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, datetime
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="file, byheart, learnop, guessop, mixing, memory, 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("--filetask_train_file", type=str, default=None)
94 parser.add_argument("--filetask_test_file", type=str, default=None)
96 ##############################
99 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
101 parser.add_argument("--rpl_max_input", type=int, default=9)
103 parser.add_argument("--rpl_prog_len", type=int, default=8)
105 parser.add_argument("--rpl_nb_runs", type=int, default=5)
107 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
109 ##############################
112 parser.add_argument("--grid_size", type=int, default=6)
114 parser.add_argument("--grid_fraction_play", type=float, default=0)
116 ##############################
119 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
121 parser.add_argument("--picoclvr_height", type=int, default=12)
123 parser.add_argument("--picoclvr_width", type=int, default=16)
125 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
127 ##############################
130 parser.add_argument("--maze_height", type=int, default=13)
132 parser.add_argument("--maze_width", type=int, default=21)
134 parser.add_argument("--maze_nb_walls", type=int, default=15)
136 ##############################
139 parser.add_argument("--snake_height", type=int, default=9)
141 parser.add_argument("--snake_width", type=int, default=12)
143 parser.add_argument("--snake_nb_colors", type=int, default=5)
145 parser.add_argument("--snake_length", type=int, default=200)
147 ##############################
150 parser.add_argument("--stack_nb_steps", type=int, default=100)
152 parser.add_argument("--stack_nb_stacks", type=int, default=3)
154 parser.add_argument("--stack_nb_digits", type=int, default=3)
156 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
158 ##############################
161 parser.add_argument("--expr_nb_variables", type=int, default=5)
163 parser.add_argument("--expr_sequence_length", type=int, default=40)
165 parser.add_argument("--expr_operand_max", type=int, default=9)
167 parser.add_argument("--expr_result_max", type=int, default=99)
169 parser.add_argument("--expr_input_file", type=str, default=None)
171 ##############################
174 parser.add_argument("--mixing_hard", action="store_true", default=False)
176 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
178 ######################################################################
180 args = parser.parse_args()
182 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
184 if args.result_dir is None:
185 args.result_dir = f"results_{args.task}"
187 ######################################################################
189 default_task_args = {
193 "nb_train_samples": 250000,
194 "nb_test_samples": 10000,
199 "nb_train_samples": 250000,
200 "nb_test_samples": 10000,
205 "nb_train_samples": 50000,
206 "nb_test_samples": 10000,
211 "nb_train_samples": 2500000,
212 "nb_test_samples": 10000,
217 "nb_train_samples": 250000,
218 "nb_test_samples": 10000,
223 "nb_train_samples": 100000,
224 "nb_test_samples": 1000,
229 "nb_train_samples": 1000000,
230 "nb_test_samples": 10000,
235 "nb_train_samples": 50000,
236 "nb_test_samples": 10000,
241 "nb_train_samples": 100000,
242 "nb_test_samples": 10000,
247 "nb_train_samples": 250000,
248 "nb_test_samples": 10000,
253 "nb_train_samples": 2500000,
254 "nb_test_samples": 10000,
259 "nb_train_samples": 250000,
260 "nb_test_samples": 10000,
265 "nb_train_samples": 100000,
266 "nb_test_samples": 1000,
271 "nb_train_samples": 50000,
272 "nb_test_samples": 10000,
277 "nb_train_samples": 25000,
278 "nb_test_samples": 1000,
283 "nb_train_samples": 250000,
284 "nb_test_samples": 10000,
289 "nb_train_samples": 60000,
290 "nb_test_samples": 10000,
294 if args.task in default_task_args:
295 for k, v in default_task_args[args.task].items():
296 if getattr(args, k) is None:
299 ######################################################################
301 default_model_args = {
339 if args.model in default_model_args:
340 for k, v in default_model_args[args.model].items():
341 if getattr(args, k) is None:
344 raise ValueError(f"Unknown model {args.model}")
346 ######################################################################
349 os.mkdir(args.result_dir)
350 except FileExistsError:
351 if not args.overwrite_results:
352 print(f"result directory {args.result_dir} already exists")
355 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
358 # torch.backends.cudnn.deterministic = True
359 # torch.backends.cudnn.benchmark = False
360 # torch.use_deterministic_algorithms(True)
361 torch.manual_seed(args.seed)
362 if torch.cuda.is_available():
363 torch.cuda.manual_seed_all(args.seed)
365 ######################################################################
369 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
371 if log_file is not None:
372 log_file.write(t + s + "\n")
379 log_string(f"argv {' '.join(sys.argv)}")
382 log_string(f"args.{n} {getattr(args, n)}")
385 ######################################################################
388 def picoclvr_pruner_horizontal_green(p):
389 return not ("green" in p and ("left" in p or "right" in p))
392 picoclvr_pruner_train = (
393 picoclvr_pruner_horizontal_green
394 if args.picocvlr_prune_properties in {"train+eval"}
398 picoclvr_pruner_eval = (
