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
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, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid",
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("--rpl_nb_starting_values", type=int, default=3)
94 parser.add_argument("--rpl_max_input", type=int, default=9)
96 parser.add_argument("--rpl_prog_len", type=int, default=8)
98 parser.add_argument("--rpl_nb_runs", type=int, default=5)
100 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
102 ##############################
105 parser.add_argument("--grid_size", type=int, default=6)
107 ##############################
110 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
112 parser.add_argument("--picoclvr_height", type=int, default=12)
114 parser.add_argument("--picoclvr_width", type=int, default=16)
116 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
118 ##############################
121 parser.add_argument("--maze_height", type=int, default=13)
123 parser.add_argument("--maze_width", type=int, default=21)
125 parser.add_argument("--maze_nb_walls", type=int, default=15)
127 ##############################
130 parser.add_argument("--snake_height", type=int, default=9)
132 parser.add_argument("--snake_width", type=int, default=12)
134 parser.add_argument("--snake_nb_colors", type=int, default=5)
136 parser.add_argument("--snake_length", type=int, default=200)
138 ##############################
141 parser.add_argument("--stack_nb_steps", type=int, default=100)
143 parser.add_argument("--stack_nb_stacks", type=int, default=3)
145 parser.add_argument("--stack_nb_digits", type=int, default=3)
147 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
149 ##############################
152 parser.add_argument("--expr_nb_variables", type=int, default=5)
154 parser.add_argument("--expr_sequence_length", type=int, default=40)
156 parser.add_argument("--expr_operand_max", type=int, default=9)
158 parser.add_argument("--expr_result_max", type=int, default=99)
160 parser.add_argument("--expr_input_file", type=str, default=None)
162 ##############################
165 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
167 ######################################################################
169 args = parser.parse_args()
171 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
173 if args.result_dir is None:
174 args.result_dir = f"results_{args.task}"
176 ######################################################################
178 default_task_args = {
182 "nb_train_samples": 250000,
183 "nb_test_samples": 10000,
188 "nb_train_samples": 50000,
189 "nb_test_samples": 10000,
194 "nb_train_samples": 2500000,
195 "nb_test_samples": 10000,
200 "nb_train_samples": 250000,
201 "nb_test_samples": 10000,
206 "nb_train_samples": 1000000,
207 "nb_test_samples": 10000,
212 "nb_train_samples": 50000,
213 "nb_test_samples": 10000,
218 "nb_train_samples": 100000,
219 "nb_test_samples": 10000,
224 "nb_train_samples": 250000,
225 "nb_test_samples": 10000,
230 "nb_train_samples": 2500000,
231 "nb_test_samples": 10000,
236 "nb_train_samples": 250000,
237 "nb_test_samples": 10000,
242 "nb_train_samples": 100000,
243 "nb_test_samples": 1000,
248 "nb_train_samples": 50000,
249 "nb_test_samples": 10000,
255 "nb_train_samples": 60000,
256 "nb_test_samples": 10000,
261 "nb_train_samples": 25000,
262 "nb_test_samples": 1000,
266 if args.task in default_task_args:
267 for k, v in default_task_args[args.task].items():
268 if getattr(args, k) is None:
271 ######################################################################
273 default_model_args = {
304 if args.model in default_model_args:
305 for k, v in default_model_args[args.model].items():
306 if getattr(args, k) is None:
309 raise ValueError(f"Unknown model {args.model}")
311 ######################################################################
314 os.mkdir(args.result_dir)
315 except FileExistsError:
316 if not args.overwrite_results:
317 print(f"result directory {args.result_dir} already exists")
320 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
323 # torch.backends.cudnn.deterministic = True
324 # torch.backends.cudnn.benchmark = False
325 # torch.use_deterministic_algorithms(True)
326 torch.manual_seed(args.seed)
327 if torch.cuda.is_available():
328 torch.cuda.manual_seed_all(args.seed)
330 ######################################################################
334 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
336 if log_file is not None:
337 log_file.write(t + s + "\n")
345 log_string(f"args.{n} {getattr(args, n)}")
348 ######################################################################
351 def picoclvr_pruner_horizontal_green(p):
