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("--gate_dropout_proba", type=float, default=0.0)
104 parser.add_argument("--gate_dropout_sync", type=str2bool, default=True)
106 parser.add_argument("--gate_dropout_replace", type=str2bool, default=True)
108 parser.add_argument("--rho_inner_loss", type=float, default=0.0)
110 parser.add_argument("--nb_blocks", type=int, default=None)
112 parser.add_argument("--dropout", type=float, default=0.1)
114 ########################################
116 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
118 parser.add_argument("--no_checkpoint", action="store_true", default=False)
120 parser.add_argument("--continue_training", action="store_true", default=False)
122 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
124 ##############################
127 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
129 parser.add_argument("--rpl_max_input", type=int, default=9)
131 parser.add_argument("--rpl_prog_len", type=int, default=8)
133 parser.add_argument("--rpl_nb_runs", type=int, default=5)
135 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
137 ##############################
140 parser.add_argument("--grid_size", type=int, default=6)
142 parser.add_argument("--grid_nb_colors", type=int, default=6)
144 parser.add_argument("--grid_nb_shapes", type=int, default=6)
146 ##############################
149 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
151 parser.add_argument("--picoclvr_height", type=int, default=12)
153 parser.add_argument("--picoclvr_width", type=int, default=16)
155 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
157 ##############################
160 parser.add_argument("--maze_height", type=int, default=13)
162 parser.add_argument("--maze_width", type=int, default=21)
164 parser.add_argument("--maze_nb_walls", type=int, default=15)
166 ##############################
169 parser.add_argument("--snake_height", type=int, default=9)
171 parser.add_argument("--snake_width", type=int, default=12)
173 parser.add_argument("--snake_nb_colors", type=int, default=5)
175 parser.add_argument("--snake_length", type=int, default=200)
177 ##############################
180 parser.add_argument("--stack_nb_steps", type=int, default=100)
182 parser.add_argument("--stack_nb_stacks", type=int, default=3)
184 parser.add_argument("--stack_nb_digits", type=int, default=3)
186 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
188 ##############################
191 parser.add_argument("--expr_nb_variables", type=int, default=5)
193 parser.add_argument("--expr_sequence_length", type=int, default=40)
195 parser.add_argument("--expr_operand_max", type=int, default=9)
197 parser.add_argument("--expr_result_max", type=int, default=99)
199 parser.add_argument("--expr_input_file", type=str, default=None)
201 ##############################
204 parser.add_argument("--memory_len_total", type=int, default=32)
206 ##############################
209 parser.add_argument("--mixing_hard", action="store_true", default=False)
211 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
213 ######################################################################
215 # args = parser.parse_args()
217 args, sup_args = parser.parse_known_args()
219 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
221 if args.result_dir is None:
222 args.result_dir = f"results_{args.task}_{args.model}"
224 ######################################################################
226 if not args.force_cpu and torch.cuda.is_available():
227 device = torch.device("cuda")
228 torch.backends.cuda.matmul.allow_tf32 = True
230 device = torch.device("cpu")
232 ######################################################################
234 default_task_args = {
238 "nb_train_samples": 250000,
239 "nb_test_samples": 10000,
244 "nb_train_samples": 50000,
245 "nb_test_samples": 10000,
250 "nb_train_samples": 2500000,
251 "nb_test_samples": 10000,
256 "nb_train_samples": 250000,
257 "nb_test_samples": 10000,
262 "nb_train_samples": 100000,
263 "nb_test_samples": 1000,
268 "nb_train_samples": 1000000,
269 "nb_test_samples": 10000,
274 "nb_train_samples": 50000,
275 "nb_test_samples": 10000,
280 "nb_train_samples": 100000,
281 "nb_test_samples": 10000,
286 "nb_train_samples": 250000,
287 "nb_test_samples": 10000,
292 "nb_train_samples": 2500000,
293 "nb_test_samples": 10000,
