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="byheart, learnop, guessop, mixing, memory, 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 parser.add_argument("--grid_fraction_play", type=float, default=0)
109 ##############################
112 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
114 parser.add_argument("--picoclvr_height", type=int, default=12)
116 parser.add_argument("--picoclvr_width", type=int, default=16)
118 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
120 ##############################
123 parser.add_argument("--maze_height", type=int, default=13)
125 parser.add_argument("--maze_width", type=int, default=21)
127 parser.add_argument("--maze_nb_walls", type=int, default=15)
129 ##############################
132 parser.add_argument("--snake_height", type=int, default=9)
134 parser.add_argument("--snake_width", type=int, default=12)
136 parser.add_argument("--snake_nb_colors", type=int, default=5)
138 parser.add_argument("--snake_length", type=int, default=200)
140 ##############################
143 parser.add_argument("--stack_nb_steps", type=int, default=100)
145 parser.add_argument("--stack_nb_stacks", type=int, default=3)
147 parser.add_argument("--stack_nb_digits", type=int, default=3)
149 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
151 ##############################
154 parser.add_argument("--expr_nb_variables", type=int, default=5)
156 parser.add_argument("--expr_sequence_length", type=int, default=40)
158 parser.add_argument("--expr_operand_max", type=int, default=9)
160 parser.add_argument("--expr_result_max", type=int, default=99)
162 parser.add_argument("--expr_input_file", type=str, default=None)
164 ##############################
167 parser.add_argument("--mixing_hard", action="store_true", default=False)
169 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
171 ######################################################################
173 args = parser.parse_args()
175 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
177 if args.result_dir is None:
178 args.result_dir = f"results_{args.task}"
180 ######################################################################
182 default_task_args = {
186 "nb_train_samples": 250000,
187 "nb_test_samples": 10000,
192 "nb_train_samples": 50000,
193 "nb_test_samples": 10000,
198 "nb_train_samples": 2500000,
199 "nb_test_samples": 10000,
204 "nb_train_samples": 250000,
205 "nb_test_samples": 10000,
210 "nb_train_samples": 100000,
211 "nb_test_samples": 1000,
216 "nb_train_samples": 1000000,
217 "nb_test_samples": 10000,
222 "nb_train_samples": 50000,
223 "nb_test_samples": 10000,
228 "nb_train_samples": 100000,
229 "nb_test_samples": 10000,
234 "nb_train_samples": 250000,
235 "nb_test_samples": 10000,
240 "nb_train_samples": 2500000,
241 "nb_test_samples": 10000,
246 "nb_train_samples": 250000,
247 "nb_test_samples": 10000,
252 "nb_train_samples": 100000,
253 "nb_test_samples": 1000,
258 "nb_train_samples": 50000,
259 "nb_test_samples": 10000,
264 "nb_train_samples": 25000,
265 "nb_test_samples": 1000,
270 "nb_train_samples": 250000,
271 "nb_test_samples": 10000,
276 "nb_train_samples": 60000,
277 "nb_test_samples": 10000,
281 if args.task in default_task_args:
282 for k, v in default_task_args[args.task].items():
283 if getattr(args, k) is None:
286 ######################################################################
288 default_model_args = {
326 if args.model in default_model_args:
327 for k, v in default_model_args[args.model].items():
328 if getattr(args, k) is None:
331 raise ValueError(f"Unknown model {args.model}")
333 ######################################################################
336 os.mkdir(args.result_dir)
337 except FileExistsError:
338 if not args.overwrite_results:
339 print(f"result directory {args.result_dir} already exists")
342 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
345 # torch.backends.cudnn.deterministic = True
346 # torch.backends.cudnn.benchmark = False
347 # torch.use_deterministic_algorithms(True)
348 torch.manual_seed(args.seed)
349 if torch.cuda.is_available():
350 torch.cuda.manual_seed_all(args.seed)
352 ######################################################################
356 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
358 if log_file is not None:
359 log_file.write(t + s + "\n")
366 log_string(f"argv {' '.join(sys.argv)}")
369 log_string(f"args.{n} {getattr(args, n)}")
372 ######################################################################
375 def picoclvr_pruner_horizontal_green(p):
376 return not ("green" in p and ("left" in p or "right" in p))
379 picoclvr_pruner_train = (
380 picoclvr_pruner_horizontal_green
381 if args.picocvlr_prune_properties in {"train+eval"}
385 picoclvr_pruner_eval = (
386 (lambda p: not picoclvr_pruner_horizontal_green(p))
387 if args.picocvlr_prune_properties in {"train+eval", "eval"}
