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, degradation, 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 ##############################
165 parser.add_argument("--degradation_hard", action="store_true", default=False)
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 = {
182 "nb_train_samples": 250000,
183 "nb_test_samples": 10000,
188 "nb_train_samples": 50000,
189 "nb_test_samples": 10000,
194 "nb_train_samples": 2500000,
195 "nb_test_samples": 10000,
200 "nb_train_samples": 250000,
201 "nb_test_samples": 10000,
206 "nb_train_samples": 100000,
207 "nb_test_samples": 1000,
212 "nb_train_samples": 1000000,
213 "nb_test_samples": 10000,
218 "nb_train_samples": 50000,
219 "nb_test_samples": 10000,
224 "nb_train_samples": 100000,
225 "nb_test_samples": 10000,
230 "nb_train_samples": 250000,
231 "nb_test_samples": 10000,
236 "nb_train_samples": 2500000,
237 "nb_test_samples": 10000,
242 "nb_train_samples": 250000,
243 "nb_test_samples": 10000,
248 "nb_train_samples": 100000,
249 "nb_test_samples": 1000,
254 "nb_train_samples": 50000,
255 "nb_test_samples": 10000,
260 "nb_train_samples": 100000,
261 "nb_test_samples": 10000,
266 "nb_train_samples": 60000,
267 "nb_test_samples": 10000,
271 if args.task in default_task_args:
272 for k, v in default_task_args[args.task].items():
273 if getattr(args, k) is None:
276 ######################################################################
278 default_model_args = {
309 if args.model in default_model_args:
310 for k, v in default_model_args[args.model].items():
311 if getattr(args, k) is None:
314 raise ValueError(f"Unknown model {args.model}")
316 ######################################################################
319 os.mkdir(args.result_dir)
320 except FileExistsError:
321 if not args.overwrite_results:
322 print(f"result directory {args.result_dir} already exists")
325 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
328 # torch.backends.cudnn.deterministic = True
329 # torch.backends.cudnn.benchmark = False
330 # torch.use_deterministic_algorithms(True)
331 torch.manual_seed(args.seed)
332 if torch.cuda.is_available():
333 torch.cuda.manual_seed_all(args.seed)
335 ######################################################################
339 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
341 if log_file is not None:
342 log_file.write(t + s + "\n")
350 log_string(f"args.{n} {getattr(args, n)}")
353 ######################################################################
356 def picoclvr_pruner_horizontal_green(p):
357 return not ("green" in p and ("left" in p or "right" in p))
360 picoclvr_pruner_train = (
361 picoclvr_pruner_horizontal_green
362 if args.picocvlr_prune_properties in {"train+eval"}
366 picoclvr_pruner_eval = (
367 (lambda p: not picoclvr_pruner_horizontal_green(p))
368 if args.picocvlr_prune_properties in {"train+eval", "eval"}
372 ######################################################################
374 if args.task == "byheart":
375 task = tasks.SandBox(
376 problem=problems.ProblemByHeart(),
377 nb_train_samples=args.nb_train_samples,
378 nb_test_samples=args.nb_test_samples,
379 batch_size=args.batch_size,
383 args.max_percents_of_test_in_train = -1
385 elif args.task == "learnop":
386 task = tasks.SandBox(
387 problem=problems.ProblemLearnOperator(),
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 == "guessop":
397 task = tasks.SandBox(
398 problem=problems.ProblemGuessOperator(),
399 nb_train_samples=args.nb_train_samples,
400 nb_test_samples=args.nb_test_samples,
401 batch_size=args.batch_size,
407 elif args.task == "twotargets":
408 task = tasks.SandBox(
409 problem=problems.ProblemTwoTargets(),
410 nb_train_samples=args.nb_train_samples,
