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, 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 ######################################################################
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 = {
177 "nb_train_samples": 250000,
178 "nb_test_samples": 10000,
183 "nb_train_samples": 50000,
184 "nb_test_samples": 10000,
189 "nb_train_samples": 2500000,
190 "nb_test_samples": 10000,
195 "nb_train_samples": 250000,
196 "nb_test_samples": 10000,
201 "nb_train_samples": 100000,
202 "nb_test_samples": 1000,
207 "nb_train_samples": 1000000,
208 "nb_test_samples": 10000,
213 "nb_train_samples": 50000,
214 "nb_test_samples": 10000,
219 "nb_train_samples": 100000,
220 "nb_test_samples": 10000,
225 "nb_train_samples": 250000,
226 "nb_test_samples": 10000,
231 "nb_train_samples": 2500000,
232 "nb_test_samples": 10000,
237 "nb_train_samples": 250000,
238 "nb_test_samples": 10000,
243 "nb_train_samples": 100000,
244 "nb_test_samples": 1000,
249 "nb_train_samples": 50000,
250 "nb_test_samples": 10000,
255 "nb_train_samples": 60000,
256 "nb_test_samples": 10000,
260 if args.task in default_task_args:
261 for k, v in default_task_args[args.task].items():
262 if getattr(args, k) is None:
265 ######################################################################
267 default_model_args = {
298 if args.model in default_model_args:
299 for k, v in default_model_args[args.model].items():
300 if getattr(args, k) is None:
303 raise ValueError(f"Unknown model {args.model}")
305 ######################################################################
308 os.mkdir(args.result_dir)
309 except FileExistsError:
310 if not args.overwrite_results:
311 print(f"result directory {args.result_dir} already exists")
314 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
317 # torch.backends.cudnn.deterministic = True
318 # torch.backends.cudnn.benchmark = False
319 # torch.use_deterministic_algorithms(True)
320 torch.manual_seed(args.seed)
321 if torch.cuda.is_available():
322 torch.cuda.manual_seed_all(args.seed)
324 ######################################################################
328 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
330 if log_file is not None:
331 log_file.write(t + s + "\n")
339 log_string(f"args.{n} {getattr(args, n)}")
342 ######################################################################
345 def picoclvr_pruner_horizontal_green(p):
346 return not ("green" in p and ("left" in p or "right" in p))
349 picoclvr_pruner_train = (
350 picoclvr_pruner_horizontal_green
351 if args.picocvlr_prune_properties in {"train+eval"}
355 picoclvr_pruner_eval = (
356 (lambda p: not picoclvr_pruner_horizontal_green(p))
357 if args.picocvlr_prune_properties in {"train+eval", "eval"}
361 ######################################################################
363 if args.task == "byheart":
364 task = tasks.SandBox(
365 problem=problems.ProblemByHeart(),
366 nb_train_samples=args.nb_train_samples,
367 nb_test_samples=args.nb_test_samples,
368 batch_size=args.batch_size,
372 args.max_percents_of_test_in_train = -1
374 elif args.task == "learnop":
375 task = tasks.SandBox(
376 problem=problems.ProblemLearnOperator(),
377 nb_train_samples=args.nb_train_samples,
378 nb_test_samples=args.nb_test_samples,
379 batch_size=args.batch_size,
385 elif args.task == "guessop":
386 task = tasks.SandBox(
387 problem=problems.ProblemGuessOperator(),
388 nb_train_samples=args.nb_train_samples,
389 nb_test_samples=args.nb_test_samples,
390 batch_size=args.batch_size,
396 elif args.task == "twotargets":