298 "nb_train_samples": 250000,
299 "nb_test_samples": 10000,
304 "nb_train_samples": 100000,
305 "nb_test_samples": 1000,
310 "nb_train_samples": 50000,
311 "nb_test_samples": 10000,
316 "nb_train_samples": 25000,
317 "nb_test_samples": 10000,
322 "nb_train_samples": 250000,
323 "nb_test_samples": 10000,
328 "nb_train_samples": 60000,
329 "nb_test_samples": 10000,
333 if args.task in default_task_args:
334 for k, v in default_task_args[args.task].items():
335 if getattr(args, k) is None:
338 ######################################################################
340 default_model_args = {
350 "attention": "caterpillar",
356 "caterpillar_height": 4,
368 "attention": "caterpillar",
374 "caterpillar_height": 4,
386 "attention": "caterpillar",
392 "caterpillar_height": 32,
404 "attention": "caterpillar",
421 "attention": "caterpillar",
431 if args.model in default_model_args:
432 for k, v in default_model_args[args.model].items():
433 if getattr(args, k) is None:
436 raise ValueError(f"Unknown model {args.model}")
438 ######################################################################
441 os.mkdir(args.result_dir)
442 except FileExistsError:
443 if not args.continue_training:
444 print(f"result directory {args.result_dir} already exists")
447 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
449 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
452 # torch.backends.cudnn.deterministic = True
453 # torch.backends.cudnn.benchmark = False
454 # torch.use_deterministic_algorithms(True)
455 torch.manual_seed(args.seed)
456 if torch.cuda.is_available():
457 torch.cuda.manual_seed_all(args.seed)
459 ######################################################################
463 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
465 if log_file is not None:
466 log_file.write(t + s + "\n")
473 with os.popen("sha256sum *.py") as f:
475 log_string(f"sha256sum {l.strip()}")
477 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
478 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
480 log_string(f"argv {' '.join(sys.argv)}")
483 log_string(f"args.{n} {getattr(args, n)}")
485 for k, v in sup_args.items():
486 log_string(f'sup_args["{k}"] "{v}"')
489 ######################################################################
492 def get_lr(n_epoch, it):
493 if args.legacy_lr_schedule:
494 # my crude scheduling to compare to previous baseline, added
497 if it < args.nb_warmup_iter:
498 return args.legacy_large_lr * it / args.nb_warmup_iter
499 elif n_epoch < args.legacy_nb_epoch_large_lr:
500 return args.legacy_large_lr
502 return args.legacy_small_lr
506 # 1) linear warmup for warmup_iter steps
507 if it < args.nb_warmup_iter:
508 return args.learning_rate * it / args.nb_warmup_iter
509 # 2) if it > nb_decay_iter, return min learning rate
510 if it > args.nb_decay_iter:
511 return args.min_learning_rate
512 # 3) in between, use cosine decay down to min learning rate
513 decay_ratio = (it - args.nb_warmup_iter) / (
514 args.nb_decay_iter - args.nb_warmup_iter
516 coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
517 return args.min_learning_rate + coeff * (
518 args.learning_rate - args.min_learning_rate
522 ######################################################################
525 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
528 def picoclvr_pruner_horizontal_green(p):
529 return not ("green" in p and ("left" in p or "right" in p))
532 picoclvr_pruner_train = (
533 picoclvr_pruner_horizontal_green
534 if args.picocvlr_prune_properties in {"train+eval"}
538 picoclvr_pruner_eval = (
539 (lambda p: not picoclvr_pruner_horizontal_green(p))
540 if args.picocvlr_prune_properties in {"train+eval", "eval"}
544 ######################################################################
548 if args.task == "byheart":
549 task = tasks.SandBox(
550 problem=problems.ProblemByHeart(),
551 nb_train_samples=args.nb_train_samples,
552 nb_test_samples=args.nb_test_samples,
553 batch_size=args.batch_size,
557 args.max_percents_of_test_in_train = -1
559 elif args.task == "learnop":
560 task = tasks.SandBox(
561 problem=problems.ProblemLearnOperator(),
562 nb_train_samples=args.nb_train_samples,
563 nb_test_samples=args.nb_test_samples,
564 batch_size=args.batch_size,
570 elif args.task == "guessop":
571 task = tasks.SandBox(
572 problem=problems.ProblemGuessOperator(),
573 nb_train_samples=args.nb_train_samples,