411 nb_test_samples=args.nb_test_samples,
412 batch_size=args.batch_size,
417 elif args.task == "degradation":
418 task = tasks.SandBox(
419 problem=problems.ProblemDegradation(hard=args.degradation_hard),
420 nb_train_samples=args.nb_train_samples,
421 nb_test_samples=args.nb_test_samples,
422 batch_size=args.batch_size,
427 elif args.task == "addition":
428 task = tasks.SandBox(
429 problem=problems.ProblemAddition(),
430 nb_train_samples=args.nb_train_samples,
431 nb_test_samples=args.nb_test_samples,
432 batch_size=args.batch_size,
437 elif args.task == "picoclvr":
438 task = tasks.PicoCLVR(
439 nb_train_samples=args.nb_train_samples,
440 nb_test_samples=args.nb_test_samples,
441 batch_size=args.batch_size,
442 height=args.picoclvr_height,
443 width=args.picoclvr_width,
444 nb_colors=args.picoclvr_nb_colors,
447 pruner_train=picoclvr_pruner_train,
448 pruner_eval=picoclvr_pruner_eval,
451 elif args.task == "mnist":
453 nb_train_samples=args.nb_train_samples,
454 nb_test_samples=args.nb_test_samples,
455 batch_size=args.batch_size,
459 elif args.task == "maze":
461 nb_train_samples=args.nb_train_samples,
462 nb_test_samples=args.nb_test_samples,
463 batch_size=args.batch_size,
464 height=args.maze_height,
465 width=args.maze_width,
466 nb_walls=args.maze_nb_walls,
470 elif args.task == "snake":
472 nb_train_samples=args.nb_train_samples,
473 nb_test_samples=args.nb_test_samples,
474 batch_size=args.batch_size,
475 height=args.snake_height,
476 width=args.snake_width,
477 nb_colors=args.snake_nb_colors,
478 length=args.snake_length,
479 prompt_length=args.snake_length // 2,
483 elif args.task == "stack":
485 nb_train_samples=args.nb_train_samples,
486 nb_test_samples=args.nb_test_samples,
487 batch_size=args.batch_size,
489 nb_steps=args.stack_nb_steps,
490 nb_stacks=args.stack_nb_stacks,
491 nb_digits=args.stack_nb_digits,
492 fraction_values_for_train=args.stack_fraction_values_for_train,
496 elif args.task == "expr":
498 nb_train_samples=args.nb_train_samples,
499 nb_test_samples=args.nb_test_samples,
500 nb_variables=args.expr_nb_variables,
501 sequence_length=args.expr_sequence_length,
502 operand_max=args.expr_operand_max,
503 result_max=args.expr_result_max,
504 batch_size=args.batch_size,
508 elif args.task == "rpl":
510 nb_train_samples=args.nb_train_samples,
511 nb_test_samples=args.nb_test_samples,
512 batch_size=args.batch_size,
513 nb_starting_values=args.rpl_nb_starting_values,
514 max_input=args.rpl_max_input,
515 prog_len=args.rpl_prog_len,
516 nb_runs=args.rpl_nb_runs,
517 no_prog=args.rpl_no_prog,
522 elif args.task == "grid":
524 nb_train_samples=args.nb_train_samples,
525 nb_test_samples=args.nb_test_samples,
526 batch_size=args.batch_size,
532 elif args.task == "qmlp":
534 nb_train_samples=args.nb_train_samples,
535 nb_test_samples=args.nb_test_samples,
536 batch_size=args.batch_size,
537 result_dir=args.result_dir,
543 raise ValueError(f"Unknown task {args.task}")
545 ######################################################################
547 log_string(f"device {device}")
549 vocabulary_size = task.vocabulary_size()
551 log_string(f"vocabulary_size {vocabulary_size}")
553 ##############################
556 vocabulary_size=vocabulary_size,
557 dim_model=args.dim_model,
558 dim_keys=args.dim_keys,
559 dim_hidden=args.dim_hidden,
560 nb_heads=args.nb_heads,
561 nb_blocks=args.nb_blocks,
563 dropout=args.dropout,
568 nb_parameters = sum(p.numel() for p in model.parameters())
569 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
571 ######################################################################
573 nb_epochs_finished = 0
575 if args.no_checkpoint:
576 log_string(f"not trying to load checkpoint.")