397 task = tasks.SandBox(
398 problem=problems.ProblemTwoTargets(),
399 nb_train_samples=args.nb_train_samples,
400 nb_test_samples=args.nb_test_samples,
401 batch_size=args.batch_size,
406 elif args.task == "addition":
407 task = tasks.SandBox(
408 problem=problems.ProblemAddition(),
409 nb_train_samples=args.nb_train_samples,
410 nb_test_samples=args.nb_test_samples,
411 batch_size=args.batch_size,
416 elif args.task == "picoclvr":
417 task = tasks.PicoCLVR(
418 nb_train_samples=args.nb_train_samples,
419 nb_test_samples=args.nb_test_samples,
420 batch_size=args.batch_size,
421 height=args.picoclvr_height,
422 width=args.picoclvr_width,
423 nb_colors=args.picoclvr_nb_colors,
426 pruner_train=picoclvr_pruner_train,
427 pruner_eval=picoclvr_pruner_eval,
430 elif args.task == "mnist":
432 nb_train_samples=args.nb_train_samples,
433 nb_test_samples=args.nb_test_samples,
434 batch_size=args.batch_size,
438 elif args.task == "maze":
440 nb_train_samples=args.nb_train_samples,
441 nb_test_samples=args.nb_test_samples,
442 batch_size=args.batch_size,
443 height=args.maze_height,
444 width=args.maze_width,
445 nb_walls=args.maze_nb_walls,
449 elif args.task == "snake":
451 nb_train_samples=args.nb_train_samples,
452 nb_test_samples=args.nb_test_samples,
453 batch_size=args.batch_size,
454 height=args.snake_height,
455 width=args.snake_width,
456 nb_colors=args.snake_nb_colors,
457 length=args.snake_length,
458 prompt_length=args.snake_length // 2,
462 elif args.task == "stack":
464 nb_train_samples=args.nb_train_samples,
465 nb_test_samples=args.nb_test_samples,
466 batch_size=args.batch_size,
468 nb_steps=args.stack_nb_steps,
469 nb_stacks=args.stack_nb_stacks,
470 nb_digits=args.stack_nb_digits,
471 fraction_values_for_train=args.stack_fraction_values_for_train,
475 elif args.task == "expr":
477 nb_train_samples=args.nb_train_samples,
478 nb_test_samples=args.nb_test_samples,
479 nb_variables=args.expr_nb_variables,
480 sequence_length=args.expr_sequence_length,
481 operand_max=args.expr_operand_max,
482 result_max=args.expr_result_max,
483 batch_size=args.batch_size,
487 elif args.task == "rpl":
489 nb_train_samples=args.nb_train_samples,
490 nb_test_samples=args.nb_test_samples,
491 batch_size=args.batch_size,
492 nb_starting_values=args.rpl_nb_starting_values,
493 max_input=args.rpl_max_input,
494 prog_len=args.rpl_prog_len,
495 nb_runs=args.rpl_nb_runs,
496 no_prog=args.rpl_no_prog,
501 elif args.task == "grid":
503 nb_train_samples=args.nb_train_samples,
504 nb_test_samples=args.nb_test_samples,
505 batch_size=args.batch_size,
511 elif args.task == "qmlp":
513 nb_train_samples=args.nb_train_samples,
514 nb_test_samples=args.nb_test_samples,
515 batch_size=args.batch_size,
516 result_dir=args.result_dir,
522 raise ValueError(f"Unknown task {args.task}")
524 ######################################################################
526 log_string(f"device {device}")
528 vocabulary_size = task.vocabulary_size()
530 log_string(f"vocabulary_size {vocabulary_size}")
532 ##############################
535 vocabulary_size=vocabulary_size,
536 dim_model=args.dim_model,
537 dim_keys=args.dim_keys,
538 dim_hidden=args.dim_hidden,
539 nb_heads=args.nb_heads,
540 nb_blocks=args.nb_blocks,
542 dropout=args.dropout,
547 nb_parameters = sum(p.numel() for p in model.parameters())
548 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
550 ######################################################################
552 nb_epochs_finished = 0
554 if args.no_checkpoint:
555 log_string(f"not trying to load checkpoint.")