574 nb_test_samples=args.nb_test_samples,
575 batch_size=args.batch_size,
581 elif args.task == "twotargets":
582 task = tasks.SandBox(
583 problem=problems.ProblemTwoTargets(),
584 nb_train_samples=args.nb_train_samples,
585 nb_test_samples=args.nb_test_samples,
586 batch_size=args.batch_size,
591 elif args.task == "memory":
592 task = tasks.SandBox(
593 problem=problems.ProblemMemory(len_total=args.memory_len_total),
594 nb_train_samples=args.nb_train_samples,
595 nb_test_samples=args.nb_test_samples,
596 batch_size=args.batch_size,
601 elif args.task == "mixing":
602 task = tasks.SandBox(
603 problem=problems.ProblemMixing(
604 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
606 nb_train_samples=args.nb_train_samples,
607 nb_test_samples=args.nb_test_samples,
608 batch_size=args.batch_size,
613 elif args.task == "addition":
614 task = tasks.SandBox(
615 problem=problems.ProblemAddition(),
616 nb_train_samples=args.nb_train_samples,
617 nb_test_samples=args.nb_test_samples,
618 batch_size=args.batch_size,
623 elif args.task == "picoclvr":
624 task = tasks.PicoCLVR(
625 nb_train_samples=args.nb_train_samples,
626 nb_test_samples=args.nb_test_samples,
627 batch_size=args.batch_size,
628 height=args.picoclvr_height,
629 width=args.picoclvr_width,
630 nb_colors=args.picoclvr_nb_colors,
633 pruner_train=picoclvr_pruner_train,
634 pruner_eval=picoclvr_pruner_eval,
637 elif args.task == "mnist":
639 nb_train_samples=args.nb_train_samples,
640 nb_test_samples=args.nb_test_samples,
641 batch_size=args.batch_size,
645 elif args.task == "maze":
647 nb_train_samples=args.nb_train_samples,
648 nb_test_samples=args.nb_test_samples,
649 batch_size=args.batch_size,
650 height=args.maze_height,
651 width=args.maze_width,
652 nb_walls=args.maze_nb_walls,
656 elif args.task == "snake":
658 nb_train_samples=args.nb_train_samples,
659 nb_test_samples=args.nb_test_samples,
660 batch_size=args.batch_size,
661 height=args.snake_height,
662 width=args.snake_width,
663 nb_colors=args.snake_nb_colors,
664 length=args.snake_length,
665 prompt_length=args.snake_length // 2,
669 elif args.task == "stack":
671 nb_train_samples=args.nb_train_samples,
672 nb_test_samples=args.nb_test_samples,
673 batch_size=args.batch_size,
675 nb_steps=args.stack_nb_steps,
676 nb_stacks=args.stack_nb_stacks,
677 nb_digits=args.stack_nb_digits,
678 fraction_values_for_train=args.stack_fraction_values_for_train,
682 elif args.task == "expr":
684 nb_train_samples=args.nb_train_samples,
685 nb_test_samples=args.nb_test_samples,
686 nb_variables=args.expr_nb_variables,
687 sequence_length=args.expr_sequence_length,
688 operand_max=args.expr_operand_max,
689 result_max=args.expr_result_max,
690 batch_size=args.batch_size,
694 elif args.task == "rpl":
696 nb_train_samples=args.nb_train_samples,
697 nb_test_samples=args.nb_test_samples,
698 batch_size=args.batch_size,
699 nb_starting_values=args.rpl_nb_starting_values,
700 max_input=args.rpl_max_input,
701 prog_len=args.rpl_prog_len,
702 nb_runs=args.rpl_nb_runs,
703 no_prog=args.rpl_no_prog,
708 elif args.task == "grid":
710 nb_train_samples=args.nb_train_samples,
711 nb_test_samples=args.nb_test_samples,
712 batch_size=args.batch_size,
714 nb_shapes=args.grid_nb_shapes,
715 nb_colors=args.grid_nb_colors,
720 elif args.task == "qmlp":
722 nb_train_samples=args.nb_train_samples,
723 nb_test_samples=args.nb_test_samples,
724 batch_size=args.batch_size,
725 result_dir=args.result_dir,
731 raise ValueError(f"Unknown task {args.task}")
733 ######################################################################
735 log_string(f"device {device}")
737 vocabulary_size = task.vocabulary_size()
739 log_string(f"vocabulary_size {vocabulary_size}")
741 ##############################
744 vocabulary_size=vocabulary_size,
745 dim_model=args.dim_model,
746 dim_keys=args.dim_keys,
747 dim_hidden=args.dim_hidden,
748 nb_heads=args.nb_heads,
749 nb_lines=args.nb_lines,
750 caterpillar_height=args.caterpillar_height,
751 nb_blocks=args.nb_blocks,
753 dropout=args.dropout,
754 attention_layer=args.attention,
761 nb_parameters = sum(p.numel() for p in model.parameters())
762 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
764 ######################################################################
766 nb_epochs_finished = 0
768 if args.no_checkpoint:
769 log_string(f"not trying to load checkpoint.")
