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 ######################################################################
164 args = parser.parse_args()
166 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
168 if args.result_dir is None:
169 args.result_dir = f"results_{args.task}"
171 ######################################################################
173 default_task_args = {
177 "nb_train_samples": 250000,
178 "nb_test_samples": 10000,
183 "nb_train_samples": 50000,
184 "nb_test_samples": 10000,
189 "nb_train_samples": 2500000,
190 "nb_test_samples": 10000,
195 "nb_train_samples": 250000,
196 "nb_test_samples": 10000,
201 "nb_train_samples": 1000000,
202 "nb_test_samples": 10000,
207 "nb_train_samples": 50000,
208 "nb_test_samples": 10000,
213 "nb_train_samples": 100000,
214 "nb_test_samples": 10000,
219 "nb_train_samples": 250000,
220 "nb_test_samples": 10000,
225 "nb_train_samples": 2500000,
226 "nb_test_samples": 10000,
231 "nb_train_samples": 250000,
232 "nb_test_samples": 10000,
237 "nb_train_samples": 100000,
238 "nb_test_samples": 1000,
243 "nb_train_samples": 50000,
244 "nb_test_samples": 10000,
249 "nb_train_samples": 60000,
250 "nb_test_samples": 10000,
254 if args.task in default_task_args:
255 for k, v in default_task_args[args.task].items():
256 if getattr(args, k) is None:
259 ######################################################################
261 default_model_args = {
292 if args.model in default_model_args:
293 for k, v in default_model_args[args.model].items():
294 if getattr(args, k) is None:
297 raise ValueError(f"Unknown model {args.model}")
299 ######################################################################
302 os.mkdir(args.result_dir)
303 except FileExistsError:
304 if not args.overwrite_results:
305 print(f"result directory {args.result_dir} already exists")
308 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
311 # torch.backends.cudnn.deterministic = True
312 # torch.backends.cudnn.benchmark = False
313 # torch.use_deterministic_algorithms(True)
314 torch.manual_seed(args.seed)
315 if torch.cuda.is_available():
316 torch.cuda.manual_seed_all(args.seed)
318 ######################################################################
322 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
324 if log_file is not None:
325 log_file.write(t + s + "\n")
333 log_string(f"args.{n} {getattr(args, n)}")
336 ######################################################################
339 def picoclvr_pruner_horizontal_green(p):
340 return not ("green" in p and ("left" in p or "right" in p))
343 picoclvr_pruner_train = (
344 picoclvr_pruner_horizontal_green
345 if args.picocvlr_prune_properties in {"train+eval"}
349 picoclvr_pruner_eval = (
350 (lambda p: not picoclvr_pruner_horizontal_green(p))
351 if args.picocvlr_prune_properties in {"train+eval", "eval"}
355 ######################################################################
357 if args.task == "byheart":
358 task = tasks.SandBox(
359 problem=problems.ProblemByHeart(),
360 nb_train_samples=args.nb_train_samples,
361 nb_test_samples=args.nb_test_samples,
362 batch_size=args.batch_size,
366 args.max_percents_of_test_in_train = -1
368 elif args.task == "learnop":
369 task = tasks.SandBox(
370 problem=problems.ProblemLearnOperator(),
371 nb_train_samples=args.nb_train_samples,
372 nb_test_samples=args.nb_test_samples,
373 batch_size=args.batch_size,
379 elif args.task == "guessop":
380 task = tasks.SandBox(
381 problem=problems.ProblemGuessOperator(),
