return not ("green" in p and ("left" in p or "right" in p))
409 picoclvr_pruner_train = (
410 picoclvr_pruner_horizontal_green
411 if args.picocvlr_prune_properties in {"train+eval"}
415 picoclvr_pruner_eval = (
416 (lambda p: not picoclvr_pruner_horizontal_green(p))
417 if args.picocvlr_prune_properties in {"train+eval", "eval"}
421 ######################################################################
423 if args.task == "file":
425 args.filetask_train_file is not None and args.filetask_test_file is not None
426 ), "You have to specify the task train and test files"
427 task = tasks.TaskFromFile(
428 args.filetask_train_file,
429 args.filetask_test_file,
430 nb_train_samples=args.nb_train_samples,
431 nb_test_samples=args.nb_test_samples,
432 batch_size=args.batch_size,
436 args.max_percents_of_test_in_train = 0
438 elif args.task == "byheart":
439 task = tasks.SandBox(
440 problem=problems.ProblemByHeart(),
441 nb_train_samples=args.nb_train_samples,
442 nb_test_samples=args.nb_test_samples,
443 batch_size=args.batch_size,
447 args.max_percents_of_test_in_train = -1
449 elif args.task == "learnop":
450 task = tasks.SandBox(
451 problem=problems.ProblemLearnOperator(),
452 nb_train_samples=args.nb_train_samples,
453 nb_test_samples=args.nb_test_samples,
454 batch_size=args.batch_size,
460 elif args.task == "guessop":
461 task = tasks.SandBox(
462 problem=problems.ProblemGuessOperator(),
463 nb_train_samples=args.nb_train_samples,
464 nb_test_samples=args.nb_test_samples,
465 batch_size=args.batch_size,
471 elif args.task == "twotargets":
472 task = tasks.SandBox(
473 problem=problems.ProblemTwoTargets(),
474 nb_train_samples=args.nb_train_samples,
475 nb_test_samples=args.nb_test_samples,
476 batch_size=args.batch_size,
481 elif args.task == "memory":
482 task = tasks.SandBox(
483 problem=problems.ProblemMemory(),
484 nb_train_samples=args.nb_train_samples,
485 nb_test_samples=args.nb_test_samples,
486 batch_size=args.batch_size,
491 elif args.task == "mixing":
492 task = tasks.SandBox(
493 problem=problems.ProblemMixing(
494 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
496 nb_train_samples=args.nb_train_samples,
497 nb_test_samples=args.nb_test_samples,
498 batch_size=args.batch_size,
503 elif args.task == "addition":
504 task = tasks.SandBox(
505 problem=problems.ProblemAddition(),
506 nb_train_samples=args.nb_train_samples,
507 nb_test_samples=args.nb_test_samples,
508 batch_size=args.batch_size,
513 elif args.task == "picoclvr":
514 task = tasks.PicoCLVR(
515 nb_train_samples=args.nb_train_samples,
516 nb_test_samples=args.nb_test_samples,
517 batch_size=args.batch_size,
518 height=args.picoclvr_height,
519 width=args.picoclvr_width,
520 nb_colors=args.picoclvr_nb_colors,
523 pruner_train=picoclvr_pruner_train,
524 pruner_eval=picoclvr_pruner_eval,
527 elif args.task == "mnist":
529 nb_train_samples=args.nb_train_samples,
530 nb_test_samples=args.nb_test_samples,
531 batch_size=args.batch_size,
535 elif args.task == "maze":
537 nb_train_samples=args.nb_train_samples,
538 nb_test_samples=args.nb_test_samples,
539 batch_size=args.batch_size,
540 height=args.maze_height,
541 width=args.maze_width,
542 nb_walls=args.maze_nb_walls,
546 elif args.task == "snake":
548 nb_train_samples=args.nb_train_samples,
549 nb_test_samples=args.nb_test_samples,
550 batch_size=args.batch_size,
551 height=args.snake_height,
552 width=args.snake_width,
553 nb_colors=args.snake_nb_colors,
554 length=args.snake_length,
555 prompt_length=args.snake_length // 2,
559 elif args.task == "stack":
561 nb_train_samples=args.nb_train_samples,
562 nb_test_samples=args.nb_test_samples,
563 batch_size=args.batch_size,
565 nb_steps=args.stack_nb_steps,
566 nb_stacks=args.stack_nb_stacks,
567 nb_digits=args.stack_nb_digits,
568 fraction_values_for_train=args.stack_fraction_values_for_train,
572 elif args.task == "expr":
574 nb_train_samples=args.nb_train_samples,
575 nb_test_samples=args.nb_test_samples,
576 nb_variables=args.expr_nb_variables,
577 sequence_length=args.expr_sequence_length,
578 operand_max=args.expr_operand_max,
579 result_max=args.expr_result_max,
580 batch_size=args.batch_size,
584 elif args.task == "rpl":
586 nb_train_samples=args.nb_train_samples,
587 nb_test_samples=args.nb_test_samples,
588 batch_size=args.batch_size,
589 nb_starting_values=args.rpl_nb_starting_values,
590 max_input=args.rpl_max_input,
591 prog_len=args.rpl_prog_len,
592 nb_runs=args.rpl_nb_runs,
593 no_prog=args.rpl_no_prog,
598 elif args.task == "grid":
600 nb_train_samples=args.nb_train_samples,
601 nb_test_samples=args.nb_test_samples,
602 batch_size=args.batch_size,
604 fraction_play=args.grid_fraction_play,
609 elif args.task == "qmlp":
611 nb_train_samples=args.nb_train_samples,
612 nb_test_samples=args.nb_test_samples,
613 batch_size=args.batch_size,
614 result_dir=args.result_dir,
619 elif args.task == "escape":
621 nb_train_samples=args.nb_train_samples,
622 nb_test_samples=args.nb_test_samples,
623 batch_size=args.batch_size,
624 height=args.escape_height,
625 width=args.escape_width,
627 nb_walls=args.escape_nb_walls,
633 raise ValueError(f"Unknown task {args.task}")
635 ######################################################################
637 log_string(f"device {device}")
639 vocabulary_size = task.vocabulary_size()
641 log_string(f"vocabulary_size {vocabulary_size}")
643 ##############################
646 vocabulary_size=vocabulary_size,
647 dim_model=args.dim_model,
648 dim_keys=args.dim_keys,
649 dim_hidden=args.dim_hidden,
650 nb_heads=args.nb_heads,
651 nb_blocks=args.nb_blocks,
653 dropout=args.dropout,
658 nb_parameters = sum(p.numel() for p in model.parameters())
659 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
661 ######################################################################
663 nb_epochs_finished = 0
665 if args.no_checkpoint:
666 log_string(f"not trying to load checkpoint.")
