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 ######################################################################
22 if x in {"1", "true", "yes"}:
24 elif x in {"0", "false", "no"}:
30 parser = argparse.ArgumentParser(
31 description="An implementation of GPT with cache.",
32 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
39 help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
42 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
44 parser.add_argument("--result_dir", type=str, default=None)
46 parser.add_argument("--seed", type=int, default=0)
48 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
50 parser.add_argument("--force_cpu", type=str2bool, default=False)
52 ########################################
54 parser.add_argument("--nb_epochs", type=int, default=50)
56 parser.add_argument("--batch_size", type=int, default=None)
58 parser.add_argument("--nb_train_samples", type=int, default=None)
60 parser.add_argument("--nb_test_samples", type=int, default=None)
62 parser.add_argument("--optim", type=str, default="adam")
64 ########################################
66 parser.add_argument("--nb_warmup_iter", type=int, default=100)
68 parser.add_argument("--nb_decay_iter", type=int, default=5000)
70 parser.add_argument("--learning_rate", type=float, default=6e-4)
72 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
76 parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
78 parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
80 parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
82 parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
84 ########################################
86 parser.add_argument("--model", type=str, default=None)
88 parser.add_argument("--attention", type=str, default=None)
90 parser.add_argument("--dim_model", type=int, default=None)
92 parser.add_argument("--dim_keys", type=int, default=None)
94 parser.add_argument("--dim_hidden", type=int, default=None)
96 parser.add_argument("--nb_heads", type=int, default=None)
98 parser.add_argument("--nb_lines", type=int, default=None)
100 parser.add_argument("--caterpillar_height", type=int, default=None)
102 parser.add_argument("--rho", type=float, default=0.0)
104 parser.add_argument("--nb_blocks", type=int, default=None)
106 parser.add_argument("--dropout", type=float, default=0.1)
108 ########################################
110 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
112 parser.add_argument("--no_checkpoint", action="store_true", default=False)
114 parser.add_argument("--continue_training", action="store_true", default=False)
116 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
118 ##############################
121 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
123 parser.add_argument("--rpl_max_input", type=int, default=9)
125 parser.add_argument("--rpl_prog_len", type=int, default=8)
127 parser.add_argument("--rpl_nb_runs", type=int, default=5)
129 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
131 ##############################
134 parser.add_argument("--grid_size", type=int, default=6)
136 parser.add_argument("--grid_nb_colors", type=int, default=6)
138 parser.add_argument("--grid_nb_shapes", type=int, default=6)
140 ##############################
143 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
145 parser.add_argument("--picoclvr_height", type=int, default=12)
147 parser.add_argument("--picoclvr_width", type=int, default=16)
149 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
151 ##############################
154 parser.add_argument("--maze_height", type=int, default=13)
156 parser.add_argument("--maze_width", type=int, default=21)
158 parser.add_argument("--maze_nb_walls", type=int, default=15)
160 ##############################
163 parser.add_argument("--snake_height", type=int, default=9)
165 parser.add_argument("--snake_width", type=int, default=12)
167 parser.add_argument("--snake_nb_colors", type=int, default=5)
169 parser.add_argument("--snake_length", type=int, default=200)
171 ##############################
174 parser.add_argument("--stack_nb_steps", type=int, default=100)
176 parser.add_argument("--stack_nb_stacks", type=int, default=3)
178 parser.add_argument("--stack_nb_digits", type=int, default=3)
180 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
182 ##############################
185 parser.add_argument("--expr_nb_variables", type=int, default=5)
187 parser.add_argument("--expr_sequence_length", type=int, default=40)
189 parser.add_argument("--expr_operand_max", type=int, default=9)
191 parser.add_argument("--expr_result_max", type=int, default=99)
193 parser.add_argument("--expr_input_file", type=str, default=None)
195 ##############################
198 parser.add_argument("--memory_len_total", type=int, default=32)
200 ##############################
203 parser.add_argument("--mixing_hard", action="store_true", default=False)
205 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
