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, warnings
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, 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=50)
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 ########################################
61 parser.add_argument("--nb_warmup_iter", type=int, default=100)
63 parser.add_argument("--nb_decay_iter", type=int, default=5000)
65 parser.add_argument("--learning_rate", type=float, default=6e-4)
67 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
69 ########################################
71 parser.add_argument("--model", type=str, default=None)
73 parser.add_argument("--attention", type=str, default=None)
75 parser.add_argument("--dim_model", type=int, default=None)
77 parser.add_argument("--dim_keys", type=int, default=None)
79 parser.add_argument("--dim_hidden", type=int, default=None)
81 parser.add_argument("--nb_heads", type=int, default=None)
83 parser.add_argument("--nb_lines", type=int, default=None)
85 parser.add_argument("--caterpillar_height", type=int, default=None)
87 parser.add_argument("--rho", type=float, default=0.0)
89 parser.add_argument("--dim_rec_v", type=int, default=None)
91 parser.add_argument("--nb_blocks", type=int, default=None)
93 parser.add_argument("--dropout", type=float, default=0.1)
95 ########################################
97 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
99 parser.add_argument("--no_checkpoint", action="store_true", default=False)
101 parser.add_argument("--overwrite_results", action="store_true", default=False)
103 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
105 ##############################
108 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
110 parser.add_argument("--rpl_max_input", type=int, default=9)
112 parser.add_argument("--rpl_prog_len", type=int, default=8)
114 parser.add_argument("--rpl_nb_runs", type=int, default=5)
116 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
118 ##############################
121 parser.add_argument("--grid_size", type=int, default=6)
123 ##############################
126 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
128 parser.add_argument("--picoclvr_height", type=int, default=12)
130 parser.add_argument("--picoclvr_width", type=int, default=16)
132 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
134 ##############################
137 parser.add_argument("--maze_height", type=int, default=13)
139 parser.add_argument("--maze_width", type=int, default=21)
141 parser.add_argument("--maze_nb_walls", type=int, default=15)
143 ##############################
146 parser.add_argument("--snake_height", type=int, default=9)
148 parser.add_argument("--snake_width", type=int, default=12)
150 parser.add_argument("--snake_nb_colors", type=int, default=5)
152 parser.add_argument("--snake_length", type=int, default=200)
154 ##############################
157 parser.add_argument("--stack_nb_steps", type=int, default=100)
159 parser.add_argument("--stack_nb_stacks", type=int, default=3)
161 parser.add_argument("--stack_nb_digits", type=int, default=3)
163 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
165 ##############################
168 parser.add_argument("--expr_nb_variables", type=int, default=5)
170 parser.add_argument("--expr_sequence_length", type=int, default=40)
172 parser.add_argument("--expr_operand_max", type=int, default=9)
174 parser.add_argument("--expr_result_max", type=int, default=99)
176 parser.add_argument("--expr_input_file", type=str, default=None)
178 ##############################
181 parser.add_argument("--memory_len_total", type=int, default=32)
183 ##############################
186 parser.add_argument("--mixing_hard", action="store_true", default=False)
188 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
190 ######################################################################
192 args = parser.parse_args()
194 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
196 if args.result_dir is None:
197 args.result_dir = f"results_{args.task}_{args.model}"
199 ######################################################################
201 default_task_args = {
205 "nb_train_samples": 250000,
206 "nb_test_samples": 10000,
211 "nb_train_samples": 50000,
212 "nb_test_samples": 10000,
217 "nb_train_samples": 2500000,
