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 ##############################
139 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
141 parser.add_argument("--picoclvr_height", type=int, default=12)
143 parser.add_argument("--picoclvr_width", type=int, default=16)
145 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
147 ##############################
150 parser.add_argument("--maze_height", type=int, default=13)
152 parser.add_argument("--maze_width", type=int, default=21)
154 parser.add_argument("--maze_nb_walls", type=int, default=15)
156 ##############################
159 parser.add_argument("--snake_height", type=int, default=9)
161 parser.add_argument("--snake_width", type=int, default=12)
163 parser.add_argument("--snake_nb_colors", type=int, default=5)
165 parser.add_argument("--snake_length", type=int, default=200)
167 ##############################
170 parser.add_argument("--stack_nb_steps", type=int, default=100)
172 parser.add_argument("--stack_nb_stacks", type=int, default=3)
174 parser.add_argument("--stack_nb_digits", type=int, default=3)
176 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
178 ##############################
181 parser.add_argument("--expr_nb_variables", type=int, default=5)
183 parser.add_argument("--expr_sequence_length", type=int, default=40)
185 parser.add_argument("--expr_operand_max", type=int, default=9)
187 parser.add_argument("--expr_result_max", type=int, default=99)
189 parser.add_argument("--expr_input_file", type=str, default=None)
191 ##############################
194 parser.add_argument("--memory_len_total", type=int, default=32)
196 ##############################
199 parser.add_argument("--mixing_hard", action="store_true", default=False)
201 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
203 ######################################################################
205 # args = parser.parse_args()
207 args, sup_args = parser.parse_known_args()
209 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
211 if args.result_dir is None:
212 args.result_dir = f"results_{args.task}_{args.model}"
214 ######################################################################
216 if not args.force_cpu and torch.cuda.is_available():
217 device = torch.device("cuda")
218 torch.backends.cuda.matmul.allow_tf32 = True
220 device = torch.device("cpu")
222 ######################################################################
224 default_task_args = {
228 "nb_train_samples": 250000,
229 "nb_test_samples": 10000,
234 "nb_train_samples": 50000,
235 "nb_test_samples": 10000,
240 "nb_train_samples": 2500000,
241 "nb_test_samples": 10000,
246 "nb_train_samples": 250000,
247 "nb_test_samples": 10000,
252 "nb_train_samples": 100000,
253 "nb_test_samples": 1000,
258 "nb_train_samples": 1000000,
259 "nb_test_samples": 10000,
264 "nb_train_samples": 50000,
265 "nb_test_samples": 10000,
270 "nb_train_samples": 100000,
271 "nb_test_samples": 10000,
276 "nb_train_samples": 250000,
277 "nb_test_samples": 10000,
282 "nb_train_samples": 2500000,
283 "nb_test_samples": 10000,
288 "nb_train_samples": 250000,
289 "nb_test_samples": 10000,
294 "nb_train_samples": 100000,
295 "nb_test_samples": 1000,
300 "nb_train_samples": 50000,
301 "nb_test_samples": 10000,
306 "nb_train_samples": 25000,
307 "nb_test_samples": 10000,
312 "nb_train_samples": 250000,
313 "nb_test_samples": 10000,
318 "nb_train_samples": 60000,
319 "nb_test_samples": 10000,
323 if args.task in default_task_args:
324 for k, v in default_task_args[args.task].items():
325 if getattr(args, k) is None:
328 ######################################################################
330 default_model_args = {
340 "attention": "caterpillar",
346 "caterpillar_height": 4,
358 "attention": "caterpillar",
364 "caterpillar_height": 4,
376 "attention": "caterpillar",
382 "caterpillar_height": 32,
394 "attention": "caterpillar",
411 "attention": "caterpillar",
421 if args.model in default_model_args:
422 for k, v in default_model_args[args.model].items():
