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("--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("--world_vqae_nb_epochs", type=int, default=25)
167 ######################################################################
169 args = parser.parse_args()
171 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
173 if args.result_dir is None:
174 args.result_dir = f"results_{args.task}"
176 ######################################################################
178 default_task_args = {
183 "nb_train_samples": 50000,
184 "nb_test_samples": 10000,
190 "nb_train_samples": 50000,
191 "nb_test_samples": 10000,
197 "nb_train_samples": 1000000,
198 "nb_test_samples": 10000,
204 "nb_train_samples": 50000,
205 "nb_test_samples": 10000,
211 "nb_train_samples": 250000,
212 "nb_test_samples": 10000,
218 "nb_train_samples": 250000,
219 "nb_test_samples": 10000,
225 "nb_train_samples": 60000,
226 "nb_test_samples": 10000,
232 "nb_train_samples": 100000,
233 "nb_test_samples": 10000,
239 "nb_train_samples": 250000,
240 "nb_test_samples": 10000,
246 "nb_train_samples": 100000,
247 "nb_test_samples": 1000,
253 "nb_train_samples": 2500000,
254 "nb_test_samples": 10000,
260 "nb_train_samples": 1000000,
261 "nb_test_samples": 10000,
267 "nb_train_samples": 25000,
268 "nb_test_samples": 1000,
274 "nb_train_samples": 250000,
275 "nb_test_samples": 10000,
279 if args.task in default_task_args:
280 for k, v in default_task_args[args.task].items():
281 if getattr(args, k) is None:
284 ######################################################################
286 default_model_args = {
317 if args.model in default_model_args:
318 for k, v in default_model_args[args.model].items():
319 if getattr(args, k) is None:
322 raise ValueError(f"Unknown model {args.model}")
324 ######################################################################
327 os.mkdir(args.result_dir)
328 except FileExistsError:
329 if not args.overwrite_results:
330 print(f"result directory {args.result_dir} already exists")
333 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
336 # torch.backends.cudnn.deterministic = True
337 # torch.backends.cudnn.benchmark = False
338 # torch.use_deterministic_algorithms(True)
339 torch.manual_seed(args.seed)
340 if torch.cuda.is_available():
341 torch.cuda.manual_seed_all(args.seed)
343 ######################################################################
347 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
349 if log_file is not None:
350 log_file.write(t + s + "\n")
358 log_string(f"args.{n} {getattr(args, n)}")
361 ######################################################################
364 def picoclvr_pruner_horizontal_green(p):
365 return not ("green" in p and ("left" in p or "right" in p))
368 picoclvr_pruner_train = (
369 picoclvr_pruner_horizontal_green
370 if args.picocvlr_prune_properties in {"train+eval"}
374 picoclvr_pruner_eval = (
375 (lambda p: not picoclvr_pruner_horizontal_green(p))
376 if args.picocvlr_prune_properties in {"train+eval", "eval"}
380 ######################################################################
382 if args.task == "byheart":
383 task = tasks.SandBox(
384 problem=problems.ProblemByHeart(),
385 nb_train_samples=args.nb_train_samples,
386 nb_test_samples=args.nb_test_samples,
387 batch_size=args.batch_size,
391 args.max_percents_of_test_in_train = -1
393 elif args.task == "learnop":
394 task = tasks.SandBox(
395 problem=problems.ProblemLearnOperator(),
396 nb_train_samples=args.nb_train_samples,
397 nb_test_samples=args.nb_test_samples,
398 batch_size=args.batch_size,
404 elif args.task == "guessop":
405 task = tasks.SandBox(