559 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
560 checkpoint = torch.load(checkpoint_name)
561 nb_epochs_finished = checkpoint["nb_epochs_finished"]
562 model.load_state_dict(checkpoint["model_state"])
563 torch.set_rng_state(checkpoint["rng_state"])
564 if torch.cuda.is_available():
565 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
567 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
569 except FileNotFoundError:
570 log_string("starting from scratch.")
573 log_string("error when loading the checkpoint.")
576 ######################################################################
578 if args.task == "expr" and args.expr_input_file is not None:
579 task.produce_results(
580 n_epoch=nb_epochs_finished,
582 result_dir=args.result_dir,
584 deterministic_synthesis=args.deterministic_synthesis,
585 input_file=args.expr_input_file,
590 ######################################################################
592 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
594 # Compute the entropy of the training tokens
597 for input in task.batches(split="train"):
598 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
599 token_probas = token_count / token_count.sum()
600 entropy = -torch.xlogy(token_probas, token_probas).sum()
601 train_set_perplexity = math.exp(entropy)
603 ######################################################################
604 # A bit of paranoia never hurts
606 if args.max_percents_of_test_in_train >= 0:
608 def subsets_as_tuples(batches, cs):
610 for batch in batches:
612 s.add(tuple([v.item() for v in x]))
618 nb_test, nb_in_train = 0, 0
619 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
621 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
622 in_train.update(test_subset.intersection(train_subset))
623 nb_in_train += len(in_train)
624 nb_test += len(test_subset)
627 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
631 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
632 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
634 ##############################
636 if args.learning_rate_schedule == "cos":
637 learning_rate_schedule = {}
638 for n_epoch in range(args.nb_epochs):
639 u = n_epoch / args.nb_epochs * math.pi
640 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
645 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
649 learning_rate_schedule = {}
650 learning_rate = args.learning_rate
651 for n_epoch in range(args.nb_epochs):
653 learning_rate = u[n_epoch]
654 learning_rate_schedule[n_epoch] = learning_rate
656 log_string(f"learning_rate_schedule {learning_rate_schedule}")
658 ##############################
662 if nb_epochs_finished >= nb_epochs:
663 task.produce_results(
664 n_epoch=nb_epochs_finished,
666 result_dir=args.result_dir,
668 deterministic_synthesis=args.deterministic_synthesis,
671 for n_epoch in range(nb_epochs_finished, nb_epochs):
672 learning_rate = learning_rate_schedule[n_epoch]
674 log_string(f"learning_rate {learning_rate}")
676 if args.optim == "sgd":
677 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
678 elif args.optim == "adam":
679 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
680 elif args.optim == "adamw":
681 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
683 raise ValueError(f"Unknown optimizer {args.optim}.")
687 nb_train_samples, acc_train_loss = 0, 0.0
689 for input in task.batches(split="train"):
690 input = input.to(device)
691 output = model(mygpt.BracketedSequence(input)).x
692 loss = F.cross_entropy(output.transpose(1, 2), input)
693 acc_train_loss += loss.item() * input.size(0)
694 nb_train_samples += input.size(0)
695 nb_samples_seen += input.size(0)
697 optimizer.zero_grad()
701 with torch.autograd.no_grad():
704 nb_test_samples, acc_test_loss = 0, 0.0
706 for input in task.batches(split="test"):
707 input = input.to(device)
709 output = model(mygpt.BracketedSequence(input)).x
710 loss = F.cross_entropy(output.transpose(1, 2), input)
711 acc_test_loss += loss.item() * input.size(0)
712 nb_test_samples += input.size(0)
714 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
715 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
718 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
721 task.produce_results(
724 result_dir=args.result_dir,
726 deterministic_synthesis=args.deterministic_synthesis,
730 "nb_epochs_finished": n_epoch + 1,
731 "model_state": model.state_dict(),
732 "rng_state": torch.get_rng_state(),
735 if torch.cuda.is_available():
736 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
738 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
739 torch.save(checkpoint, checkpoint_name)
740 log_string(f"saved checkpoint {checkpoint_name}")
742 ######################################################################