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, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid",
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=None)
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("--picoclvr_nb_colors", type=int, default=5)
107 parser.add_argument("--picoclvr_height", type=int, default=12)
109 parser.add_argument("--picoclvr_width", type=int, default=16)
111 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
113 ##############################
116 parser.add_argument("--maze_height", type=int, default=13)
118 parser.add_argument("--maze_width", type=int, default=21)
120 parser.add_argument("--maze_nb_walls", type=int, default=15)
122 ##############################
125 parser.add_argument("--snake_height", type=int, default=9)
127 parser.add_argument("--snake_width", type=int, default=12)
129 parser.add_argument("--snake_nb_colors", type=int, default=5)
131 parser.add_argument("--snake_length", type=int, default=200)
133 ##############################
136 parser.add_argument("--stack_nb_steps", type=int, default=100)
138 parser.add_argument("--stack_nb_stacks", type=int, default=3)
140 parser.add_argument("--stack_nb_digits", type=int, default=3)
142 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
144 ##############################
147 parser.add_argument("--expr_nb_variables", type=int, default=5)
149 parser.add_argument("--expr_sequence_length", type=int, default=40)
151 parser.add_argument("--expr_operand_max", type=int, default=9)
153 parser.add_argument("--expr_result_max", type=int, default=99)
155 parser.add_argument("--expr_input_file", type=str, default=None)
157 ##############################
160 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
162 ######################################################################
164 args = parser.parse_args()
166 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
168 if args.result_dir is None:
169 args.result_dir = f"results_{args.task}"
171 ######################################################################
173 default_task_args = {
178 "nb_train_samples": 50000,
179 "nb_test_samples": 10000,
185 "nb_train_samples": 50000,
186 "nb_test_samples": 10000,
192 "nb_train_samples": 1000000,
193 "nb_test_samples": 10000,
199 "nb_train_samples": 50000,
200 "nb_test_samples": 10000,
206 "nb_train_samples": 250000,
207 "nb_test_samples": 10000,
213 "nb_train_samples": 250000,
214 "nb_test_samples": 10000,
220 "nb_train_samples": 60000,
221 "nb_test_samples": 10000,
227 "nb_train_samples": 100000,
228 "nb_test_samples": 10000,
234 "nb_train_samples": 250000,
235 "nb_test_samples": 10000,
241 "nb_train_samples": 100000,
242 "nb_test_samples": 1000,
248 "nb_train_samples": 2500000,
249 "nb_test_samples": 10000,
255 "nb_train_samples": 1000000,
256 "nb_test_samples": 10000,
262 "nb_train_samples": 25000,
263 "nb_test_samples": 1000,
269 "nb_train_samples": 250000,
270 "nb_test_samples": 10000,
274 if args.task in default_task_args:
275 for k, v in default_task_args[args.task].items():
276 if getattr(args, k) is None:
279 ######################################################################
281 default_model_args = {
312 if args.model in default_model_args:
313 for k, v in default_model_args[args.model].items():
314 if getattr(args, k) is None:
317 raise ValueError(f"Unknown model {args.model}")
319 ######################################################################
322 os.mkdir(args.result_dir)
323 except FileExistsError:
324 if not args.overwrite_results:
325 print(f"result directory {args.result_dir} already exists")
328 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
331 # torch.backends.cudnn.deterministic = True
332 # torch.backends.cudnn.benchmark = False
333 # torch.use_deterministic_algorithms(True)
334 torch.manual_seed(args.seed)
335 if torch.cuda.is_available():
336 torch.cuda.manual_seed_all(args.seed)
338 ######################################################################
342 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
344 if log_file is not None:
345 log_file.write(t + s + "\n")
353 log_string(f"args.{n} {getattr(args, n)}")
356 ######################################################################
359 def picoclvr_pruner_horizontal_green(p):
360 return not ("green" in p and ("left" in p or "right" in p))
363 picoclvr_pruner_train = (
364 picoclvr_pruner_horizontal_green
365 if args.picocvlr_prune_properties in {"train+eval"}
369 picoclvr_pruner_eval = (
370 (lambda p: not picoclvr_pruner_horizontal_green(p))