406 problem=problems.ProblemGuessOperator(),
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 == "twotargets":
416 task = tasks.SandBox(
417 problem=problems.ProblemTwoTargets(),
418 nb_train_samples=args.nb_train_samples,
419 nb_test_samples=args.nb_test_samples,
420 batch_size=args.batch_size,
425 elif args.task == "addition":
426 task = tasks.SandBox(
427 problem=problems.ProblemAddition(),
428 nb_train_samples=args.nb_train_samples,
429 nb_test_samples=args.nb_test_samples,
430 batch_size=args.batch_size,
435 elif args.task == "picoclvr":
436 task = tasks.PicoCLVR(
437 nb_train_samples=args.nb_train_samples,
438 nb_test_samples=args.nb_test_samples,
439 batch_size=args.batch_size,
440 height=args.picoclvr_height,
441 width=args.picoclvr_width,
442 nb_colors=args.picoclvr_nb_colors,
445 pruner_train=picoclvr_pruner_train,
446 pruner_eval=picoclvr_pruner_eval,
449 elif args.task == "mnist":
451 nb_train_samples=args.nb_train_samples,
452 nb_test_samples=args.nb_test_samples,
453 batch_size=args.batch_size,
457 elif args.task == "maze":
459 nb_train_samples=args.nb_train_samples,
460 nb_test_samples=args.nb_test_samples,
461 batch_size=args.batch_size,
462 height=args.maze_height,
463 width=args.maze_width,
464 nb_walls=args.maze_nb_walls,
468 elif args.task == "snake":
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.snake_height,
474 width=args.snake_width,
475 nb_colors=args.snake_nb_colors,
476 length=args.snake_length,
477 prompt_length=args.snake_length // 2,
481 elif args.task == "stack":
483 nb_train_samples=args.nb_train_samples,
484 nb_test_samples=args.nb_test_samples,
485 batch_size=args.batch_size,
487 nb_steps=args.stack_nb_steps,
488 nb_stacks=args.stack_nb_stacks,
489 nb_digits=args.stack_nb_digits,
490 fraction_values_for_train=args.stack_fraction_values_for_train,
494 elif args.task == "expr":
496 nb_train_samples=args.nb_train_samples,
497 nb_test_samples=args.nb_test_samples,
498 nb_variables=args.expr_nb_variables,
499 sequence_length=args.expr_sequence_length,
500 operand_max=args.expr_operand_max,
501 result_max=args.expr_result_max,
502 batch_size=args.batch_size,
506 elif args.task == "rpl":
508 nb_train_samples=args.nb_train_samples,
509 nb_test_samples=args.nb_test_samples,
510 batch_size=args.batch_size,
511 nb_starting_values=args.rpl_nb_starting_values,
512 max_input=args.rpl_max_input,
513 prog_len=args.rpl_prog_len,
514 nb_runs=args.rpl_nb_runs,
515 no_prog=args.rpl_no_prog,
520 elif args.task == "grid":
522 nb_train_samples=args.nb_train_samples,
523 nb_test_samples=args.nb_test_samples,
524 batch_size=args.batch_size,
530 elif args.task == "world":
532 nb_train_samples=args.nb_train_samples,
533 nb_test_samples=args.nb_test_samples,
534 batch_size=args.batch_size,
535 vqae_nb_epochs=args.world_vqae_nb_epochs,
541 raise ValueError(f"Unknown task {args.task}")
543 ######################################################################
545 log_string(f"device {device}")
547 vocabulary_size = task.vocabulary_size()
549 log_string(f"vocabulary_size {vocabulary_size}")
551 ##############################
554 vocabulary_size=vocabulary_size,
555 dim_model=args.dim_model,
556 dim_keys=args.dim_keys,
557 dim_hidden=args.dim_hidden,
558 nb_heads=args.nb_heads,
559 nb_blocks=args.nb_blocks,
561 dropout=args.dropout,
566 nb_parameters = sum(p.numel() for p in model.parameters())
567 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
569 ######################################################################
571 nb_epochs_finished = 0
573 if args.no_checkpoint:
574 log_string(f"not trying to load checkpoint.")