207 ######################################################################
209 # args = parser.parse_args()
211 args, sup_args = parser.parse_known_args()
213 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
215 if args.result_dir is None:
216 args.result_dir = f"results_{args.task}_{args.model}"
218 ######################################################################
220 if not args.force_cpu and torch.cuda.is_available():
221 device = torch.device("cuda")
222 torch.backends.cuda.matmul.allow_tf32 = True
224 device = torch.device("cpu")
226 ######################################################################
228 default_task_args = {
232 "nb_train_samples": 250000,
233 "nb_test_samples": 10000,
238 "nb_train_samples": 50000,
239 "nb_test_samples": 10000,
244 "nb_train_samples": 2500000,
245 "nb_test_samples": 10000,
250 "nb_train_samples": 250000,
251 "nb_test_samples": 10000,
256 "nb_train_samples": 100000,
257 "nb_test_samples": 1000,
262 "nb_train_samples": 1000000,
263 "nb_test_samples": 10000,
268 "nb_train_samples": 50000,
269 "nb_test_samples": 10000,
274 "nb_train_samples": 100000,
275 "nb_test_samples": 10000,
280 "nb_train_samples": 250000,
281 "nb_test_samples": 10000,
286 "nb_train_samples": 2500000,
287 "nb_test_samples": 10000,
292 "nb_train_samples": 250000,
293 "nb_test_samples": 10000,
298 "nb_train_samples": 100000,
299 "nb_test_samples": 1000,
304 "nb_train_samples": 50000,
305 "nb_test_samples": 10000,
310 "nb_train_samples": 25000,
311 "nb_test_samples": 10000,
316 "nb_train_samples": 250000,
317 "nb_test_samples": 10000,
322 "nb_train_samples": 60000,
323 "nb_test_samples": 10000,
327 if args.task in default_task_args:
328 for k, v in default_task_args[args.task].items():
329 if getattr(args, k) is None:
332 ######################################################################
334 default_model_args = {
344 "attention": "caterpillar",
350 "caterpillar_height": 4,
362 "attention": "caterpillar",
368 "caterpillar_height": 4,
380 "attention": "caterpillar",
386 "caterpillar_height": 32,
398 "attention": "caterpillar",
415 "attention": "caterpillar",
425 if args.model in default_model_args:
426 for k, v in default_model_args[args.model].items():
427 if getattr(args, k) is None:
430 raise ValueError(f"Unknown model {args.model}")
432 ######################################################################
435 os.mkdir(args.result_dir)
436 except FileExistsError:
437 if not args.continue_training:
438 print(f"result directory {args.result_dir} already exists")
441 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
443 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
446 # torch.backends.cudnn.deterministic = True
447 # torch.backends.cudnn.benchmark = False
448 # torch.use_deterministic_algorithms(True)
449 torch.manual_seed(args.seed)
450 if torch.cuda.is_available():
451 torch.cuda.manual_seed_all(args.seed)
453 ######################################################################
457 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
459 if log_file is not None:
460 log_file.write(t + s + "\n")
467 with os.popen("sha256sum *.py") as f:
469 log_string(f"sha256sum {l.strip()}")
471 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
472 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
474 log_string(f"argv {' '.join(sys.argv)}")
477 log_string(f"args.{n} {getattr(args, n)}")
479 for k, v in sup_args.items():
480 log_string(f'sup_args["{k}"] "{v}"')
483 ######################################################################
486 def get_lr(n_epoch, it):
487 if args.legacy_lr_schedule:
488 # my crude scheduling to compare to previous baseline, added
491 if it < args.nb_warmup_iter:
492 return args.legacy_large_lr * it / args.nb_warmup_iter
493 elif n_epoch < args.legacy_nb_epoch_large_lr:
494 return args.legacy_large_lr
496 return args.legacy_small_lr
500 # 1) linear warmup for warmup_iter steps
501 if it < args.nb_warmup_iter:
502 return args.learning_rate * it / args.nb_warmup_iter
503 # 2) if it > nb_decay_iter, return min learning rate
504 if it > args.nb_decay_iter:
505 return args.min_learning_rate
506 # 3) in between, use cosine decay down to min learning rate
507 decay_ratio = (it - args.nb_warmup_iter) / (
508 args.nb_decay_iter - args.nb_warmup_iter
510 coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
511 return args.min_learning_rate + coeff * (
512 args.learning_rate - args.min_learning_rate
516 ######################################################################
519 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
522 def picoclvr_pruner_horizontal_green(p):
523 return not ("green" in p and ("left" in p or "right" in p))
526 picoclvr_pruner_train = (
527 picoclvr_pruner_horizontal_green
528 if args.picocvlr_prune_properties in {"train+eval"}
532 picoclvr_pruner_eval = (
533 (lambda p: not picoclvr_pruner_horizontal_green(p))
534 if args.picocvlr_prune_properties in {"train+eval", "eval"}
538 ######################################################################