423 if getattr(args, k) is None:
426 raise ValueError(f"Unknown model {args.model}")
428 ######################################################################
431 os.mkdir(args.result_dir)
432 except FileExistsError:
433 if not args.continue_training:
434 print(f"result directory {args.result_dir} already exists")
437 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
439 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
442 # torch.backends.cudnn.deterministic = True
443 # torch.backends.cudnn.benchmark = False
444 # torch.use_deterministic_algorithms(True)
445 torch.manual_seed(args.seed)
446 if torch.cuda.is_available():
447 torch.cuda.manual_seed_all(args.seed)
449 ######################################################################
453 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
455 if log_file is not None:
456 log_file.write(t + s + "\n")
463 with os.popen("sha256sum *.py") as f:
465 log_string(f"sha256sum {l.strip()}")
467 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
468 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
470 log_string(f"argv {' '.join(sys.argv)}")
473 log_string(f"args.{n} {getattr(args, n)}")
475 for k, v in sup_args.items():
476 log_string(f'sup_args["{k}"] "{v}"')
479 ######################################################################
482 def get_lr(n_epoch, it):
483 if args.legacy_lr_schedule:
484 # my crude scheduling to compare to previous baseline, added
487 if it < args.nb_warmup_iter:
488 return args.legacy_large_lr * it / args.nb_warmup_iter
489 elif n_epoch < args.legacy_nb_epoch_large_lr:
490 return args.legacy_large_lr
492 return args.legacy_small_lr
496 # 1) linear warmup for warmup_iter steps
497 if it < args.nb_warmup_iter:
498 return args.learning_rate * it / args.nb_warmup_iter
499 # 2) if it > nb_decay_iter, return min learning rate
500 if it > args.nb_decay_iter:
501 return args.min_learning_rate
502 # 3) in between, use cosine decay down to min learning rate
503 decay_ratio = (it - args.nb_warmup_iter) / (
504 args.nb_decay_iter - args.nb_warmup_iter
506 coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
507 return args.min_learning_rate + coeff * (
508 args.learning_rate - args.min_learning_rate
512 ######################################################################
515 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
518 def picoclvr_pruner_horizontal_green(p):
519 return not ("green" in p and ("left" in p or "right" in p))
522 picoclvr_pruner_train = (
523 picoclvr_pruner_horizontal_green
524 if args.picocvlr_prune_properties in {"train+eval"}
528 picoclvr_pruner_eval = (
529 (lambda p: not picoclvr_pruner_horizontal_green(p))
530 if args.picocvlr_prune_properties in {"train+eval", "eval"}
534 ######################################################################
538 if args.task == "byheart":
539 task = tasks.SandBox(
540 problem=problems.ProblemByHeart(),
541 nb_train_samples=args.nb_train_samples,
542 nb_test_samples=args.nb_test_samples,
543 batch_size=args.batch_size,
547 args.max_percents_of_test_in_train = -1
549 elif args.task == "learnop":
550 task = tasks.SandBox(
551 problem=problems.ProblemLearnOperator(),
552 nb_train_samples=args.nb_train_samples,
553 nb_test_samples=args.nb_test_samples,
554 batch_size=args.batch_size,
560 elif args.task == "guessop":
561 task = tasks.SandBox(
562 problem=problems.ProblemGuessOperator(),
563 nb_train_samples=args.nb_train_samples,
564 nb_test_samples=args.nb_test_samples,
565 batch_size=args.batch_size,
571 elif args.task == "twotargets":
572 task = tasks.SandBox(
573 problem=problems.ProblemTwoTargets(),
574 nb_train_samples=args.nb_train_samples,
575 nb_test_samples=args.nb_test_samples,
576 batch_size=args.batch_size,
581 elif args.task == "memory":
582 task = tasks.SandBox(
583 problem=problems.ProblemMemory(len_total=args.memory_len_total),
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 == "mixing":
592 task = tasks.SandBox(
593 problem=problems.ProblemMixing(
