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
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, 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=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("--rpl_nb_starting_values", type=int, default=3)
94 parser.add_argument("--rpl_max_input", type=int, default=9)
96 parser.add_argument("--rpl_prog_len", type=int, default=8)
98 parser.add_argument("--rpl_nb_runs", type=int, default=5)
100 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
102 ##############################
105 parser.add_argument("--grid_size", type=int, default=6)
107 ##############################
110 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
112 parser.add_argument("--picoclvr_height", type=int, default=12)
114 parser.add_argument("--picoclvr_width", type=int, default=16)
116 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
118 ##############################
121 parser.add_argument("--maze_height", type=int, default=13)
123 parser.add_argument("--maze_width", type=int, default=21)
125 parser.add_argument("--maze_nb_walls", type=int, default=15)
127 ##############################
130 parser.add_argument("--snake_height", type=int, default=9)
132 parser.add_argument("--snake_width", type=int, default=12)
134 parser.add_argument("--snake_nb_colors", type=int, default=5)
136 parser.add_argument("--snake_length", type=int, default=200)
138 ##############################
141 parser.add_argument("--stack_nb_steps", type=int, default=100)
143 parser.add_argument("--stack_nb_stacks", type=int, default=3)
145 parser.add_argument("--stack_nb_digits", type=int, default=3)
147 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
149 ##############################
152 parser.add_argument("--expr_nb_variables", type=int, default=5)
154 parser.add_argument("--expr_sequence_length", type=int, default=40)
156 parser.add_argument("--expr_operand_max", type=int, default=9)
158 parser.add_argument("--expr_result_max", type=int, default=99)
160 parser.add_argument("--expr_input_file", type=str, default=None)
162 ##############################
165 parser.add_argument("--mixing_hard", action="store_true", default=False)
167 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
169 ######################################################################
171 args = parser.parse_args()
173 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
175 if args.result_dir is None:
176 args.result_dir = f"results_{args.task}"
178 ######################################################################
180 default_task_args = {
184 "nb_train_samples": 250000,
185 "nb_test_samples": 10000,
190 "nb_train_samples": 50000,
191 "nb_test_samples": 10000,
196 "nb_train_samples": 2500000,
197 "nb_test_samples": 10000,
202 "nb_train_samples": 250000,
203 "nb_test_samples": 10000,
208 "nb_train_samples": 100000,
209 "nb_test_samples": 1000,
214 "nb_train_samples": 1000000,
215 "nb_test_samples": 10000,
220 "nb_train_samples": 50000,
221 "nb_test_samples": 10000,
226 "nb_train_samples": 100000,
227 "nb_test_samples": 10000,
232 "nb_train_samples": 250000,
233 "nb_test_samples": 10000,
238 "nb_train_samples": 2500000,
239 "nb_test_samples": 10000,
244 "nb_train_samples": 250000,
245 "nb_test_samples": 10000,
250 "nb_train_samples": 100000,
251 "nb_test_samples": 1000,
256 "nb_train_samples": 50000,
257 "nb_test_samples": 10000,
262 "nb_train_samples": 250000,
263 "nb_test_samples": 10000,
268 "nb_train_samples": 60000,
269 "nb_test_samples": 10000,
273 if args.task in default_task_args:
274 for k, v in default_task_args[args.task].items():
275 if getattr(args, k) is None:
278 ######################################################################
280 default_model_args = {
311 if args.model in default_model_args:
312 for k, v in default_model_args[args.model].items():
313 if getattr(args, k) is None:
316 raise ValueError(f"Unknown model {args.model}")
318 ######################################################################
321 os.mkdir(args.result_dir)
322 except FileExistsError:
323 if not args.overwrite_results:
324 print(f"result directory {args.result_dir} already exists")
327 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
330 # torch.backends.cudnn.deterministic = True
331 # torch.backends.cudnn.benchmark = False
332 # torch.use_deterministic_algorithms(True)
333 torch.manual_seed(args.seed)
334 if torch.cuda.is_available():
335 torch.cuda.manual_seed_all(args.seed)
337 ######################################################################
341 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
343 if log_file is not None:
344 log_file.write(t + s + "\n")
351 log_string(f"argv {' '.join(sys.argv)}")
354 log_string(f"args.{n} {getattr(args, n)}")
357 ######################################################################
360 def picoclvr_pruner_horizontal_green(p):
361 return not ("green" in p and ("left" in p or "right" in p))