578 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
579 checkpoint = torch.load(checkpoint_name)
580 nb_epochs_finished = checkpoint["nb_epochs_finished"]
581 model.load_state_dict(checkpoint["model_state"])
582 torch.set_rng_state(checkpoint["rng_state"])
583 if torch.cuda.is_available():
584 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
586 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
588 except FileNotFoundError:
589 log_string("starting from scratch.")
592 log_string("error when loading the checkpoint.")
595 ######################################################################
597 if args.task == "expr" and args.expr_input_file is not None:
598 task.produce_results(
599 n_epoch=nb_epochs_finished,
601 result_dir=args.result_dir,
603 deterministic_synthesis=args.deterministic_synthesis,
604 input_file=args.expr_input_file,
609 ######################################################################
611 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
613 # Compute the entropy of the training tokens
616 for input in task.batches(split="train"):
617 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
618 token_probas = token_count / token_count.sum()
619 entropy = -torch.xlogy(token_probas, token_probas).sum()
620 train_set_perplexity = math.exp(entropy)
622 ######################################################################
623 # A bit of paranoia never hurts
625 if args.max_percents_of_test_in_train >= 0:
627 def subsets_as_tuples(batches, cs):
629 for batch in batches:
631 s.add(tuple([v.item() for v in x]))
637 nb_test, nb_in_train = 0, 0
638 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
640 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
641 in_train.update(test_subset.intersection(train_subset))
642 nb_in_train += len(in_train)
643 nb_test += len(test_subset)
646 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
650 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
651 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
653 ##############################
655 if args.learning_rate_schedule == "cos":
656 learning_rate_schedule = {}
657 for n_epoch in range(args.nb_epochs):
658 u = n_epoch / args.nb_epochs * math.pi
659 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
664 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
668 learning_rate_schedule = {}
669 learning_rate = args.learning_rate
670 for n_epoch in range(args.nb_epochs):
672 learning_rate = u[n_epoch]
673 learning_rate_schedule[n_epoch] = learning_rate
675 log_string(f"learning_rate_schedule {learning_rate_schedule}")
677 ##############################
681 if nb_epochs_finished >= nb_epochs:
682 task.produce_results(
683 n_epoch=nb_epochs_finished,
685 result_dir=args.result_dir,
687 deterministic_synthesis=args.deterministic_synthesis,
690 for n_epoch in range(nb_epochs_finished, nb_epochs):
691 learning_rate = learning_rate_schedule[n_epoch]
693 log_string(f"learning_rate {learning_rate}")
695 if args.optim == "sgd":
696 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
697 elif args.optim == "adam":
698 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
699 elif args.optim == "adamw":
700 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
702 raise ValueError(f"Unknown optimizer {args.optim}.")
706 nb_train_samples, acc_train_loss = 0, 0.0
708 for input in task.batches(split="train"):
709 input = input.to(device)
710 output = model(mygpt.BracketedSequence(input)).x
711 loss = F.cross_entropy(output.transpose(1, 2), input)
712 acc_train_loss += loss.item() * input.size(0)
713 nb_train_samples += input.size(0)
714 nb_samples_seen += input.size(0)
716 optimizer.zero_grad()
720 with torch.autograd.no_grad():
723 nb_test_samples, acc_test_loss = 0, 0.0
725 for input in task.batches(split="test"):
726 input = input.to(device)
728 output = model(mygpt.BracketedSequence(input)).x
729 loss = F.cross_entropy(output.transpose(1, 2), input)
730 acc_test_loss += loss.item() * input.size(0)
731 nb_test_samples += input.size(0)
733 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
734 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
737 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
740 task.produce_results(
743 result_dir=args.result_dir,
745 deterministic_synthesis=args.deterministic_synthesis,
749 "nb_epochs_finished": n_epoch + 1,
750 "model_state": model.state_dict(),
751 "rng_state": torch.get_rng_state(),
754 if torch.cuda.is_available():
755 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
757 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
758 torch.save(checkpoint, checkpoint_name)
759 log_string(f"saved checkpoint {checkpoint_name}")
761 ######################################################################