594 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
596 nb_train_samples=args.nb_train_samples,
597 nb_test_samples=args.nb_test_samples,
598 batch_size=args.batch_size,
603 elif args.task == "addition":
604 task = tasks.SandBox(
605 problem=problems.ProblemAddition(),
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 == "picoclvr":
614 task = tasks.PicoCLVR(
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.picoclvr_height,
619 width=args.picoclvr_width,
620 nb_colors=args.picoclvr_nb_colors,
623 pruner_train=picoclvr_pruner_train,
624 pruner_eval=picoclvr_pruner_eval,
627 elif args.task == "mnist":
629 nb_train_samples=args.nb_train_samples,
630 nb_test_samples=args.nb_test_samples,
631 batch_size=args.batch_size,
635 elif args.task == "maze":
637 nb_train_samples=args.nb_train_samples,
638 nb_test_samples=args.nb_test_samples,
639 batch_size=args.batch_size,
640 height=args.maze_height,
641 width=args.maze_width,
642 nb_walls=args.maze_nb_walls,
646 elif args.task == "snake":
648 nb_train_samples=args.nb_train_samples,
649 nb_test_samples=args.nb_test_samples,
650 batch_size=args.batch_size,
651 height=args.snake_height,
652 width=args.snake_width,
653 nb_colors=args.snake_nb_colors,
654 length=args.snake_length,
655 prompt_length=args.snake_length // 2,
659 elif args.task == "stack":
661 nb_train_samples=args.nb_train_samples,
662 nb_test_samples=args.nb_test_samples,
663 batch_size=args.batch_size,
665 nb_steps=args.stack_nb_steps,
666 nb_stacks=args.stack_nb_stacks,
667 nb_digits=args.stack_nb_digits,
668 fraction_values_for_train=args.stack_fraction_values_for_train,
672 elif args.task == "expr":
674 nb_train_samples=args.nb_train_samples,
675 nb_test_samples=args.nb_test_samples,
676 nb_variables=args.expr_nb_variables,
677 sequence_length=args.expr_sequence_length,
678 operand_max=args.expr_operand_max,
679 result_max=args.expr_result_max,
680 batch_size=args.batch_size,
684 elif args.task == "rpl":
686 nb_train_samples=args.nb_train_samples,
687 nb_test_samples=args.nb_test_samples,
688 batch_size=args.batch_size,
689 nb_starting_values=args.rpl_nb_starting_values,
690 max_input=args.rpl_max_input,
691 prog_len=args.rpl_prog_len,
692 nb_runs=args.rpl_nb_runs,
693 no_prog=args.rpl_no_prog,
698 elif args.task == "grid":
700 nb_train_samples=args.nb_train_samples,
701 nb_test_samples=args.nb_test_samples,
702 batch_size=args.batch_size,
708 elif args.task == "qmlp":
710 nb_train_samples=args.nb_train_samples,
711 nb_test_samples=args.nb_test_samples,
712 batch_size=args.batch_size,
713 result_dir=args.result_dir,
719 raise ValueError(f"Unknown task {args.task}")
721 ######################################################################
723 log_string(f"device {device}")
725 vocabulary_size = task.vocabulary_size()
727 log_string(f"vocabulary_size {vocabulary_size}")
729 ##############################
732 vocabulary_size=vocabulary_size,
733 dim_model=args.dim_model,
734 dim_keys=args.dim_keys,
735 dim_hidden=args.dim_hidden,
736 nb_heads=args.nb_heads,
737 nb_lines=args.nb_lines,
738 caterpillar_height=args.caterpillar_height,
739 nb_blocks=args.nb_blocks,
741 dropout=args.dropout,
742 attention_layer=args.attention,
749 nb_parameters = sum(p.numel() for p in model.parameters())
750 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
752 ######################################################################
754 nb_epochs_finished = 0
756 if args.no_checkpoint:
757 log_string(f"not trying to load checkpoint.")
761 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
762 checkpoint = torch.load(checkpoint_name)
763 nb_epochs_finished = checkpoint["nb_epochs_finished"]
764 model.load_state_dict(checkpoint["model_state"])
765 torch.set_rng_state(checkpoint["rng_state"])
766 if torch.cuda.is_available():
767 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
769 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
771 except FileNotFoundError:
772 log_string("starting from scratch.")