218 "nb_test_samples": 10000,
223 "nb_train_samples": 250000,
224 "nb_test_samples": 10000,
229 "nb_train_samples": 100000,
230 "nb_test_samples": 1000,
235 "nb_train_samples": 1000000,
236 "nb_test_samples": 10000,
241 "nb_train_samples": 50000,
242 "nb_test_samples": 10000,
247 "nb_train_samples": 100000,
248 "nb_test_samples": 10000,
253 "nb_train_samples": 250000,
254 "nb_test_samples": 10000,
259 "nb_train_samples": 2500000,
260 "nb_test_samples": 10000,
265 "nb_train_samples": 250000,
266 "nb_test_samples": 10000,
271 "nb_train_samples": 100000,
272 "nb_test_samples": 1000,
277 "nb_train_samples": 50000,
278 "nb_test_samples": 10000,
283 "nb_train_samples": 25000,
284 "nb_test_samples": 10000,
289 "nb_train_samples": 250000,
290 "nb_test_samples": 10000,
295 "nb_train_samples": 60000,
296 "nb_test_samples": 10000,
300 if args.task in default_task_args:
301 for k, v in default_task_args[args.task].items():
302 if getattr(args, k) is None:
305 ######################################################################
307 default_model_args = {
318 "attention": "caterpillar",
324 "caterpillar_height": 4,
338 "attention": "caterpillar",
344 "caterpillar_height": 4,
345 "dim_rec_v": 64, # dim_model / nb_heads
357 "attention": "caterpillar",
363 "caterpillar_height": 32,
377 "attention": "caterpillar",
396 "attention": "caterpillar",
407 if args.model in default_model_args:
408 for k, v in default_model_args[args.model].items():
409 if getattr(args, k) is None:
412 raise ValueError(f"Unknown model {args.model}")
414 ######################################################################
417 os.mkdir(args.result_dir)
418 except FileExistsError:
419 if not args.overwrite_results:
420 print(f"result directory {args.result_dir} already exists")
423 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
426 # torch.backends.cudnn.deterministic = True
427 # torch.backends.cudnn.benchmark = False
428 # torch.use_deterministic_algorithms(True)
429 torch.manual_seed(args.seed)
430 if torch.cuda.is_available():
431 torch.cuda.manual_seed_all(args.seed)
433 ######################################################################
437 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
439 if log_file is not None:
440 log_file.write(t + s + "\n")
447 with os.popen("sha256sum *.py") as f:
449 log_string(f"sha256sum {l.strip()}")
451 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
452 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
454 log_string(f"argv {' '.join(sys.argv)}")
457 log_string(f"args.{n} {getattr(args, n)}")
460 ######################################################################
466 # 1) linear warmup for warmup_iter steps
467 if it < args.nb_warmup_iter:
468 return args.learning_rate * it / args.nb_warmup_iter
469 # 2) if it > nb_decay_iter, return min learning rate
470 if it > args.nb_decay_iter:
471 return args.min_learning_rate
472 # 3) in between, use cosine decay down to min learning rate
473 decay_ratio = (it - args.nb_warmup_iter) / (
474 args.nb_decay_iter - args.nb_warmup_iter
476 coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
477 return args.min_learning_rate + coeff * (
478 args.learning_rate - args.min_learning_rate
482 ######################################################################
485 def picoclvr_pruner_horizontal_green(p):
486 return not ("green" in p and ("left" in p or "right" in p))
489 picoclvr_pruner_train = (
490 picoclvr_pruner_horizontal_green
491 if args.picocvlr_prune_properties in {"train+eval"}
495 picoclvr_pruner_eval = (
496 (lambda p: not picoclvr_pruner_horizontal_green(p))
497 if args.picocvlr_prune_properties in {"train+eval", "eval"}
501 ######################################################################
505 if args.task == "byheart":
506 task = tasks.SandBox(
507 problem=problems.ProblemByHeart(),
508 nb_train_samples=args.nb_train_samples,
509 nb_test_samples=args.nb_test_samples,
510 batch_size=args.batch_size,
514 args.max_percents_of_test_in_train = -1
516 elif args.task == "learnop":
517 task = tasks.SandBox(
518 problem=problems.ProblemLearnOperator(),
519 nb_train_samples=args.nb_train_samples,
520 nb_test_samples=args.nb_test_samples,
521 batch_size=args.batch_size,
527 elif args.task == "guessop":
528 task = tasks.SandBox(
529 problem=problems.ProblemGuessOperator(),
530 nb_train_samples=args.nb_train_samples,