399 (lambda p: not picoclvr_pruner_horizontal_green(p))
400 if args.picocvlr_prune_properties in {"train+eval", "eval"}
404 ######################################################################
406 if args.task == "file":
408 args.filetask_train_file is not None and args.filetask_test_file is not None
409 ), "You have to specify the task train and test files"
410 task = tasks.TaskFromFile(
411 args.filetask_train_file,
412 args.filetask_test_file,
413 nb_train_samples=args.nb_train_samples,
414 nb_test_samples=args.nb_test_samples,
415 batch_size=args.batch_size,
418 args.max_percents_of_test_in_train = 0
420 elif args.task == "byheart":
421 task = tasks.SandBox(
422 problem=problems.ProblemByHeart(),
423 nb_train_samples=args.nb_train_samples,
424 nb_test_samples=args.nb_test_samples,
425 batch_size=args.batch_size,
429 args.max_percents_of_test_in_train = -1
431 elif args.task == "learnop":
432 task = tasks.SandBox(
433 problem=problems.ProblemLearnOperator(),
434 nb_train_samples=args.nb_train_samples,
435 nb_test_samples=args.nb_test_samples,
436 batch_size=args.batch_size,
442 elif args.task == "guessop":
443 task = tasks.SandBox(
444 problem=problems.ProblemGuessOperator(),
445 nb_train_samples=args.nb_train_samples,
446 nb_test_samples=args.nb_test_samples,
447 batch_size=args.batch_size,
453 elif args.task == "twotargets":
454 task = tasks.SandBox(
455 problem=problems.ProblemTwoTargets(),
456 nb_train_samples=args.nb_train_samples,
457 nb_test_samples=args.nb_test_samples,
458 batch_size=args.batch_size,
463 elif args.task == "memory":
464 task = tasks.SandBox(
465 problem=problems.ProblemMemory(),
466 nb_train_samples=args.nb_train_samples,
467 nb_test_samples=args.nb_test_samples,
468 batch_size=args.batch_size,
473 elif args.task == "mixing":
474 task = tasks.SandBox(
475 problem=problems.ProblemMixing(
476 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
478 nb_train_samples=args.nb_train_samples,
479 nb_test_samples=args.nb_test_samples,
480 batch_size=args.batch_size,
485 elif args.task == "addition":
486 task = tasks.SandBox(
487 problem=problems.ProblemAddition(),
488 nb_train_samples=args.nb_train_samples,
489 nb_test_samples=args.nb_test_samples,
490 batch_size=args.batch_size,
495 elif args.task == "picoclvr":
496 task = tasks.PicoCLVR(
497 nb_train_samples=args.nb_train_samples,
498 nb_test_samples=args.nb_test_samples,
499 batch_size=args.batch_size,
500 height=args.picoclvr_height,
501 width=args.picoclvr_width,
502 nb_colors=args.picoclvr_nb_colors,
505 pruner_train=picoclvr_pruner_train,
506 pruner_eval=picoclvr_pruner_eval,
509 elif args.task == "mnist":
511 nb_train_samples=args.nb_train_samples,
512 nb_test_samples=args.nb_test_samples,
513 batch_size=args.batch_size,
517 elif args.task == "maze":
519 nb_train_samples=args.nb_train_samples,
520 nb_test_samples=args.nb_test_samples,
521 batch_size=args.batch_size,
522 height=args.maze_height,
523 width=args.maze_width,
524 nb_walls=args.maze_nb_walls,
528 elif args.task == "snake":
530 nb_train_samples=args.nb_train_samples,
531 nb_test_samples=args.nb_test_samples,
532 batch_size=args.batch_size,
533 height=args.snake_height,
534 width=args.snake_width,
535 nb_colors=args.snake_nb_colors,
536 length=args.snake_length,
537 prompt_length=args.snake_length // 2,
541 elif args.task == "stack":
543 nb_train_samples=args.nb_train_samples,
544 nb_test_samples=args.nb_test_samples,
545 batch_size=args.batch_size,
547 nb_steps=args.stack_nb_steps,
548 nb_stacks=args.stack_nb_stacks,
549 nb_digits=args.stack_nb_digits,
550 fraction_values_for_train=args.stack_fraction_values_for_train,
554 elif args.task == "expr":
556 nb_train_samples=args.nb_train_samples,
557 nb_test_samples=args.nb_test_samples,
558 nb_variables=args.expr_nb_variables,
559 sequence_length=args.expr_sequence_length,
560 operand_max=args.expr_operand_max,
561 result_max=args.expr_result_max,
562 batch_size=args.batch_size,
566 elif args.task == "rpl":
568 nb_train_samples=args.nb_train_samples,
569 nb_test_samples=args.nb_test_samples,
570 batch_size=args.batch_size,
571 nb_starting_values=args.rpl_nb_starting_values,
572 max_input=args.rpl_max_input,
573 prog_len=args.rpl_prog_len,
574 nb_runs=args.rpl_nb_runs,
575 no_prog=args.rpl_no_prog,
580 elif args.task == "grid":
582 nb_train_samples=args.nb_train_samples,
583 nb_test_samples=args.nb_test_samples,
584 batch_size=args.batch_size,
586 fraction_play=args.grid_fraction_play,
591 elif args.task == "qmlp":
593 nb_train_samples=args.nb_train_samples,
594 nb_test_samples=args.nb_test_samples,
595 batch_size=args.batch_size,
596 result_dir=args.result_dir,
602 raise ValueError(f"Unknown task {args.task}")
604 ######################################################################
606 log_string(f"device {device}")
608 vocabulary_size = task.vocabulary_size()
610 log_string(f"vocabulary_size {vocabulary_size}")
612 ##############################
615 vocabulary_size=vocabulary_size,
616 dim_model=args.dim_model,
617 dim_keys=args.dim_keys,
618 dim_hidden=args.dim_hidden,
619 nb_heads=args.nb_heads,
620 nb_blocks=args.nb_blocks,
622 dropout=args.dropout,
627 nb_parameters = sum(p.numel() for p in model.parameters())
628 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
630 ######################################################################
632 nb_epochs_finished = 0
634 if args.no_checkpoint:
635 log_string(f"not trying to load checkpoint.")