670 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
671 checkpoint = torch.load(checkpoint_name)
672 nb_epochs_finished = checkpoint["nb_epochs_finished"]
673 model.load_state_dict(checkpoint["model_state"])
674 torch.set_rng_state(checkpoint["rng_state"])
675 if torch.cuda.is_available():
676 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
678 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
680 except FileNotFoundError:
681 log_string("starting from scratch.")
684 log_string("error when loading the checkpoint.")
687 ######################################################################
689 if args.task == "expr" and args.expr_input_file is not None:
690 task.produce_results(
691 n_epoch=nb_epochs_finished,
693 result_dir=args.result_dir,
695 deterministic_synthesis=args.deterministic_synthesis,
696 input_file=args.expr_input_file,
701 ######################################################################
703 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
705 # Compute the entropy of the training tokens
708 for input in task.batches(split="train"):
709 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
710 token_probas = token_count / token_count.sum()
711 entropy = -torch.xlogy(token_probas, token_probas).sum()
712 train_set_perplexity = math.exp(entropy)
714 ######################################################################
715 # A bit of paranoia never hurts
717 if args.max_percents_of_test_in_train >= 0:
719 def subsets_as_tuples(batches, cs):
721 for batch in batches:
723 s.add(tuple([v.item() for v in x]))
729 nb_test, nb_in_train = 0, 0
730 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
732 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
733 in_train.update(test_subset.intersection(train_subset))
734 nb_in_train += len(in_train)
735 nb_test += len(test_subset)
738 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
742 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
743 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
745 ##############################
747 if args.learning_rate_schedule == "cos":
748 learning_rate_schedule = {}
749 for n_epoch in range(args.nb_epochs):
750 u = n_epoch / args.nb_epochs * math.pi
751 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
756 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
760 learning_rate_schedule = {}
761 learning_rate = args.learning_rate
762 for n_epoch in range(args.nb_epochs):
764 learning_rate = u[n_epoch]
765 learning_rate_schedule[n_epoch] = learning_rate
767 log_string(f"learning_rate_schedule {learning_rate_schedule}")
769 ##############################
773 if nb_epochs_finished >= nb_epochs:
774 task.produce_results(
775 n_epoch=nb_epochs_finished,
777 result_dir=args.result_dir,
779 deterministic_synthesis=args.deterministic_synthesis,
782 time_pred_result = None
784 for n_epoch in range(nb_epochs_finished, nb_epochs):
785 learning_rate = learning_rate_schedule[n_epoch]
787 log_string(f"learning_rate {learning_rate}")
789 if args.optim == "sgd":
790 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
791 elif args.optim == "adam":
792 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
793 elif args.optim == "adamw":
794 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
796 raise ValueError(f"Unknown optimizer {args.optim}.")
800 nb_train_samples, acc_train_loss = 0, 0.0
802 for input in task.batches(split="train"):
803 input = input.to(device)
804 output = model(mygpt.BracketedSequence(input)).x
805 loss = F.cross_entropy(output.transpose(1, 2), input)
806 acc_train_loss += loss.item() * input.size(0)
807 nb_train_samples += input.size(0)
808 nb_samples_seen += input.size(0)
810 optimizer.zero_grad()
814 with torch.autograd.no_grad():
817 nb_test_samples, acc_test_loss = 0, 0.0
819 for input in task.batches(split="test"):
820 input = input.to(device)
822 output = model(mygpt.BracketedSequence(input)).x
823 loss = F.cross_entropy(output.transpose(1, 2), input)
824 acc_test_loss += loss.item() * input.size(0)
825 nb_test_samples += input.size(0)
827 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
828 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
831 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
834 task.produce_results(
837 result_dir=args.result_dir,
839 deterministic_synthesis=args.deterministic_synthesis,
842 time_current_result = datetime.datetime.now()
843 if time_pred_result is not None:
845 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
847 time_pred_result = time_current_result
850 "nb_epochs_finished": n_epoch + 1,
851 "model_state": model.state_dict(),
852 "rng_state": torch.get_rng_state(),
855 if torch.cuda.is_available():
856 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
858 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
859 torch.save(checkpoint, checkpoint_name)
860 log_string(f"saved checkpoint {checkpoint_name}")
862 ######################################################################