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",
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=5)
94 parser.add_argument("--rpl_max_input", type=int, default=9)
96 parser.add_argument("--rpl_prog_len", type=int, default=10)
98 parser.add_argument("--rpl_nb_runs", type=int, default=8)
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=23)
118 parser.add_argument("--maze_width", type=int, default=39)
120 parser.add_argument("--maze_nb_walls", type=int, default=45)
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": 250000,
193 "nb_test_samples": 10000,
199 "nb_train_samples": 50000,
200 "nb_test_samples": 10000,
206 "nb_train_samples": 50000,
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": 250000,
228 "nb_test_samples": 10000,
234 "nb_train_samples": 50000,
235 "nb_test_samples": 10000,
241 "nb_train_samples": 100000,
242 "nb_test_samples": 1000,
248 "nb_train_samples": 1000000,
249 "nb_test_samples": 10000,
255 "nb_train_samples": 100000,
256 "nb_test_samples": 10000,
262 "nb_train_samples": 25000,
263 "nb_test_samples": 1000,
267 if args.task in default_task_args:
268 for k, v in default_task_args[args.task].items():
269 if getattr(args, k) is None:
272 ######################################################################
274 default_model_args = {
305 if args.model in default_model_args:
306 for k, v in default_model_args[args.model].items():
307 if getattr(args, k) is None:
310 raise ValueError(f"Unknown model {args.model}")
312 ######################################################################
315 os.mkdir(args.result_dir)
316 except FileExistsError:
317 if not args.overwrite_results:
318 print(f"result directory {args.result_dir} already exists")
321 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
324 # torch.backends.cudnn.deterministic = True
325 # torch.backends.cudnn.benchmark = False
326 # torch.use_deterministic_algorithms(True)
327 torch.manual_seed(args.seed)
328 if torch.cuda.is_available():
329 torch.cuda.manual_seed_all(args.seed)
331 ######################################################################
335 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
337 if log_file is not None:
338 log_file.write(t + s + "\n")
346 log_string(f"args.{n} {getattr(args, n)}")
349 ######################################################################
352 def picoclvr_pruner_horizontal_green(p):
353 return not ("green" in p and ("left" in p or "right" in p))
356 picoclvr_pruner_train = (
357 picoclvr_pruner_horizontal_green
358 if args.picocvlr_prune_properties in {"train+eval"}
362 picoclvr_pruner_eval = (
363 (lambda p: not picoclvr_pruner_horizontal_green(p))
364 if args.picocvlr_prune_properties in {"train+eval", "eval"}
368 ######################################################################
370 if args.task == "byheart":
371 task = tasks.SandBox(
372 problem=problems.ProblemByHeart(),
373 nb_train_samples=args.nb_train_samples,
374 nb_test_samples=args.nb_test_samples,
375 batch_size=args.batch_size,
379 args.max_percents_of_test_in_train = -1
381 elif args.task == "learnop":
382 task = tasks.SandBox(
383 problem=problems.ProblemLearnOperator(),
384 nb_train_samples=args.nb_train_samples,
385 nb_test_samples=args.nb_test_samples,
386 batch_size=args.batch_size,
392 elif args.task == "guessop":
393 task = tasks.SandBox(
394 problem=problems.ProblemGuessOperator(),
395 nb_train_samples=args.nb_train_samples,
396 nb_test_samples=args.nb_test_samples,
397 batch_size=args.batch_size,
403 elif args.task == "twotargets":
404 task = tasks.SandBox(
405 problem=problems.ProblemTwoTargets(),
406 nb_train_samples=args.nb_train_samples,
407 nb_test_samples=args.nb_test_samples,
408 batch_size=args.batch_size,
413 elif args.task == "addition":
414 task = tasks.SandBox(
415 problem=problems.ProblemAddition(),
416 nb_train_samples=args.nb_train_samples,
417 nb_test_samples=args.nb_test_samples,
418 batch_size=args.batch_size,
423 elif args.task == "picoclvr":
424 task = tasks.PicoCLVR(
425 nb_train_samples=args.nb_train_samples,
426 nb_test_samples=args.nb_test_samples,
427 batch_size=args.batch_size,
428 height=args.picoclvr_height,
429 width=args.picoclvr_width,
430 nb_colors=args.picoclvr_nb_colors,
433 pruner_train=picoclvr_pruner_train,
434 pruner_eval=picoclvr_pruner_eval,
437 elif args.task == "mnist":
439 nb_train_samples=args.nb_train_samples,
440 nb_test_samples=args.nb_test_samples,
441 batch_size=args.batch_size,
445 elif args.task == "maze":
447 nb_train_samples=args.nb_train_samples,
448 nb_test_samples=args.nb_test_samples,
449 batch_size=args.batch_size,
450 height=args.maze_height,
451 width=args.maze_width,
452 nb_walls=args.maze_nb_walls,
456 elif args.task == "snake":
458 nb_train_samples=args.nb_train_samples,
459 nb_test_samples=args.nb_test_samples,
460 batch_size=args.batch_size,
461 height=args.snake_height,
462 width=args.snake_width,
463 nb_colors=args.snake_nb_colors,
464 length=args.snake_length,
465 prompt_length=args.snake_length // 2,
469 elif args.task == "stack":
471 nb_train_samples=args.nb_train_samples,
472 nb_test_samples=args.nb_test_samples,
473 batch_size=args.batch_size,
475 nb_steps=args.stack_nb_steps,
476 nb_stacks=args.stack_nb_stacks,
477 nb_digits=args.stack_nb_digits,
478 fraction_values_for_train=args.stack_fraction_values_for_train,
482 elif args.task == "expr":
484 nb_train_samples=args.nb_train_samples,
485 nb_test_samples=args.nb_test_samples,
486 nb_variables=args.expr_nb_variables,
487 sequence_length=args.expr_sequence_length,
488 operand_max=args.expr_operand_max,
489 result_max=args.expr_result_max,
490 batch_size=args.batch_size,
494 elif args.task == "rpl":
496 nb_train_samples=args.nb_train_samples,
497 nb_test_samples=args.nb_test_samples,
498 batch_size=args.batch_size,
499 nb_starting_values=args.rpl_nb_starting_values,
500 max_input=args.rpl_max_input,
501 prog_len=args.rpl_prog_len,
502 nb_runs=args.rpl_nb_runs,
503 no_prog=args.rpl_no_prog,
508 elif args.task == "world":
510 nb_train_samples=args.nb_train_samples,
511 nb_test_samples=args.nb_test_samples,
512 batch_size=args.batch_size,
513 vqae_nb_epochs=args.world_vqae_nb_epochs,
519 raise ValueError(f"Unknown task {args.task}")
521 ######################################################################
523 log_string(f"device {device}")
525 vocabulary_size = task.vocabulary_size()
527 log_string(f"vocabulary_size {vocabulary_size}")
529 ##############################
532 vocabulary_size=vocabulary_size,
533 dim_model=args.dim_model,
534 dim_keys=args.dim_keys,
535 dim_hidden=args.dim_hidden,
536 nb_heads=args.nb_heads,
537 nb_blocks=args.nb_blocks,
539 dropout=args.dropout,
544 nb_parameters = sum(p.numel() for p in model.parameters())
545 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
547 ######################################################################
549 nb_epochs_finished = 0
551 if args.no_checkpoint:
552 log_string(f"not trying to load checkpoint.")