382 nb_train_samples=args.nb_train_samples,
383 nb_test_samples=args.nb_test_samples,
384 batch_size=args.batch_size,
390 elif args.task == "twotargets":
391 task = tasks.SandBox(
392 problem=problems.ProblemTwoTargets(),
393 nb_train_samples=args.nb_train_samples,
394 nb_test_samples=args.nb_test_samples,
395 batch_size=args.batch_size,
400 elif args.task == "addition":
401 task = tasks.SandBox(
402 problem=problems.ProblemAddition(),
403 nb_train_samples=args.nb_train_samples,
404 nb_test_samples=args.nb_test_samples,
405 batch_size=args.batch_size,
410 elif args.task == "picoclvr":
411 task = tasks.PicoCLVR(
412 nb_train_samples=args.nb_train_samples,
413 nb_test_samples=args.nb_test_samples,
414 batch_size=args.batch_size,
415 height=args.picoclvr_height,
416 width=args.picoclvr_width,
417 nb_colors=args.picoclvr_nb_colors,
420 pruner_train=picoclvr_pruner_train,
421 pruner_eval=picoclvr_pruner_eval,
424 elif args.task == "mnist":
426 nb_train_samples=args.nb_train_samples,
427 nb_test_samples=args.nb_test_samples,
428 batch_size=args.batch_size,
432 elif args.task == "maze":
434 nb_train_samples=args.nb_train_samples,
435 nb_test_samples=args.nb_test_samples,
436 batch_size=args.batch_size,
437 height=args.maze_height,
438 width=args.maze_width,
439 nb_walls=args.maze_nb_walls,
443 elif args.task == "snake":
445 nb_train_samples=args.nb_train_samples,
446 nb_test_samples=args.nb_test_samples,
447 batch_size=args.batch_size,
448 height=args.snake_height,
449 width=args.snake_width,
450 nb_colors=args.snake_nb_colors,
451 length=args.snake_length,
452 prompt_length=args.snake_length // 2,
456 elif args.task == "stack":
458 nb_train_samples=args.nb_train_samples,
459 nb_test_samples=args.nb_test_samples,
460 batch_size=args.batch_size,
462 nb_steps=args.stack_nb_steps,
463 nb_stacks=args.stack_nb_stacks,
464 nb_digits=args.stack_nb_digits,
465 fraction_values_for_train=args.stack_fraction_values_for_train,
469 elif args.task == "expr":
471 nb_train_samples=args.nb_train_samples,
472 nb_test_samples=args.nb_test_samples,
473 nb_variables=args.expr_nb_variables,
474 sequence_length=args.expr_sequence_length,
475 operand_max=args.expr_operand_max,
476 result_max=args.expr_result_max,
477 batch_size=args.batch_size,
481 elif args.task == "rpl":
483 nb_train_samples=args.nb_train_samples,
484 nb_test_samples=args.nb_test_samples,
485 batch_size=args.batch_size,
486 nb_starting_values=args.rpl_nb_starting_values,
487 max_input=args.rpl_max_input,
488 prog_len=args.rpl_prog_len,
489 nb_runs=args.rpl_nb_runs,
490 no_prog=args.rpl_no_prog,
495 elif args.task == "grid":
497 nb_train_samples=args.nb_train_samples,
498 nb_test_samples=args.nb_test_samples,
499 batch_size=args.batch_size,
506 raise ValueError(f"Unknown task {args.task}")
508 ######################################################################
510 log_string(f"device {device}")
512 vocabulary_size = task.vocabulary_size()
514 log_string(f"vocabulary_size {vocabulary_size}")
516 ##############################
519 vocabulary_size=vocabulary_size,
520 dim_model=args.dim_model,
521 dim_keys=args.dim_keys,
522 dim_hidden=args.dim_hidden,
523 nb_heads=args.nb_heads,
524 nb_blocks=args.nb_blocks,
526 dropout=args.dropout,
531 nb_parameters = sum(p.numel() for p in model.parameters())
532 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
534 ######################################################################
536 nb_epochs_finished = 0
538 if args.no_checkpoint:
539 log_string(f"not trying to load checkpoint.")