580 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
581 checkpoint = torch.load(checkpoint_name)
582 nb_epochs_finished = checkpoint["nb_epochs_finished"]
583 model.load_state_dict(checkpoint["model_state"])
584 torch.set_rng_state(checkpoint["rng_state"])
585 if torch.cuda.is_available():
586 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
588 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
590 except FileNotFoundError:
591 log_string("starting from scratch.")
594 log_string("error when loading the checkpoint.")
597 ######################################################################
599 if args.task == "expr" and args.expr_input_file is not None:
600 task.produce_results(
601 n_epoch=nb_epochs_finished,
603 result_dir=args.result_dir,
605 deterministic_synthesis=args.deterministic_synthesis,
606 input_file=args.expr_input_file,
611 ######################################################################
613 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
615 # Compute the entropy of the training tokens
618 for input in task.batches(split="train"):
619 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
620 token_probas = token_count / token_count.sum()
621 entropy = -torch.xlogy(token_probas, token_probas).sum()
622 train_set_perplexity = math.exp(entropy)
624 ######################################################################
625 # A bit of paranoia never hurts
627 if args.max_percents_of_test_in_train >= 0:
629 def subsets_as_tuples(batches, cs):
631 for batch in batches:
633 s.add(tuple([v.item() for v in x]))
639 nb_test, nb_in_train = 0, 0
640 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
642 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
643 in_train.update(test_subset.intersection(train_subset))
644 nb_in_train += len(in_train)
645 nb_test += len(test_subset)
648 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
652 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
653 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
655 ##############################
657 if args.learning_rate_schedule == "cos":
658 learning_rate_schedule = {}
659 for n_epoch in range(args.nb_epochs):
660 u = n_epoch / args.nb_epochs * math.pi
661 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
666 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
670 learning_rate_schedule = {}
671 learning_rate = args.learning_rate
672 for n_epoch in range(args.nb_epochs):
674 learning_rate = u[n_epoch]
675 learning_rate_schedule[n_epoch] = learning_rate
677 log_string(f"learning_rate_schedule {learning_rate_schedule}")
679 ##############################
683 if nb_epochs_finished >= nb_epochs:
684 task.produce_results(
685 n_epoch=nb_epochs_finished,
687 result_dir=args.result_dir,
689 deterministic_synthesis=args.deterministic_synthesis,
692 for n_epoch in range(nb_epochs_finished, nb_epochs):
693 learning_rate = learning_rate_schedule[n_epoch]
695 log_string(f"learning_rate {learning_rate}")
697 if args.optim == "sgd":
698 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
699 elif args.optim == "adam":
700 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
701 elif args.optim == "adamw":
702 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
704 raise ValueError(f"Unknown optimizer {args.optim}.")
708 nb_train_samples, acc_train_loss = 0, 0.0
710 for input in task.batches(split="train"):
711 input = input.to(device)
712 output = model(mygpt.BracketedSequence(input)).x
713 loss = F.cross_entropy(output.transpose(1, 2), input)
714 acc_train_loss += loss.item() * input.size(0)
715 nb_train_samples += input.size(0)
716 nb_samples_seen += input.size(0)
718 optimizer.zero_grad()
722 with torch.autograd.no_grad():
725 nb_test_samples, acc_test_loss = 0, 0.0
727 for input in task.batches(split="test"):
728 input = input.to(device)
730 output = model(mygpt.BracketedSequence(input)).x
731 loss = F.cross_entropy(output.transpose(1, 2), input)
732 acc_test_loss += loss.item() * input.size(0)
733 nb_test_samples += input.size(0)
735 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
736 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
739 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
742 task.produce_results(
745 result_dir=args.result_dir,
747 deterministic_synthesis=args.deterministic_synthesis,
751 "nb_epochs_finished": n_epoch + 1,
752 "model_state": model.state_dict(),
753 "rng_state": torch.get_rng_state(),
756 if torch.cuda.is_available():
757 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
759 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
760 torch.save(checkpoint, checkpoint_name)
761 log_string(f"saved checkpoint {checkpoint_name}")
763 ######################################################################