391 ######################################################################
393 if args.task == "byheart":
394 task = tasks.SandBox(
395 problem=problems.ProblemByHeart(),
396 nb_train_samples=args.nb_train_samples,
397 nb_test_samples=args.nb_test_samples,
398 batch_size=args.batch_size,
402 args.max_percents_of_test_in_train = -1
404 elif args.task == "learnop":
405 task = tasks.SandBox(
406 problem=problems.ProblemLearnOperator(),
407 nb_train_samples=args.nb_train_samples,
408 nb_test_samples=args.nb_test_samples,
409 batch_size=args.batch_size,
415 elif args.task == "guessop":
416 task = tasks.SandBox(
417 problem=problems.ProblemGuessOperator(),
418 nb_train_samples=args.nb_train_samples,
419 nb_test_samples=args.nb_test_samples,
420 batch_size=args.batch_size,
426 elif args.task == "twotargets":
427 task = tasks.SandBox(
428 problem=problems.ProblemTwoTargets(),
429 nb_train_samples=args.nb_train_samples,
430 nb_test_samples=args.nb_test_samples,
431 batch_size=args.batch_size,
436 elif args.task == "memory":
437 task = tasks.SandBox(
438 problem=problems.ProblemMemory(),
439 nb_train_samples=args.nb_train_samples,
440 nb_test_samples=args.nb_test_samples,
441 batch_size=args.batch_size,
446 elif args.task == "mixing":
447 task = tasks.SandBox(
448 problem=problems.ProblemMixing(
449 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
451 nb_train_samples=args.nb_train_samples,
452 nb_test_samples=args.nb_test_samples,
453 batch_size=args.batch_size,
458 elif args.task == "addition":
459 task = tasks.SandBox(
460 problem=problems.ProblemAddition(),
461 nb_train_samples=args.nb_train_samples,
462 nb_test_samples=args.nb_test_samples,
463 batch_size=args.batch_size,
468 elif args.task == "picoclvr":
469 task = tasks.PicoCLVR(
470 nb_train_samples=args.nb_train_samples,
471 nb_test_samples=args.nb_test_samples,
472 batch_size=args.batch_size,
473 height=args.picoclvr_height,
474 width=args.picoclvr_width,
475 nb_colors=args.picoclvr_nb_colors,
478 pruner_train=picoclvr_pruner_train,
479 pruner_eval=picoclvr_pruner_eval,
482 elif args.task == "mnist":
484 nb_train_samples=args.nb_train_samples,
485 nb_test_samples=args.nb_test_samples,
486 batch_size=args.batch_size,
490 elif args.task == "maze":
492 nb_train_samples=args.nb_train_samples,
493 nb_test_samples=args.nb_test_samples,
494 batch_size=args.batch_size,
495 height=args.maze_height,
496 width=args.maze_width,
497 nb_walls=args.maze_nb_walls,
501 elif args.task == "snake":
503 nb_train_samples=args.nb_train_samples,
504 nb_test_samples=args.nb_test_samples,
505 batch_size=args.batch_size,
506 height=args.snake_height,
507 width=args.snake_width,
508 nb_colors=args.snake_nb_colors,
509 length=args.snake_length,
510 prompt_length=args.snake_length // 2,
514 elif args.task == "stack":
516 nb_train_samples=args.nb_train_samples,
517 nb_test_samples=args.nb_test_samples,
518 batch_size=args.batch_size,
520 nb_steps=args.stack_nb_steps,
521 nb_stacks=args.stack_nb_stacks,
522 nb_digits=args.stack_nb_digits,
523 fraction_values_for_train=args.stack_fraction_values_for_train,
527 elif args.task == "expr":
529 nb_train_samples=args.nb_train_samples,
530 nb_test_samples=args.nb_test_samples,
531 nb_variables=args.expr_nb_variables,
532 sequence_length=args.expr_sequence_length,
533 operand_max=args.expr_operand_max,
534 result_max=args.expr_result_max,
535 batch_size=args.batch_size,
539 elif args.task == "rpl":
541 nb_train_samples=args.nb_train_samples,
542 nb_test_samples=args.nb_test_samples,
543 batch_size=args.batch_size,
544 nb_starting_values=args.rpl_nb_starting_values,
545 max_input=args.rpl_max_input,
546 prog_len=args.rpl_prog_len,
547 nb_runs=args.rpl_nb_runs,
548 no_prog=args.rpl_no_prog,
553 elif args.task == "grid":
555 nb_train_samples=args.nb_train_samples,
556 nb_test_samples=args.nb_test_samples,
557 batch_size=args.batch_size,
559 fraction_play=args.grid_fraction_play,
564 elif args.task == "qmlp":
566 nb_train_samples=args.nb_train_samples,
567 nb_test_samples=args.nb_test_samples,
568 batch_size=args.batch_size,
569 result_dir=args.result_dir,
575 raise ValueError(f"Unknown task {args.task}")
577 ######################################################################
579 log_string(f"device {device}")
581 vocabulary_size = task.vocabulary_size()
583 log_string(f"vocabulary_size {vocabulary_size}")
585 ##############################
588 vocabulary_size=vocabulary_size,
589 dim_model=args.dim_model,
590 dim_keys=args.dim_keys,
591 dim_hidden=args.dim_hidden,
592 nb_heads=args.nb_heads,
593 nb_blocks=args.nb_blocks,
595 dropout=args.dropout,
600 nb_parameters = sum(p.numel() for p in model.parameters())
601 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
603 ######################################################################
605 nb_epochs_finished = 0
607 if args.no_checkpoint:
608 log_string(f"not trying to load checkpoint.")