531 nb_test_samples=args.nb_test_samples,
532 batch_size=args.batch_size,
538 elif args.task == "twotargets":
539 task = tasks.SandBox(
540 problem=problems.ProblemTwoTargets(),
541 nb_train_samples=args.nb_train_samples,
542 nb_test_samples=args.nb_test_samples,
543 batch_size=args.batch_size,
548 elif args.task == "memory":
549 task = tasks.SandBox(
550 problem=problems.ProblemMemory(len_total=args.memory_len_total),
551 nb_train_samples=args.nb_train_samples,
552 nb_test_samples=args.nb_test_samples,
553 batch_size=args.batch_size,
558 elif args.task == "mixing":
559 task = tasks.SandBox(
560 problem=problems.ProblemMixing(
561 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
563 nb_train_samples=args.nb_train_samples,
564 nb_test_samples=args.nb_test_samples,
565 batch_size=args.batch_size,
570 elif args.task == "addition":
571 task = tasks.SandBox(
572 problem=problems.ProblemAddition(),
573 nb_train_samples=args.nb_train_samples,
574 nb_test_samples=args.nb_test_samples,
575 batch_size=args.batch_size,
580 elif args.task == "picoclvr":
581 task = tasks.PicoCLVR(
582 nb_train_samples=args.nb_train_samples,
583 nb_test_samples=args.nb_test_samples,
584 batch_size=args.batch_size,
585 height=args.picoclvr_height,
586 width=args.picoclvr_width,
587 nb_colors=args.picoclvr_nb_colors,
590 pruner_train=picoclvr_pruner_train,
591 pruner_eval=picoclvr_pruner_eval,
594 elif args.task == "mnist":
596 nb_train_samples=args.nb_train_samples,
597 nb_test_samples=args.nb_test_samples,
598 batch_size=args.batch_size,
602 elif args.task == "maze":
604 nb_train_samples=args.nb_train_samples,
605 nb_test_samples=args.nb_test_samples,
606 batch_size=args.batch_size,
607 height=args.maze_height,
608 width=args.maze_width,
609 nb_walls=args.maze_nb_walls,
613 elif args.task == "snake":
615 nb_train_samples=args.nb_train_samples,
616 nb_test_samples=args.nb_test_samples,
617 batch_size=args.batch_size,
618 height=args.snake_height,
619 width=args.snake_width,
620 nb_colors=args.snake_nb_colors,
621 length=args.snake_length,
622 prompt_length=args.snake_length // 2,
626 elif args.task == "stack":
628 nb_train_samples=args.nb_train_samples,
629 nb_test_samples=args.nb_test_samples,
630 batch_size=args.batch_size,
632 nb_steps=args.stack_nb_steps,
633 nb_stacks=args.stack_nb_stacks,
634 nb_digits=args.stack_nb_digits,
635 fraction_values_for_train=args.stack_fraction_values_for_train,
639 elif args.task == "expr":
641 nb_train_samples=args.nb_train_samples,
642 nb_test_samples=args.nb_test_samples,
643 nb_variables=args.expr_nb_variables,
644 sequence_length=args.expr_sequence_length,
645 operand_max=args.expr_operand_max,
646 result_max=args.expr_result_max,
647 batch_size=args.batch_size,
651 elif args.task == "rpl":
653 nb_train_samples=args.nb_train_samples,
654 nb_test_samples=args.nb_test_samples,
655 batch_size=args.batch_size,
656 nb_starting_values=args.rpl_nb_starting_values,
657 max_input=args.rpl_max_input,
658 prog_len=args.rpl_prog_len,
659 nb_runs=args.rpl_nb_runs,
660 no_prog=args.rpl_no_prog,
665 elif args.task == "grid":
667 nb_train_samples=args.nb_train_samples,
668 nb_test_samples=args.nb_test_samples,
669 batch_size=args.batch_size,
675 elif args.task == "qmlp":
677 nb_train_samples=args.nb_train_samples,
678 nb_test_samples=args.nb_test_samples,
679 batch_size=args.batch_size,
680 result_dir=args.result_dir,
686 raise ValueError(f"Unknown task {args.task}")
688 ######################################################################
690 log_string(f"device {device}")
692 vocabulary_size = task.vocabulary_size()
694 log_string(f"vocabulary_size {vocabulary_size}")
696 ##############################
699 vocabulary_size=vocabulary_size,
700 dim_model=args.dim_model,
701 dim_keys=args.dim_keys,
702 dim_hidden=args.dim_hidden,
703 nb_heads=args.nb_heads,
704 nb_lines=args.nb_lines,
705 caterpillar_height=args.caterpillar_height,
706 dim_rec_v=args.dim_rec_v,
707 nb_blocks=args.nb_blocks,
709 dropout=args.dropout,
710 attention_layer=args.attention,
715 nb_parameters = sum(p.numel() for p in model.parameters())
716 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
718 ######################################################################
720 nb_epochs_finished = 0
722 if args.no_checkpoint:
723 log_string(f"not trying to load checkpoint.")