352 return not ("green" in p and ("left" in p or "right" in p))
355 picoclvr_pruner_train = (
356 picoclvr_pruner_horizontal_green
357 if args.picocvlr_prune_properties in {"train+eval"}
361 picoclvr_pruner_eval = (
362 (lambda p: not picoclvr_pruner_horizontal_green(p))
363 if args.picocvlr_prune_properties in {"train+eval", "eval"}
367 ######################################################################
369 if args.task == "byheart":
370 task = tasks.SandBox(
371 problem=problems.ProblemByHeart(),
372 nb_train_samples=args.nb_train_samples,
373 nb_test_samples=args.nb_test_samples,
374 batch_size=args.batch_size,
378 args.max_percents_of_test_in_train = -1
380 elif args.task == "learnop":
381 task = tasks.SandBox(
382 problem=problems.ProblemLearnOperator(),
383 nb_train_samples=args.nb_train_samples,
384 nb_test_samples=args.nb_test_samples,
385 batch_size=args.batch_size,
391 elif args.task == "guessop":
392 task = tasks.SandBox(
393 problem=problems.ProblemGuessOperator(),
394 nb_train_samples=args.nb_train_samples,
395 nb_test_samples=args.nb_test_samples,
396 batch_size=args.batch_size,
402 elif args.task == "twotargets":
403 task = tasks.SandBox(
404 problem=problems.ProblemTwoTargets(),
405 nb_train_samples=args.nb_train_samples,
406 nb_test_samples=args.nb_test_samples,
407 batch_size=args.batch_size,
412 elif args.task == "addition":
413 task = tasks.SandBox(
414 problem=problems.ProblemAddition(),
415 nb_train_samples=args.nb_train_samples,
416 nb_test_samples=args.nb_test_samples,
417 batch_size=args.batch_size,
422 elif args.task == "picoclvr":
423 task = tasks.PicoCLVR(
424 nb_train_samples=args.nb_train_samples,
425 nb_test_samples=args.nb_test_samples,
426 batch_size=args.batch_size,
427 height=args.picoclvr_height,
428 width=args.picoclvr_width,
429 nb_colors=args.picoclvr_nb_colors,
432 pruner_train=picoclvr_pruner_train,
433 pruner_eval=picoclvr_pruner_eval,
436 elif args.task == "mnist":
438 nb_train_samples=args.nb_train_samples,
439 nb_test_samples=args.nb_test_samples,
440 batch_size=args.batch_size,
444 elif args.task == "maze":
446 nb_train_samples=args.nb_train_samples,
447 nb_test_samples=args.nb_test_samples,
448 batch_size=args.batch_size,
449 height=args.maze_height,
450 width=args.maze_width,
451 nb_walls=args.maze_nb_walls,
455 elif args.task == "snake":
457 nb_train_samples=args.nb_train_samples,
458 nb_test_samples=args.nb_test_samples,
459 batch_size=args.batch_size,
460 height=args.snake_height,
461 width=args.snake_width,
462 nb_colors=args.snake_nb_colors,
463 length=args.snake_length,
464 prompt_length=args.snake_length // 2,
468 elif args.task == "stack":
470 nb_train_samples=args.nb_train_samples,
471 nb_test_samples=args.nb_test_samples,
472 batch_size=args.batch_size,
474 nb_steps=args.stack_nb_steps,
475 nb_stacks=args.stack_nb_stacks,
476 nb_digits=args.stack_nb_digits,
477 fraction_values_for_train=args.stack_fraction_values_for_train,
481 elif args.task == "expr":
483 nb_train_samples=args.nb_train_samples,
484 nb_test_samples=args.nb_test_samples,
485 nb_variables=args.expr_nb_variables,
486 sequence_length=args.expr_sequence_length,
487 operand_max=args.expr_operand_max,
488 result_max=args.expr_result_max,
489 batch_size=args.batch_size,
493 elif args.task == "rpl":
495 nb_train_samples=args.nb_train_samples,
496 nb_test_samples=args.nb_test_samples,
497 batch_size=args.batch_size,
498 nb_starting_values=args.rpl_nb_starting_values,
499 max_input=args.rpl_max_input,
500 prog_len=args.rpl_prog_len,
501 nb_runs=args.rpl_nb_runs,
502 no_prog=args.rpl_no_prog,
507 elif args.task == "grid":
509 nb_train_samples=args.nb_train_samples,
510 nb_test_samples=args.nb_test_samples,
511 batch_size=args.batch_size,
517 elif args.task == "world":
519 nb_train_samples=args.nb_train_samples,
520 nb_test_samples=args.nb_test_samples,
521 batch_size=args.batch_size,
522 vqae_nb_epochs=args.world_vqae_nb_epochs,
528 raise ValueError(f"Unknown task {args.task}")
530 ######################################################################
532 log_string(f"device {device}")
534 vocabulary_size = task.vocabulary_size()
536 log_string(f"vocabulary_size {vocabulary_size}")
538 ##############################
541 vocabulary_size=vocabulary_size,
542 dim_model=args.dim_model,
543 dim_keys=args.dim_keys,
544 dim_hidden=args.dim_hidden,
545 nb_heads=args.nb_heads,
546 nb_blocks=args.nb_blocks,
548 dropout=args.dropout,
553 nb_parameters = sum(p.numel() for p in model.parameters())
554 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
556 ######################################################################
558 nb_epochs_finished = 0
560 if args.no_checkpoint:
561 log_string(f"not trying to load checkpoint.")