612 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
613 checkpoint = torch.load(checkpoint_name)
614 nb_epochs_finished = checkpoint["nb_epochs_finished"]
615 model.load_state_dict(checkpoint["model_state"])
616 torch.set_rng_state(checkpoint["rng_state"])
617 if torch.cuda.is_available():
618 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
620 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
622 except FileNotFoundError:
623 log_string("starting from scratch.")
626 log_string("error when loading the checkpoint.")
629 ######################################################################
631 if args.task == "expr" and args.expr_input_file is not None:
632 task.produce_results(
633 n_epoch=nb_epochs_finished,
635 result_dir=args.result_dir,
637 deterministic_synthesis=args.deterministic_synthesis,
638 input_file=args.expr_input_file,
643 ######################################################################
645 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
647 # Compute the entropy of the training tokens
650 for input in task.batches(split="train"):
651 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
652 token_probas = token_count / token_count.sum()
653 entropy = -torch.xlogy(token_probas, token_probas).sum()
654 train_set_perplexity = math.exp(entropy)
656 ######################################################################
657 # A bit of paranoia never hurts
659 if args.max_percents_of_test_in_train >= 0:
661 def subsets_as_tuples(batches, cs):
663 for batch in batches:
665 s.add(tuple([v.item() for v in x]))
671 nb_test, nb_in_train = 0, 0
672 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
674 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
675 in_train.update(test_subset.intersection(train_subset))
676 nb_in_train += len(in_train)
677 nb_test += len(test_subset)
680 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
684 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
685 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
687 ##############################
689 if args.learning_rate_schedule == "cos":
690 learning_rate_schedule = {}
691 for n_epoch in range(args.nb_epochs):
692 u = n_epoch / args.nb_epochs * math.pi
693 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
698 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
702 learning_rate_schedule = {}
703 learning_rate = args.learning_rate
704 for n_epoch in range(args.nb_epochs):
706 learning_rate = u[n_epoch]
707 learning_rate_schedule[n_epoch] = learning_rate
709 log_string(f"learning_rate_schedule {learning_rate_schedule}")
711 ##############################
715 if nb_epochs_finished >= nb_epochs:
716 task.produce_results(
717 n_epoch=nb_epochs_finished,
719 result_dir=args.result_dir,
721 deterministic_synthesis=args.deterministic_synthesis,
724 time_pred_result = None
726 for n_epoch in range(nb_epochs_finished, nb_epochs):
727 learning_rate = learning_rate_schedule[n_epoch]
729 log_string(f"learning_rate {learning_rate}")
731 if args.optim == "sgd":
732 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
733 elif args.optim == "adam":
734 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
735 elif args.optim == "adamw":
736 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
738 raise ValueError(f"Unknown optimizer {args.optim}.")
742 nb_train_samples, acc_train_loss = 0, 0.0
744 for input in task.batches(split="train"):
745 input = input.to(device)
746 output = model(mygpt.BracketedSequence(input)).x
747 loss = F.cross_entropy(output.transpose(1, 2), input)
748 acc_train_loss += loss.item() * input.size(0)
749 nb_train_samples += input.size(0)
750 nb_samples_seen += input.size(0)
752 optimizer.zero_grad()
756 with torch.autograd.no_grad():
759 nb_test_samples, acc_test_loss = 0, 0.0
761 for input in task.batches(split="test"):
762 input = input.to(device)
764 output = model(mygpt.BracketedSequence(input)).x
765 loss = F.cross_entropy(output.transpose(1, 2), input)
766 acc_test_loss += loss.item() * input.size(0)
767 nb_test_samples += input.size(0)
769 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
770 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
773 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
776 task.produce_results(
779 result_dir=args.result_dir,
781 deterministic_synthesis=args.deterministic_synthesis,
784 time_current_result = datetime.datetime.now()
785 if time_pred_result is not None:
787 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
789 time_pred_result = time_current_result
792 "nb_epochs_finished": n_epoch + 1,
793 "model_state": model.state_dict(),
794 "rng_state": torch.get_rng_state(),
797 if torch.cuda.is_available():
798 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
800 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
801 torch.save(checkpoint, checkpoint_name)
802 log_string(f"saved checkpoint {checkpoint_name}")
804 ######################################################################