371 if args.picocvlr_prune_properties in {"train+eval", "eval"}
375 ######################################################################
377 if args.task == "byheart":
378 task = tasks.SandBox(
379 problem=problems.ProblemByHeart(),
380 nb_train_samples=args.nb_train_samples,
381 nb_test_samples=args.nb_test_samples,
382 batch_size=args.batch_size,
386 args.max_percents_of_test_in_train = -1
388 elif args.task == "learnop":
389 task = tasks.SandBox(
390 problem=problems.ProblemLearnOperator(),
391 nb_train_samples=args.nb_train_samples,
392 nb_test_samples=args.nb_test_samples,
393 batch_size=args.batch_size,
399 elif args.task == "guessop":
400 task = tasks.SandBox(
401 problem=problems.ProblemGuessOperator(),
402 nb_train_samples=args.nb_train_samples,
403 nb_test_samples=args.nb_test_samples,
404 batch_size=args.batch_size,
410 elif args.task == "twotargets":
411 task = tasks.SandBox(
412 problem=problems.ProblemTwoTargets(),
413 nb_train_samples=args.nb_train_samples,
414 nb_test_samples=args.nb_test_samples,
415 batch_size=args.batch_size,
420 elif args.task == "addition":
421 task = tasks.SandBox(
422 problem=problems.ProblemAddition(),
423 nb_train_samples=args.nb_train_samples,
424 nb_test_samples=args.nb_test_samples,
425 batch_size=args.batch_size,
430 elif args.task == "picoclvr":
431 task = tasks.PicoCLVR(
432 nb_train_samples=args.nb_train_samples,
433 nb_test_samples=args.nb_test_samples,
434 batch_size=args.batch_size,
435 height=args.picoclvr_height,
436 width=args.picoclvr_width,
437 nb_colors=args.picoclvr_nb_colors,
440 pruner_train=picoclvr_pruner_train,
441 pruner_eval=picoclvr_pruner_eval,
444 elif args.task == "mnist":
446 nb_train_samples=args.nb_train_samples,
447 nb_test_samples=args.nb_test_samples,
448 batch_size=args.batch_size,
452 elif args.task == "maze":
454 nb_train_samples=args.nb_train_samples,
455 nb_test_samples=args.nb_test_samples,
456 batch_size=args.batch_size,
457 height=args.maze_height,
458 width=args.maze_width,
459 nb_walls=args.maze_nb_walls,
463 elif args.task == "snake":
465 nb_train_samples=args.nb_train_samples,
466 nb_test_samples=args.nb_test_samples,
467 batch_size=args.batch_size,
468 height=args.snake_height,
469 width=args.snake_width,
470 nb_colors=args.snake_nb_colors,
471 length=args.snake_length,
472 prompt_length=args.snake_length // 2,
476 elif args.task == "stack":
478 nb_train_samples=args.nb_train_samples,
479 nb_test_samples=args.nb_test_samples,
480 batch_size=args.batch_size,
482 nb_steps=args.stack_nb_steps,
483 nb_stacks=args.stack_nb_stacks,
484 nb_digits=args.stack_nb_digits,
485 fraction_values_for_train=args.stack_fraction_values_for_train,
489 elif args.task == "expr":
491 nb_train_samples=args.nb_train_samples,
492 nb_test_samples=args.nb_test_samples,
493 nb_variables=args.expr_nb_variables,
494 sequence_length=args.expr_sequence_length,
495 operand_max=args.expr_operand_max,
496 result_max=args.expr_result_max,
497 batch_size=args.batch_size,
501 elif args.task == "rpl":
503 nb_train_samples=args.nb_train_samples,
504 nb_test_samples=args.nb_test_samples,
505 batch_size=args.batch_size,
506 nb_starting_values=args.rpl_nb_starting_values,
507 max_input=args.rpl_max_input,
508 prog_len=args.rpl_prog_len,
509 nb_runs=args.rpl_nb_runs,
510 no_prog=args.rpl_no_prog,
515 elif args.task == "grid":
517 nb_train_samples=args.nb_train_samples,
518 nb_test_samples=args.nb_test_samples,
519 batch_size=args.batch_size,
520 height=args.picoclvr_height,
521 width=args.picoclvr_width,
526 elif args.task == "world":
528 nb_train_samples=args.nb_train_samples,
529 nb_test_samples=args.nb_test_samples,
530 batch_size=args.batch_size,
531 vqae_nb_epochs=args.world_vqae_nb_epochs,
537 raise ValueError(f"Unknown task {args.task}")
539 ######################################################################
541 log_string(f"device {device}")
543 vocabulary_size = task.vocabulary_size()
545 log_string(f"vocabulary_size {vocabulary_size}")
547 ##############################
550 vocabulary_size=vocabulary_size,
551 dim_model=args.dim_model,
552 dim_keys=args.dim_keys,
553 dim_hidden=args.dim_hidden,
554 nb_heads=args.nb_heads,
555 nb_blocks=args.nb_blocks,
557 dropout=args.dropout,
562 nb_parameters = sum(p.numel() for p in model.parameters())
563 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
565 ######################################################################
567 nb_epochs_finished = 0
569 if args.no_checkpoint:
570 log_string(f"not trying to load checkpoint.")