565 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
566 checkpoint = torch.load(checkpoint_name)
567 nb_epochs_finished = checkpoint["nb_epochs_finished"]
568 model.load_state_dict(checkpoint["model_state"])
569 torch.set_rng_state(checkpoint["rng_state"])
570 if torch.cuda.is_available():
571 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
573 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
575 except FileNotFoundError:
576 log_string("starting from scratch.")
579 log_string("error when loading the checkpoint.")
582 ######################################################################
584 if args.task == "expr" and args.expr_input_file is not None:
585 task.produce_results(
586 n_epoch=nb_epochs_finished,
588 result_dir=args.result_dir,
590 deterministic_synthesis=args.deterministic_synthesis,
591 input_file=args.expr_input_file,
596 ######################################################################
598 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
600 # Compute the entropy of the training tokens
603 for input in task.batches(split="train"):
604 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
605 token_probas = token_count / token_count.sum()
606 entropy = -torch.xlogy(token_probas, token_probas).sum()
607 train_set_perplexity = math.exp(entropy)
609 ######################################################################
610 # A bit of paranoia never hurts
612 if args.max_percents_of_test_in_train >= 0:
614 def subsets_as_tuples(batches, cs):
616 for batch in batches:
618 s.add(tuple([v.item() for v in x]))
624 nb_test, nb_in_train = 0, 0
625 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
627 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
628 in_train.update(test_subset.intersection(train_subset))
629 nb_in_train += len(in_train)
630 nb_test += len(test_subset)
633 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
637 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
638 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
640 ##############################
642 if args.learning_rate_schedule == "cos":
643 learning_rate_schedule = {}
644 for n_epoch in range(args.nb_epochs):
645 u = n_epoch / args.nb_epochs * math.pi
646 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
651 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
655 learning_rate_schedule = {}
656 learning_rate = args.learning_rate
657 for n_epoch in range(args.nb_epochs):
659 learning_rate = u[n_epoch]
660 learning_rate_schedule[n_epoch] = learning_rate
662 log_string(f"learning_rate_schedule {learning_rate_schedule}")
664 ##############################
668 if nb_epochs_finished >= nb_epochs:
669 task.produce_results(
670 n_epoch=nb_epochs_finished,
672 result_dir=args.result_dir,
674 deterministic_synthesis=args.deterministic_synthesis,
677 for n_epoch in range(nb_epochs_finished, nb_epochs):
678 learning_rate = learning_rate_schedule[n_epoch]
680 log_string(f"learning_rate {learning_rate}")
682 if args.optim == "sgd":
683 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
684 elif args.optim == "adam":
685 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
686 elif args.optim == "adamw":
687 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
689 raise ValueError(f"Unknown optimizer {args.optim}.")
693 nb_train_samples, acc_train_loss = 0, 0.0
695 for input in task.batches(split="train"):
696 input = input.to(device)
697 output = model(mygpt.BracketedSequence(input)).x
698 loss = F.cross_entropy(output.transpose(1, 2), input)
699 acc_train_loss += loss.item() * input.size(0)
700 nb_train_samples += input.size(0)
701 nb_samples_seen += input.size(0)
703 optimizer.zero_grad()
707 with torch.autograd.no_grad():
710 nb_test_samples, acc_test_loss = 0, 0.0
712 for input in task.batches(split="test"):
713 input = input.to(device)
715 output = model(mygpt.BracketedSequence(input)).x
716 loss = F.cross_entropy(output.transpose(1, 2), input)
717 acc_test_loss += loss.item() * input.size(0)
718 nb_test_samples += input.size(0)
720 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
721 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
724 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
727 task.produce_results(
730 result_dir=args.result_dir,
732 deterministic_synthesis=args.deterministic_synthesis,
736 "nb_epochs_finished": n_epoch + 1,
737 "model_state": model.state_dict(),
738 "rng_state": torch.get_rng_state(),
741 if torch.cuda.is_available():
742 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
744 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
745 torch.save(checkpoint, checkpoint_name)
746 log_string(f"saved checkpoint {checkpoint_name}")
748 ######################################################################