542 if args.task == "byheart":
543 task = tasks.SandBox(
544 problem=problems.ProblemByHeart(),
545 nb_train_samples=args.nb_train_samples,
546 nb_test_samples=args.nb_test_samples,
547 batch_size=args.batch_size,
551 args.max_percents_of_test_in_train = -1
553 elif args.task == "learnop":
554 task = tasks.SandBox(
555 problem=problems.ProblemLearnOperator(),
556 nb_train_samples=args.nb_train_samples,
557 nb_test_samples=args.nb_test_samples,
558 batch_size=args.batch_size,
564 elif args.task == "guessop":
565 task = tasks.SandBox(
566 problem=problems.ProblemGuessOperator(),
567 nb_train_samples=args.nb_train_samples,
568 nb_test_samples=args.nb_test_samples,
569 batch_size=args.batch_size,
575 elif args.task == "twotargets":
576 task = tasks.SandBox(
577 problem=problems.ProblemTwoTargets(),
578 nb_train_samples=args.nb_train_samples,
579 nb_test_samples=args.nb_test_samples,
580 batch_size=args.batch_size,
585 elif args.task == "memory":
586 task = tasks.SandBox(
587 problem=problems.ProblemMemory(len_total=args.memory_len_total),
588 nb_train_samples=args.nb_train_samples,
589 nb_test_samples=args.nb_test_samples,
590 batch_size=args.batch_size,
595 elif args.task == "mixing":
596 task = tasks.SandBox(
597 problem=problems.ProblemMixing(
598 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
600 nb_train_samples=args.nb_train_samples,
601 nb_test_samples=args.nb_test_samples,
602 batch_size=args.batch_size,
607 elif args.task == "addition":
608 task = tasks.SandBox(
609 problem=problems.ProblemAddition(),
610 nb_train_samples=args.nb_train_samples,
611 nb_test_samples=args.nb_test_samples,
612 batch_size=args.batch_size,
617 elif args.task == "picoclvr":
618 task = tasks.PicoCLVR(
619 nb_train_samples=args.nb_train_samples,
620 nb_test_samples=args.nb_test_samples,
621 batch_size=args.batch_size,
622 height=args.picoclvr_height,
623 width=args.picoclvr_width,
624 nb_colors=args.picoclvr_nb_colors,
627 pruner_train=picoclvr_pruner_train,
628 pruner_eval=picoclvr_pruner_eval,
631 elif args.task == "mnist":
633 nb_train_samples=args.nb_train_samples,
634 nb_test_samples=args.nb_test_samples,
635 batch_size=args.batch_size,
639 elif args.task == "maze":
641 nb_train_samples=args.nb_train_samples,
642 nb_test_samples=args.nb_test_samples,
643 batch_size=args.batch_size,
644 height=args.maze_height,
645 width=args.maze_width,
646 nb_walls=args.maze_nb_walls,
650 elif args.task == "snake":
652 nb_train_samples=args.nb_train_samples,
653 nb_test_samples=args.nb_test_samples,
654 batch_size=args.batch_size,
655 height=args.snake_height,
656 width=args.snake_width,
657 nb_colors=args.snake_nb_colors,
658 length=args.snake_length,
659 prompt_length=args.snake_length // 2,
663 elif args.task == "stack":
665 nb_train_samples=args.nb_train_samples,
666 nb_test_samples=args.nb_test_samples,
667 batch_size=args.batch_size,
669 nb_steps=args.stack_nb_steps,
670 nb_stacks=args.stack_nb_stacks,
671 nb_digits=args.stack_nb_digits,
672 fraction_values_for_train=args.stack_fraction_values_for_train,
676 elif args.task == "expr":
678 nb_train_samples=args.nb_train_samples,
679 nb_test_samples=args.nb_test_samples,
680 nb_variables=args.expr_nb_variables,
681 sequence_length=args.expr_sequence_length,
682 operand_max=args.expr_operand_max,
683 result_max=args.expr_result_max,
684 batch_size=args.batch_size,
688 elif args.task == "rpl":
690 nb_train_samples=args.nb_train_samples,
691 nb_test_samples=args.nb_test_samples,
692 batch_size=args.batch_size,
693 nb_starting_values=args.rpl_nb_starting_values,
694 max_input=args.rpl_max_input,
695 prog_len=args.rpl_prog_len,
696 nb_runs=args.rpl_nb_runs,
697 no_prog=args.rpl_no_prog,
702 elif args.task == "grid":
704 nb_train_samples=args.nb_train_samples,
705 nb_test_samples=args.nb_test_samples,
706 batch_size=args.batch_size,
708 nb_shapes=args.grid_nb_shapes,
709 nb_colors=args.grid_nb_colors,
714 elif args.task == "qmlp":
716 nb_train_samples=args.nb_train_samples,
717 nb_test_samples=args.nb_test_samples,
718 batch_size=args.batch_size,
719 result_dir=args.result_dir,
725 raise ValueError(f"Unknown task {args.task}")
727 ######################################################################
729 log_string(f"device {device}")
731 vocabulary_size = task.vocabulary_size()
733 log_string(f"vocabulary_size {vocabulary_size}")
735 ##############################
738 vocabulary_size=vocabulary_size,
739 dim_model=args.dim_model,
740 dim_keys=args.dim_keys,
741 dim_hidden=args.dim_hidden,
742 nb_heads=args.nb_heads,
743 nb_lines=args.nb_lines,
744 caterpillar_height=args.caterpillar_height,
745 nb_blocks=args.nb_blocks,
747 dropout=args.dropout,
748 attention_layer=args.attention,
755 nb_parameters = sum(p.numel() for p in model.parameters())
756 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
758 ######################################################################
760 nb_epochs_finished = 0
762 if args.no_checkpoint:
763 log_string(f"not trying to load checkpoint.")