364 picoclvr_pruner_train = (
365 picoclvr_pruner_horizontal_green
366 if args.picocvlr_prune_properties in {"train+eval"}
370 picoclvr_pruner_eval = (
371 (lambda p: not picoclvr_pruner_horizontal_green(p))
372 if args.picocvlr_prune_properties in {"train+eval", "eval"}
376 ######################################################################
378 if args.task == "byheart":
379 task = tasks.SandBox(
380 problem=problems.ProblemByHeart(),
381 nb_train_samples=args.nb_train_samples,
382 nb_test_samples=args.nb_test_samples,
383 batch_size=args.batch_size,
387 args.max_percents_of_test_in_train = -1
389 elif args.task == "learnop":
390 task = tasks.SandBox(
391 problem=problems.ProblemLearnOperator(),
392 nb_train_samples=args.nb_train_samples,
393 nb_test_samples=args.nb_test_samples,
394 batch_size=args.batch_size,
400 elif args.task == "guessop":
401 task = tasks.SandBox(
402 problem=problems.ProblemGuessOperator(),
403 nb_train_samples=args.nb_train_samples,
404 nb_test_samples=args.nb_test_samples,
405 batch_size=args.batch_size,
411 elif args.task == "twotargets":
412 task = tasks.SandBox(
413 problem=problems.ProblemTwoTargets(),
414 nb_train_samples=args.nb_train_samples,
415 nb_test_samples=args.nb_test_samples,
416 batch_size=args.batch_size,
421 elif args.task == "mixing":
422 task = tasks.SandBox(
423 problem=problems.ProblemMixing(
424 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
426 nb_train_samples=args.nb_train_samples,
427 nb_test_samples=args.nb_test_samples,
428 batch_size=args.batch_size,
433 elif args.task == "addition":
434 task = tasks.SandBox(
435 problem=problems.ProblemAddition(),
436 nb_train_samples=args.nb_train_samples,
437 nb_test_samples=args.nb_test_samples,
438 batch_size=args.batch_size,
443 elif args.task == "picoclvr":
444 task = tasks.PicoCLVR(
445 nb_train_samples=args.nb_train_samples,
446 nb_test_samples=args.nb_test_samples,
447 batch_size=args.batch_size,
448 height=args.picoclvr_height,
449 width=args.picoclvr_width,
450 nb_colors=args.picoclvr_nb_colors,
453 pruner_train=picoclvr_pruner_train,
454 pruner_eval=picoclvr_pruner_eval,
457 elif args.task == "mnist":
459 nb_train_samples=args.nb_train_samples,
460 nb_test_samples=args.nb_test_samples,
461 batch_size=args.batch_size,
465 elif args.task == "maze":
467 nb_train_samples=args.nb_train_samples,
468 nb_test_samples=args.nb_test_samples,
469 batch_size=args.batch_size,
470 height=args.maze_height,
471 width=args.maze_width,
472 nb_walls=args.maze_nb_walls,
476 elif args.task == "snake":
478 nb_train_samples=args.nb_train_samples,
479 nb_test_samples=args.nb_test_samples,
480 batch_size=args.batch_size,
481 height=args.snake_height,
482 width=args.snake_width,
483 nb_colors=args.snake_nb_colors,
484 length=args.snake_length,
485 prompt_length=args.snake_length // 2,
489 elif args.task == "stack":
491 nb_train_samples=args.nb_train_samples,
492 nb_test_samples=args.nb_test_samples,
493 batch_size=args.batch_size,
495 nb_steps=args.stack_nb_steps,
496 nb_stacks=args.stack_nb_stacks,
497 nb_digits=args.stack_nb_digits,
498 fraction_values_for_train=args.stack_fraction_values_for_train,
502 elif args.task == "expr":
504 nb_train_samples=args.nb_train_samples,
505 nb_test_samples=args.nb_test_samples,
506 nb_variables=args.expr_nb_variables,
507 sequence_length=args.expr_sequence_length,
508 operand_max=args.expr_operand_max,
509 result_max=args.expr_result_max,
510 batch_size=args.batch_size,
514 elif args.task == "rpl":
516 nb_train_samples=args.nb_train_samples,
517 nb_test_samples=args.nb_test_samples,
518 batch_size=args.batch_size,
519 nb_starting_values=args.rpl_nb_starting_values,
520 max_input=args.rpl_max_input,
521 prog_len=args.rpl_prog_len,
522 nb_runs=args.rpl_nb_runs,
523 no_prog=args.rpl_no_prog,
528 elif args.task == "grid":
530 nb_train_samples=args.nb_train_samples,
531 nb_test_samples=args.nb_test_samples,
532 batch_size=args.batch_size,
538 elif args.task == "qmlp":
540 nb_train_samples=args.nb_train_samples,
541 nb_test_samples=args.nb_test_samples,
542 batch_size=args.batch_size,
543 result_dir=args.result_dir,
549 raise ValueError(f"Unknown task {args.task}")
551 ######################################################################
553 log_string(f"device {device}")
555 vocabulary_size = task.vocabulary_size()
557 log_string(f"vocabulary_size {vocabulary_size}")
559 ##############################
562 vocabulary_size=vocabulary_size,
563 dim_model=args.dim_model,
564 dim_keys=args.dim_keys,
565 dim_hidden=args.dim_hidden,
566 nb_heads=args.nb_heads,
567 nb_blocks=args.nb_blocks,
569 dropout=args.dropout,
574 nb_parameters = sum(p.numel() for p in model.parameters())
575 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
577 ######################################################################
579 nb_epochs_finished = 0
581 if args.no_checkpoint:
582 log_string(f"not trying to load checkpoint.")