556 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
557 checkpoint = torch.load(checkpoint_name)
558 nb_epochs_finished = checkpoint["nb_epochs_finished"]
559 model.load_state_dict(checkpoint["model_state"])
560 torch.set_rng_state(checkpoint["rng_state"])
561 if torch.cuda.is_available():
562 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
564 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
566 except FileNotFoundError:
567 log_string("starting from scratch.")
570 log_string("error when loading the checkpoint.")
573 ######################################################################
575 if args.task == "expr" and args.expr_input_file is not None:
576 task.produce_results(
577 n_epoch=nb_epochs_finished,
579 result_dir=args.result_dir,
581 deterministic_synthesis=args.deterministic_synthesis,
582 input_file=args.expr_input_file,
587 ######################################################################
589 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
591 # Compute the entropy of the training tokens
594 for input in task.batches(split="train"):
595 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
596 token_probas = token_count / token_count.sum()
597 entropy = -torch.xlogy(token_probas, token_probas).sum()
598 train_set_perplexity = math.exp(entropy)
600 ######################################################################
601 # A bit of paranoia never hurts
603 if args.max_percents_of_test_in_train >= 0:
605 def subsets_as_tuples(batches, cs):
607 for batch in batches:
609 s.add(tuple([v.item() for v in x]))
615 nb_test, nb_in_train = 0, 0
616 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
618 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
619 in_train.update(test_subset.intersection(train_subset))
620 nb_in_train += len(in_train)
621 nb_test += len(test_subset)
624 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
628 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
629 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
631 ##############################
633 if args.learning_rate_schedule == "cos":
634 learning_rate_schedule = {}
635 for n_epoch in range(args.nb_epochs):
636 u = n_epoch / args.nb_epochs * math.pi
637 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
642 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
646 learning_rate_schedule = {}
647 learning_rate = args.learning_rate
648 for n_epoch in range(args.nb_epochs):
650 learning_rate = u[n_epoch]
651 learning_rate_schedule[n_epoch] = learning_rate
653 log_string(f"learning_rate_schedule {learning_rate_schedule}")
655 ##############################
659 if nb_epochs_finished >= nb_epochs:
660 task.produce_results(
661 n_epoch=nb_epochs_finished,
663 result_dir=args.result_dir,
665 deterministic_synthesis=args.deterministic_synthesis,
668 for n_epoch in range(nb_epochs_finished, nb_epochs):
669 learning_rate = learning_rate_schedule[n_epoch]
671 log_string(f"learning_rate {learning_rate}")
673 if args.optim == "sgd":
674 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
675 elif args.optim == "adam":
676 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
677 elif args.optim == "adamw":
678 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
680 raise ValueError(f"Unknown optimizer {args.optim}.")
684 nb_train_samples, acc_train_loss = 0, 0.0
686 for input in task.batches(split="train"):
687 input = input.to(device)
688 output = model(mygpt.BracketedSequence(input)).x
689 loss = F.cross_entropy(output.transpose(1, 2), input)
690 acc_train_loss += loss.item() * input.size(0)
691 nb_train_samples += input.size(0)
692 nb_samples_seen += input.size(0)
694 optimizer.zero_grad()
698 with torch.autograd.no_grad():
701 nb_test_samples, acc_test_loss = 0, 0.0
703 for input in task.batches(split="test"):
704 input = input.to(device)
706 output = model(mygpt.BracketedSequence(input)).x
707 loss = F.cross_entropy(output.transpose(1, 2), input)
708 acc_test_loss += loss.item() * input.size(0)
709 nb_test_samples += input.size(0)
711 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
712 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
715 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
718 task.produce_results(
721 result_dir=args.result_dir,
723 deterministic_synthesis=args.deterministic_synthesis,
727 "nb_epochs_finished": n_epoch + 1,
728 "model_state": model.state_dict(),
729 "rng_state": torch.get_rng_state(),
732 if torch.cuda.is_available():
733 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
735 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
736 torch.save(checkpoint, checkpoint_name)
737 log_string(f"saved checkpoint {checkpoint_name}")
739 ######################################################################