767 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
768 checkpoint = torch.load(checkpoint_name)
769 nb_epochs_finished = checkpoint["nb_epochs_finished"]
770 model.load_state_dict(checkpoint["model_state"])
771 torch.set_rng_state(checkpoint["rng_state"])
772 if torch.cuda.is_available():
773 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
775 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
777 except FileNotFoundError:
778 log_string("starting from scratch.")
781 log_string("error when loading the checkpoint.")
784 ######################################################################
786 if args.task == "expr" and args.expr_input_file is not None:
787 task.produce_results(
788 n_epoch=nb_epochs_finished,
790 result_dir=args.result_dir,
792 deterministic_synthesis=args.deterministic_synthesis,
793 input_file=args.expr_input_file,
798 ######################################################################
800 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
802 # Compute the entropy of the training tokens
805 for input in task.batches(split="train"):
806 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
807 token_probas = token_count / token_count.sum()
808 entropy = -torch.xlogy(token_probas, token_probas).sum()
809 train_set_perplexity = math.exp(entropy)
811 ######################################################################
812 # A bit of paranoia never hurts
814 if args.max_percents_of_test_in_train >= 0:
816 def subsets_as_tuples(batches, cs):
818 for batch in batches:
820 s.add(tuple([v.item() for v in x]))
826 nb_test, nb_in_train = 0, 0
827 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
829 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
830 in_train.update(test_subset.intersection(train_subset))
831 nb_in_train += len(in_train)
832 nb_test += len(test_subset)
835 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
839 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
840 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
842 ##############################
844 if "calibrate" in sup_args:
845 for input in task.batches(split="train", desc="calibrate"):
846 input = input.to(device)
847 output = model(mygpt.BracketedSequence(input)).x
849 for n, m in model.named_modules():
852 if isinstance(x, mygpt.Calibrator):
853 print(f"####### ${n} | ${a} ########################")
854 mean, std = x.moments()
855 print("mean\n", mean, "\n")
856 print("std\n", std, "\n")
857 print(f"############################################\n\n")
861 ##############################
865 if nb_epochs_finished >= nb_epochs:
866 task.produce_results(
867 n_epoch=nb_epochs_finished,
869 result_dir=args.result_dir,
871 deterministic_synthesis=args.deterministic_synthesis,
874 time_pred_result = datetime.datetime.now()
880 for n_epoch in range(nb_epochs_finished, nb_epochs):
881 if args.optim == "sgd":
882 optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
883 elif args.optim == "adam":
884 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
885 elif args.optim == "adamw":
886 optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
888 raise ValueError(f"Unknown optimizer {args.optim}.")
892 nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
894 for input in task.batches(split="train"):
895 model.reset_inner_loss()
896 input = input.to(device)
898 output = model(mygpt.BracketedSequence(input)).x
899 loss = F.cross_entropy(output.transpose(1, 2), input)
900 inner_loss = model.get_inner_loss()
902 acc_train_loss += loss.item() * input.size(0)
903 acc_train_inner_loss += inner_loss.item() * input.size(0)
905 nb_train_samples += input.size(0)
906 nb_samples_seen += input.size(0)
908 total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
911 lr = get_lr(n_epoch, it)
912 for param_group in optimizer.param_groups:
913 param_group["lr"] = lr
915 # log_string(f"learning_rate {lr}")
917 optimizer.zero_grad()
918 total_loss.backward()
921 grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
923 loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
927 with torch.autograd.no_grad():
930 nb_test_samples, acc_test_loss = 0, 0.0
932 for input in task.batches(split="test"):
933 input = input.to(device)
935 output = model(mygpt.BracketedSequence(input)).x
936 loss = F.cross_entropy(output.transpose(1, 2), input)
937 acc_test_loss += loss.item() * input.size(0)
938 nb_test_samples += input.size(0)
941 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}"
944 task.produce_results(
947 result_dir=args.result_dir,
949 deterministic_synthesis=args.deterministic_synthesis,
952 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
953 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
956 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
959 time_current_result = datetime.datetime.now()
961 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
963 time_pred_result = time_current_result
966 "nb_epochs_finished": n_epoch + 1,
967 "model_state": model.state_dict(),
968 "rng_state": torch.get_rng_state(),
971 if torch.cuda.is_available():
972 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
974 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
975 torch.save(checkpoint, checkpoint_name)
976 log_string(f"saved checkpoint {checkpoint_name}")
978 ######################################################################