586 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
587 checkpoint = torch.load(checkpoint_name)
588 nb_epochs_finished = checkpoint["nb_epochs_finished"]
589 model.load_state_dict(checkpoint["model_state"])
590 torch.set_rng_state(checkpoint["rng_state"])
591 if torch.cuda.is_available():
592 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
594 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
596 except FileNotFoundError:
597 log_string("starting from scratch.")
600 log_string("error when loading the checkpoint.")
603 ######################################################################
605 if args.task == "expr" and args.expr_input_file is not None:
606 task.produce_results(
607 n_epoch=nb_epochs_finished,
609 result_dir=args.result_dir,
611 deterministic_synthesis=args.deterministic_synthesis,
612 input_file=args.expr_input_file,
617 ######################################################################
619 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
621 # Compute the entropy of the training tokens
624 for input in task.batches(split="train"):
625 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
626 token_probas = token_count / token_count.sum()
627 entropy = -torch.xlogy(token_probas, token_probas).sum()
628 train_set_perplexity = math.exp(entropy)
630 ######################################################################
631 # A bit of paranoia never hurts
633 if args.max_percents_of_test_in_train >= 0:
635 def subsets_as_tuples(batches, cs):
637 for batch in batches:
639 s.add(tuple([v.item() for v in x]))
645 nb_test, nb_in_train = 0, 0
646 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
648 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
649 in_train.update(test_subset.intersection(train_subset))
650 nb_in_train += len(in_train)
651 nb_test += len(test_subset)
654 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
658 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
659 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
661 ##############################
663 if args.learning_rate_schedule == "cos":
664 learning_rate_schedule = {}
665 for n_epoch in range(args.nb_epochs):
666 u = n_epoch / args.nb_epochs * math.pi
667 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
672 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
676 learning_rate_schedule = {}
677 learning_rate = args.learning_rate
678 for n_epoch in range(args.nb_epochs):
680 learning_rate = u[n_epoch]
681 learning_rate_schedule[n_epoch] = learning_rate
683 log_string(f"learning_rate_schedule {learning_rate_schedule}")
685 ##############################
689 if nb_epochs_finished >= nb_epochs:
690 task.produce_results(
691 n_epoch=nb_epochs_finished,
693 result_dir=args.result_dir,
695 deterministic_synthesis=args.deterministic_synthesis,
698 for n_epoch in range(nb_epochs_finished, nb_epochs):
699 learning_rate = learning_rate_schedule[n_epoch]
701 log_string(f"learning_rate {learning_rate}")
703 if args.optim == "sgd":
704 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
705 elif args.optim == "adam":
706 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
707 elif args.optim == "adamw":
708 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
710 raise ValueError(f"Unknown optimizer {args.optim}.")
714 nb_train_samples, acc_train_loss = 0, 0.0
716 for input in task.batches(split="train"):
717 input = input.to(device)
718 output = model(mygpt.BracketedSequence(input)).x
719 loss = F.cross_entropy(output.transpose(1, 2), input)
720 acc_train_loss += loss.item() * input.size(0)
721 nb_train_samples += input.size(0)
722 nb_samples_seen += input.size(0)
724 optimizer.zero_grad()
728 with torch.autograd.no_grad():
731 nb_test_samples, acc_test_loss = 0, 0.0
733 for input in task.batches(split="test"):
734 input = input.to(device)
736 output = model(mygpt.BracketedSequence(input)).x
737 loss = F.cross_entropy(output.transpose(1, 2), input)
738 acc_test_loss += loss.item() * input.size(0)
739 nb_test_samples += input.size(0)
741 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
742 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
745 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
748 task.produce_results(
751 result_dir=args.result_dir,
753 deterministic_synthesis=args.deterministic_synthesis,
757 "nb_epochs_finished": n_epoch + 1,
758 "model_state": model.state_dict(),
759 "rng_state": torch.get_rng_state(),
762 if torch.cuda.is_available():
763 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
765 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
766 torch.save(checkpoint, checkpoint_name)
767 log_string(f"saved checkpoint {checkpoint_name}")
769 ######################################################################