775 log_string("error when loading the checkpoint.")
778 ######################################################################
780 if args.task == "expr" and args.expr_input_file is not None:
781 task.produce_results(
782 n_epoch=nb_epochs_finished,
784 result_dir=args.result_dir,
786 deterministic_synthesis=args.deterministic_synthesis,
787 input_file=args.expr_input_file,
792 ######################################################################
794 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
796 # Compute the entropy of the training tokens
799 for input in task.batches(split="train"):
800 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
801 token_probas = token_count / token_count.sum()
802 entropy = -torch.xlogy(token_probas, token_probas).sum()
803 train_set_perplexity = math.exp(entropy)
805 ######################################################################
806 # A bit of paranoia never hurts
808 if args.max_percents_of_test_in_train >= 0:
810 def subsets_as_tuples(batches, cs):
812 for batch in batches:
814 s.add(tuple([v.item() for v in x]))
820 nb_test, nb_in_train = 0, 0
821 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
823 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
824 in_train.update(test_subset.intersection(train_subset))
825 nb_in_train += len(in_train)
826 nb_test += len(test_subset)
829 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
833 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
834 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
836 ##############################
838 for input in task.batches(split="train", desc="calibrate"):
839 input = input.to(device)
840 output = model(mygpt.BracketedSequence(input)).x
842 for n, m in model.named_modules():
845 if isinstance(x, mygpt.Calibrator):
846 print(f"####### ${n} | ${a} ########################")
847 mean, std = x.moments()
848 print("mean\n", mean, "\n")
849 print("std\n", std, "\n")
850 print(f"############################################\n\n")
854 ##############################
858 if nb_epochs_finished >= nb_epochs:
859 task.produce_results(
860 n_epoch=nb_epochs_finished,
862 result_dir=args.result_dir,
864 deterministic_synthesis=args.deterministic_synthesis,
867 time_pred_result = datetime.datetime.now()
873 for n_epoch in range(nb_epochs_finished, nb_epochs):
874 if args.optim == "sgd":
875 optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
876 elif args.optim == "adam":
877 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
878 elif args.optim == "adamw":
879 optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
881 raise ValueError(f"Unknown optimizer {args.optim}.")
885 nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
887 for input in task.batches(split="train"):
888 model.reset_inner_loss()
889 input = input.to(device)
891 output = model(mygpt.BracketedSequence(input)).x
892 loss = F.cross_entropy(output.transpose(1, 2), input)
893 inner_loss = model.get_inner_loss()
895 acc_train_loss += loss.item() * input.size(0)
896 acc_train_inner_loss += inner_loss.item() * input.size(0)
898 nb_train_samples += input.size(0)
899 nb_samples_seen += input.size(0)
901 total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
904 lr = get_lr(n_epoch, it)
905 for param_group in optimizer.param_groups:
906 param_group["lr"] = lr
908 # log_string(f"learning_rate {lr}")
910 optimizer.zero_grad()
911 total_loss.backward()
914 grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
916 loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
920 with torch.autograd.no_grad():
923 nb_test_samples, acc_test_loss = 0, 0.0
925 for input in task.batches(split="test"):
926 input = input.to(device)
928 output = model(mygpt.BracketedSequence(input)).x
929 loss = F.cross_entropy(output.transpose(1, 2), input)
930 acc_test_loss += loss.item() * input.size(0)
931 nb_test_samples += input.size(0)
934 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}"
937 task.produce_results(
940 result_dir=args.result_dir,
942 deterministic_synthesis=args.deterministic_synthesis,
945 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
946 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
949 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
952 time_current_result = datetime.datetime.now()
954 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
956 time_pred_result = time_current_result
959 "nb_epochs_finished": n_epoch + 1,
960 "model_state": model.state_dict(),
961 "rng_state": torch.get_rng_state(),
964 if torch.cuda.is_available():
965 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
967 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
968 torch.save(checkpoint, checkpoint_name)
969 log_string(f"saved checkpoint {checkpoint_name}")
971 ######################################################################