574 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
575 checkpoint = torch.load(checkpoint_name)
576 nb_epochs_finished = checkpoint["nb_epochs_finished"]
577 model.load_state_dict(checkpoint["model_state"])
578 torch.set_rng_state(checkpoint["rng_state"])
579 if torch.cuda.is_available():
580 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
582 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
584 except FileNotFoundError:
585 log_string("starting from scratch.")
588 log_string("error when loading the checkpoint.")
591 ######################################################################
593 if args.task == "expr" and args.expr_input_file is not None:
594 task.produce_results(
595 n_epoch=nb_epochs_finished,
597 result_dir=args.result_dir,
599 deterministic_synthesis=args.deterministic_synthesis,
600 input_file=args.expr_input_file,
605 ######################################################################
607 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
609 # Compute the entropy of the training tokens
612 for input in task.batches(split="train"):
613 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
614 token_probas = token_count / token_count.sum()
615 entropy = -torch.xlogy(token_probas, token_probas).sum()
616 train_set_perplexity = math.exp(entropy)
618 ######################################################################
619 # A bit of paranoia never hurts
621 if args.max_percents_of_test_in_train >= 0:
623 def subsets_as_tuples(batches, cs):
625 for batch in batches:
627 s.add(tuple([v.item() for v in x]))
633 nb_test, nb_in_train = 0, 0
634 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
636 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
637 in_train.update(test_subset.intersection(train_subset))
638 nb_in_train += len(in_train)
639 nb_test += len(test_subset)
642 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
646 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
647 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
649 ##############################
651 if args.learning_rate_schedule == "cos":
652 learning_rate_schedule = {}
653 for n_epoch in range(args.nb_epochs):
654 u = n_epoch / args.nb_epochs * math.pi
655 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
660 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
664 learning_rate_schedule = {}
665 learning_rate = args.learning_rate
666 for n_epoch in range(args.nb_epochs):
668 learning_rate = u[n_epoch]
669 learning_rate_schedule[n_epoch] = learning_rate
671 log_string(f"learning_rate_schedule {learning_rate_schedule}")
673 ##############################
677 if nb_epochs_finished >= nb_epochs:
678 task.produce_results(
679 n_epoch=nb_epochs_finished,
681 result_dir=args.result_dir,
683 deterministic_synthesis=args.deterministic_synthesis,
686 for n_epoch in range(nb_epochs_finished, nb_epochs):
687 learning_rate = learning_rate_schedule[n_epoch]
689 log_string(f"learning_rate {learning_rate}")
691 if args.optim == "sgd":
692 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
693 elif args.optim == "adam":
694 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
695 elif args.optim == "adamw":
696 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
698 raise ValueError(f"Unknown optimizer {args.optim}.")
702 nb_train_samples, acc_train_loss = 0, 0.0
704 for input in task.batches(split="train"):
705 input = input.to(device)
706 output = model(mygpt.BracketedSequence(input)).x
707 loss = F.cross_entropy(output.transpose(1, 2), input)
708 acc_train_loss += loss.item() * input.size(0)
709 nb_train_samples += input.size(0)
710 nb_samples_seen += input.size(0)
712 optimizer.zero_grad()
716 with torch.autograd.no_grad():
719 nb_test_samples, acc_test_loss = 0, 0.0
721 for input in task.batches(split="test"):
722 input = input.to(device)
724 output = model(mygpt.BracketedSequence(input)).x
725 loss = F.cross_entropy(output.transpose(1, 2), input)
726 acc_test_loss += loss.item() * input.size(0)
727 nb_test_samples += input.size(0)
729 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
730 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
733 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
736 task.produce_results(
739 result_dir=args.result_dir,
741 deterministic_synthesis=args.deterministic_synthesis,
745 "nb_epochs_finished": n_epoch + 1,
746 "model_state": model.state_dict(),
747 "rng_state": torch.get_rng_state(),
750 if torch.cuda.is_available():
751 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
753 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
754 torch.save(checkpoint, checkpoint_name)
755 log_string(f"saved checkpoint {checkpoint_name}")
757 ######################################################################