727 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
728 checkpoint = torch.load(checkpoint_name)
729 nb_epochs_finished = checkpoint["nb_epochs_finished"]
730 model.load_state_dict(checkpoint["model_state"])
731 torch.set_rng_state(checkpoint["rng_state"])
732 if torch.cuda.is_available():
733 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
735 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
737 except FileNotFoundError:
738 log_string("starting from scratch.")
741 log_string("error when loading the checkpoint.")
744 ######################################################################
746 if args.task == "expr" and args.expr_input_file is not None:
747 task.produce_results(
748 n_epoch=nb_epochs_finished,
750 result_dir=args.result_dir,
752 deterministic_synthesis=args.deterministic_synthesis,
753 input_file=args.expr_input_file,
758 ######################################################################
760 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
762 # Compute the entropy of the training tokens
765 for input in task.batches(split="train"):
766 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
767 token_probas = token_count / token_count.sum()
768 entropy = -torch.xlogy(token_probas, token_probas).sum()
769 train_set_perplexity = math.exp(entropy)
771 ######################################################################
772 # A bit of paranoia never hurts
774 if args.max_percents_of_test_in_train >= 0:
776 def subsets_as_tuples(batches, cs):
778 for batch in batches:
780 s.add(tuple([v.item() for v in x]))
786 nb_test, nb_in_train = 0, 0
787 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
789 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
790 in_train.update(test_subset.intersection(train_subset))
791 nb_in_train += len(in_train)
792 nb_test += len(test_subset)
795 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
799 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
800 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
802 ##############################
806 if nb_epochs_finished >= nb_epochs:
807 task.produce_results(
808 n_epoch=nb_epochs_finished,
810 result_dir=args.result_dir,
812 deterministic_synthesis=args.deterministic_synthesis,
815 time_pred_result = None
819 for n_epoch in range(nb_epochs_finished, nb_epochs):
820 if args.optim == "sgd":
821 optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
822 elif args.optim == "adam":
823 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
824 elif args.optim == "adamw":
825 optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
827 raise ValueError(f"Unknown optimizer {args.optim}.")
831 nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
833 for input in task.batches(split="train"):
834 model.reset_inner_loss()
835 input = input.to(device)
837 output = model(mygpt.BracketedSequence(input)).x
838 loss = F.cross_entropy(output.transpose(1, 2), input)
839 inner_loss = model.get_inner_loss()
841 acc_train_loss += loss.item() * input.size(0)
842 acc_train_inner_loss += inner_loss.item() * input.size(0)
844 nb_train_samples += input.size(0)
845 nb_samples_seen += input.size(0)
847 total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
851 for param_group in optimizer.param_groups:
852 param_group["lr"] = lr
854 # log_string(f"learning_rate {lr}")
856 optimizer.zero_grad()
857 total_loss.backward()
860 with torch.autograd.no_grad():
863 nb_test_samples, acc_test_loss = 0, 0.0
865 for input in task.batches(split="test"):
866 input = input.to(device)
868 output = model(mygpt.BracketedSequence(input)).x
869 loss = F.cross_entropy(output.transpose(1, 2), input)
870 acc_test_loss += loss.item() * input.size(0)
871 nb_test_samples += input.size(0)
874 f"loss {n_epoch} train_loss {acc_train_loss/nb_train_samples} train_inner_loss {acc_train_inner_loss/nb_train_samples} test_prediction {acc_test_loss/nb_test_samples}"
877 task.produce_results(
880 result_dir=args.result_dir,
882 deterministic_synthesis=args.deterministic_synthesis,
885 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
886 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
889 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
892 time_current_result = datetime.datetime.now()
893 if time_pred_result is not None:
895 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
897 time_pred_result = time_current_result
900 "nb_epochs_finished": n_epoch + 1,
901 "model_state": model.state_dict(),
902 "rng_state": torch.get_rng_state(),
905 if torch.cuda.is_available():
906 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
908 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
909 torch.save(checkpoint, checkpoint_name)
910 log_string(f"saved checkpoint {checkpoint_name}")
912 ######################################################################