639 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
640 checkpoint = torch.load(checkpoint_name)
641 nb_epochs_finished = checkpoint["nb_epochs_finished"]
642 model.load_state_dict(checkpoint["model_state"])
643 torch.set_rng_state(checkpoint["rng_state"])
644 if torch.cuda.is_available():
645 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
647 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
649 except FileNotFoundError:
650 log_string("starting from scratch.")
653 log_string("error when loading the checkpoint.")
656 ######################################################################
658 if args.task == "expr" and args.expr_input_file is not None:
659 task.produce_results(
660 n_epoch=nb_epochs_finished,
662 result_dir=args.result_dir,
664 deterministic_synthesis=args.deterministic_synthesis,
665 input_file=args.expr_input_file,
670 ######################################################################
672 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
674 # Compute the entropy of the training tokens
677 for input in task.batches(split="train"):
678 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
679 token_probas = token_count / token_count.sum()
680 entropy = -torch.xlogy(token_probas, token_probas).sum()
681 train_set_perplexity = math.exp(entropy)
683 ######################################################################
684 # A bit of paranoia never hurts
686 if args.max_percents_of_test_in_train >= 0:
688 def subsets_as_tuples(batches, cs):
690 for batch in batches:
692 s.add(tuple([v.item() for v in x]))
698 nb_test, nb_in_train = 0, 0
699 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
701 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
702 in_train.update(test_subset.intersection(train_subset))
703 nb_in_train += len(in_train)
704 nb_test += len(test_subset)
707 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
711 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
712 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
714 ##############################
716 if args.learning_rate_schedule == "cos":
717 learning_rate_schedule = {}
718 for n_epoch in range(args.nb_epochs):
719 u = n_epoch / args.nb_epochs * math.pi
720 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
725 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
729 learning_rate_schedule = {}
730 learning_rate = args.learning_rate
731 for n_epoch in range(args.nb_epochs):
733 learning_rate = u[n_epoch]
734 learning_rate_schedule[n_epoch] = learning_rate
736 log_string(f"learning_rate_schedule {learning_rate_schedule}")
738 ##############################
742 if nb_epochs_finished >= nb_epochs:
743 task.produce_results(
744 n_epoch=nb_epochs_finished,
746 result_dir=args.result_dir,
748 deterministic_synthesis=args.deterministic_synthesis,
751 time_pred_result = None
753 for n_epoch in range(nb_epochs_finished, nb_epochs):
754 learning_rate = learning_rate_schedule[n_epoch]
756 log_string(f"learning_rate {learning_rate}")
758 if args.optim == "sgd":
759 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
760 elif args.optim == "adam":
761 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
762 elif args.optim == "adamw":
763 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
765 raise ValueError(f"Unknown optimizer {args.optim}.")
769 nb_train_samples, acc_train_loss = 0, 0.0
771 for input in task.batches(split="train"):
772 input = input.to(device)
773 output = model(mygpt.BracketedSequence(input)).x
774 loss = F.cross_entropy(output.transpose(1, 2), input)
775 acc_train_loss += loss.item() * input.size(0)
776 nb_train_samples += input.size(0)
777 nb_samples_seen += input.size(0)
779 optimizer.zero_grad()
783 with torch.autograd.no_grad():
786 nb_test_samples, acc_test_loss = 0, 0.0
788 for input in task.batches(split="test"):
789 input = input.to(device)
791 output = model(mygpt.BracketedSequence(input)).x
792 loss = F.cross_entropy(output.transpose(1, 2), input)
793 acc_test_loss += loss.item() * input.size(0)
794 nb_test_samples += input.size(0)
796 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
797 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
800 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
803 task.produce_results(
806 result_dir=args.result_dir,
808 deterministic_synthesis=args.deterministic_synthesis,
811 time_current_result = datetime.datetime.now()
812 if time_pred_result is not None:
814 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
816 time_pred_result = time_current_result
819 "nb_epochs_finished": n_epoch + 1,
820 "model_state": model.state_dict(),
821 "rng_state": torch.get_rng_state(),
824 if torch.cuda.is_available():
825 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
827 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
828 torch.save(checkpoint, checkpoint_name)
829 log_string(f"saved checkpoint {checkpoint_name}")
831 ######################################################################