543 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
544 checkpoint = torch.load(checkpoint_name)
545 nb_epochs_finished = checkpoint["nb_epochs_finished"]
546 model.load_state_dict(checkpoint["model_state"])
547 torch.set_rng_state(checkpoint["rng_state"])
548 if torch.cuda.is_available():
549 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
551 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
553 except FileNotFoundError:
554 log_string("starting from scratch.")
557 log_string("error when loading the checkpoint.")
560 ######################################################################
562 if args.task == "expr" and args.expr_input_file is not None:
563 task.produce_results(
564 n_epoch=nb_epochs_finished,
566 result_dir=args.result_dir,
568 deterministic_synthesis=args.deterministic_synthesis,
569 input_file=args.expr_input_file,
574 ######################################################################
576 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
578 # Compute the entropy of the training tokens
581 for input in task.batches(split="train"):
582 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
583 token_probas = token_count / token_count.sum()
584 entropy = -torch.xlogy(token_probas, token_probas).sum()
585 train_set_perplexity = math.exp(entropy)
587 ######################################################################
588 # A bit of paranoia never hurts
590 if args.max_percents_of_test_in_train >= 0:
592 def subsets_as_tuples(batches, cs):
594 for batch in batches:
596 s.add(tuple([v.item() for v in x]))
602 nb_test, nb_in_train = 0, 0
603 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
605 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
606 in_train.update(test_subset.intersection(train_subset))
607 nb_in_train += len(in_train)
608 nb_test += len(test_subset)
611 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
615 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
616 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
618 ##############################
620 if args.learning_rate_schedule == "cos":
621 learning_rate_schedule = {}
622 for n_epoch in range(args.nb_epochs):
623 u = n_epoch / args.nb_epochs * math.pi
624 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
629 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
633 learning_rate_schedule = {}
634 learning_rate = args.learning_rate
635 for n_epoch in range(args.nb_epochs):
637 learning_rate = u[n_epoch]
638 learning_rate_schedule[n_epoch] = learning_rate
640 log_string(f"learning_rate_schedule {learning_rate_schedule}")
642 ##############################
646 if nb_epochs_finished >= nb_epochs:
647 task.produce_results(
648 n_epoch=nb_epochs_finished,
650 result_dir=args.result_dir,
652 deterministic_synthesis=args.deterministic_synthesis,
655 for n_epoch in range(nb_epochs_finished, nb_epochs):
656 learning_rate = learning_rate_schedule[n_epoch]
658 log_string(f"learning_rate {learning_rate}")
660 if args.optim == "sgd":
661 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
662 elif args.optim == "adam":
663 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
664 elif args.optim == "adamw":
665 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
667 raise ValueError(f"Unknown optimizer {args.optim}.")
671 nb_train_samples, acc_train_loss = 0, 0.0
673 for input in task.batches(split="train"):
674 input = input.to(device)
675 output = model(mygpt.BracketedSequence(input)).x
676 loss = F.cross_entropy(output.transpose(1, 2), input)
677 acc_train_loss += loss.item() * input.size(0)
678 nb_train_samples += input.size(0)
679 nb_samples_seen += input.size(0)
681 optimizer.zero_grad()
685 with torch.autograd.no_grad():
688 nb_test_samples, acc_test_loss = 0, 0.0
690 for input in task.batches(split="test"):
691 input = input.to(device)
693 output = model(mygpt.BracketedSequence(input)).x
694 loss = F.cross_entropy(output.transpose(1, 2), input)
695 acc_test_loss += loss.item() * input.size(0)
696 nb_test_samples += input.size(0)
698 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
699 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
702 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
705 task.produce_results(
708 result_dir=args.result_dir,
710 deterministic_synthesis=args.deterministic_synthesis,
714 "nb_epochs_finished": n_epoch + 1,
715 "model_state": model.state_dict(),
716 "rng_state": torch.get_rng_state(),
719 if torch.cuda.is_available():
720 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
722 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
723 torch.save(checkpoint, checkpoint_name)
724 log_string(f"saved checkpoint {checkpoint_name}")
726 ######################################################################