773 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
774 checkpoint = torch.load(checkpoint_name)
775 nb_epochs_finished = checkpoint["nb_epochs_finished"]
776 model.load_state_dict(checkpoint["model_state"])
777 torch.set_rng_state(checkpoint["rng_state"])
778 if torch.cuda.is_available():
779 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
781 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
783 except FileNotFoundError:
784 log_string("starting from scratch.")
787 log_string("error when loading the checkpoint.")
790 ######################################################################
792 if args.task == "expr" and args.expr_input_file is not None:
793 task.produce_results(
794 n_epoch=nb_epochs_finished,
796 result_dir=args.result_dir,
798 deterministic_synthesis=args.deterministic_synthesis,
799 input_file=args.expr_input_file,
804 ######################################################################
806 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
808 # Compute the entropy of the training tokens
811 for input in task.batches(split="train"):
812 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
813 token_probas = token_count / token_count.sum()
814 entropy = -torch.xlogy(token_probas, token_probas).sum()
815 train_set_perplexity = math.exp(entropy)
817 ######################################################################
818 # A bit of paranoia never hurts
820 if args.max_percents_of_test_in_train >= 0:
822 def subsets_as_tuples(batches, cs):
824 for batch in batches:
826 s.add(tuple([v.item() for v in x]))
832 nb_test, nb_in_train = 0, 0
833 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
835 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
836 in_train.update(test_subset.intersection(train_subset))
837 nb_in_train += len(in_train)
838 nb_test += len(test_subset)
841 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
845 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
846 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
848 ##############################
850 if "calibrate" in sup_args:
851 for input in task.batches(split="train", desc="calibrate"):
852 input = input.to(device)
853 output = model(mygpt.BracketedSequence(input)).x
855 for n, m in model.named_modules():
858 if isinstance(x, mygpt.Calibrator):
859 print(f"####### ${n} | ${a} ########################")
860 mean, std = x.moments()
861 print("mean\n", mean, "\n")
862 print("std\n", std, "\n")
863 print(f"############################################\n\n")
867 ##############################
871 if nb_epochs_finished >= nb_epochs:
872 task.produce_results(
873 n_epoch=nb_epochs_finished,
875 result_dir=args.result_dir,
877 deterministic_synthesis=args.deterministic_synthesis,
880 time_pred_result = datetime.datetime.now()
886 for n_epoch in range(nb_epochs_finished, nb_epochs):
887 if args.optim == "sgd":
888 optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
889 elif args.optim == "adam":
890 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
891 elif args.optim == "adamw":
892 optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
894 raise ValueError(f"Unknown optimizer {args.optim}.")
898 nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
900 for input in task.batches(split="train"):
901 model.reset_inner_loss()
902 input = input.to(device)
904 output = model(mygpt.BracketedSequence(input)).x
905 loss = F.cross_entropy(output.transpose(1, 2), input)
906 inner_loss = model.get_inner_loss()
908 acc_train_loss += loss.item() * input.size(0)
909 acc_train_inner_loss += inner_loss.item() * input.size(0)
911 nb_train_samples += input.size(0)
912 nb_samples_seen += input.size(0)
914 total_loss = loss + (
915 args.rho_inner_loss * inner_loss if args.rho_inner_loss > 0 else 0.0
919 lr = get_lr(n_epoch, it)
920 for param_group in optimizer.param_groups:
921 param_group["lr"] = lr
923 # log_string(f"learning_rate {lr}")
925 optimizer.zero_grad()
926 total_loss.backward()
929 grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
931 loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
935 with torch.autograd.no_grad():
938 nb_test_samples, acc_test_loss = 0, 0.0
940 for input in task.batches(split="test"):
941 input = input.to(device)
943 output = model(mygpt.BracketedSequence(input)).x
944 loss = F.cross_entropy(output.transpose(1, 2), input)
945 acc_test_loss += loss.item() * input.size(0)
946 nb_test_samples += input.size(0)
949 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}"
952 task.produce_results(
955 result_dir=args.result_dir,
957 deterministic_synthesis=args.deterministic_synthesis,
960 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
961 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
964 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
967 time_current_result = datetime.datetime.now()
969 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
971 time_pred_result = time_current_result
974 "nb_epochs_finished": n_epoch + 1,
975 "model_state": model.state_dict(),
976 "rng_state": torch.get_rng_state(),
979 if torch.cuda.is_available():
980 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
982 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
983 torch.save(checkpoint, checkpoint_name)
984 log_string(f"saved checkpoint {checkpoint_name}")
986 ######################################################################