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",
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=None)
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="37M")
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=5)
94 parser.add_argument("--rpl_max_input", type=int, default=9)
96 parser.add_argument("--rpl_prog_len", type=int, default=10)
98 parser.add_argument("--rpl_nb_runs", type=int, default=8)
100 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
102 ##############################
105 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
107 parser.add_argument("--picoclvr_height", type=int, default=12)
109 parser.add_argument("--picoclvr_width", type=int, default=16)
111 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
113 ##############################
116 parser.add_argument("--maze_height", type=int, default=23)
118 parser.add_argument("--maze_width", type=int, default=39)
120 parser.add_argument("--maze_nb_walls", type=int, default=45)
122 ##############################
125 parser.add_argument("--snake_height", type=int, default=6)
127 parser.add_argument("--snake_width", type=int, default=8)
129 parser.add_argument("--snake_nb_colors", type=int, default=5)
131 parser.add_argument("--snake_length", type=int, default=200)
133 ##############################
136 parser.add_argument("--stack_nb_steps", type=int, default=100)
138 parser.add_argument("--stack_nb_stacks", type=int, default=3)
140 parser.add_argument("--stack_nb_digits", type=int, default=3)
142 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
144 ##############################
147 parser.add_argument("--expr_nb_variables", type=int, default=5)
149 parser.add_argument("--expr_sequence_length", type=int, default=40)
151 parser.add_argument("--expr_operand_max", type=int, default=9)
153 parser.add_argument("--expr_result_max", type=int, default=99)
155 parser.add_argument("--expr_input_file", type=str, default=None)
157 ##############################
160 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
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": 100000,
178 "nb_test_samples": 10000,
183 "nb_train_samples": 250000,
184 "nb_test_samples": 10000,
189 "nb_train_samples": 250000,
190 "nb_test_samples": 10000,
195 "nb_train_samples": 250000,
196 "nb_test_samples": 10000,
201 "nb_train_samples": 250000,
202 "nb_test_samples": 10000,
207 "nb_train_samples": 100000,
208 "nb_test_samples": 1000,
213 "nb_train_samples": 1000000,
214 "nb_test_samples": 10000,
219 "nb_train_samples": 100000,
220 "nb_test_samples": 10000,
225 "nb_train_samples": 25000,
226 "nb_test_samples": 1000,
230 if args.task in default_task_args:
231 for k, v in default_task_args[args.task].items():
232 if getattr(args, k) is None:
235 ######################################################################
237 default_model_args = {
268 if args.model in default_model_args:
269 for k, v in default_model_args[args.model].items():
270 if getattr(args, k) is None:
273 raise ValueError(f"Unknown model {args.model}")
275 ######################################################################
278 os.mkdir(args.result_dir)
279 except FileExistsError:
280 if not args.overwrite_results:
281 print(f"result directory {args.result_dir} already exists")
284 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
287 # torch.backends.cudnn.deterministic = True
288 # torch.backends.cudnn.benchmark = False
289 # torch.use_deterministic_algorithms(True)
290 torch.manual_seed(args.seed)
291 if torch.cuda.is_available():
292 torch.cuda.manual_seed_all(args.seed)
294 ######################################################################
298 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
300 if log_file is not None:
301 log_file.write(t + s + "\n")
309 log_string(f"args.{n} {getattr(args, n)}")
312 ######################################################################
315 def picoclvr_pruner_horizontal_green(p):
316 return not ("green" in p and ("left" in p or "right" in p))
319 picoclvr_pruner_train = (
320 picoclvr_pruner_horizontal_green
321 if args.picocvlr_prune_properties in {"train+eval"}
325 picoclvr_pruner_eval = (
326 (lambda p: not picoclvr_pruner_horizontal_green(p))
327 if args.picocvlr_prune_properties in {"train+eval", "eval"}
331 ######################################################################
333 if args.task == "byheart":
334 task = tasks.SandBox(
335 problem=problems.ProblemByHeart(),
336 nb_train_samples=args.nb_train_samples,
337 nb_test_samples=args.nb_test_samples,
338 batch_size=args.batch_size,
344 elif args.task == "learnop":
345 task = tasks.SandBox(
346 problem=problems.ProblemLearnOperator(),
347 nb_train_samples=args.nb_train_samples,
348 nb_test_samples=args.nb_test_samples,
349 batch_size=args.batch_size,
355 elif args.task == "guessop":
356 task = tasks.SandBox(
357 problem=problems.ProblemGuessOperator(),
358 nb_train_samples=args.nb_train_samples,
359 nb_test_samples=args.nb_test_samples,
360 batch_size=args.batch_size,
366 elif args.task == "twotargets":
367 task = tasks.SandBox(
368 problem=problems.ProblemTwoTargets(),
369 nb_train_samples=args.nb_train_samples,
370 nb_test_samples=args.nb_test_samples,
371 batch_size=args.batch_size,
376 elif args.task == "addition":
377 task = tasks.SandBox(
378 problem=problems.ProblemAddition(),
379 nb_train_samples=args.nb_train_samples,
380 nb_test_samples=args.nb_test_samples,
381 batch_size=args.batch_size,
386 elif args.task == "picoclvr":
387 task = tasks.PicoCLVR(
388 nb_train_samples=args.nb_train_samples,
389 nb_test_samples=args.nb_test_samples,
390 batch_size=args.batch_size,
391 height=args.picoclvr_height,
392 width=args.picoclvr_width,
393 nb_colors=args.picoclvr_nb_colors,
396 pruner_train=picoclvr_pruner_train,
397 pruner_eval=picoclvr_pruner_eval,
400 elif args.task == "mnist":
402 nb_train_samples=args.nb_train_samples,
403 nb_test_samples=args.nb_test_samples,
404 batch_size=args.batch_size,
408 elif args.task == "maze":
410 nb_train_samples=args.nb_train_samples,
411 nb_test_samples=args.nb_test_samples,
412 batch_size=args.batch_size,
413 height=args.maze_height,
414 width=args.maze_width,
415 nb_walls=args.maze_nb_walls,
419 elif args.task == "snake":
421 nb_train_samples=args.nb_train_samples,
422 nb_test_samples=args.nb_test_samples,
423 batch_size=args.batch_size,
424 height=args.snake_height,
425 width=args.snake_width,
426 nb_colors=args.snake_nb_colors,
427 length=args.snake_length,
428 prompt_length=args.snake_length // 2,
432 elif args.task == "stack":
434 nb_train_samples=args.nb_train_samples,
435 nb_test_samples=args.nb_test_samples,
436 batch_size=args.batch_size,
438 nb_steps=args.stack_nb_steps,
439 nb_stacks=args.stack_nb_stacks,
440 nb_digits=args.stack_nb_digits,
441 fraction_values_for_train=args.stack_fraction_values_for_train,
445 elif args.task == "expr":
447 nb_train_samples=args.nb_train_samples,
448 nb_test_samples=args.nb_test_samples,
449 nb_variables=args.expr_nb_variables,
450 sequence_length=args.expr_sequence_length,
451 operand_max=args.expr_operand_max,
452 result_max=args.expr_result_max,
453 batch_size=args.batch_size,
457 elif args.task == "rpl":
459 nb_train_samples=args.nb_train_samples,
460 nb_test_samples=args.nb_test_samples,
461 batch_size=args.batch_size,
462 nb_starting_values=args.rpl_nb_starting_values,
463 max_input=args.rpl_max_input,
464 prog_len=args.rpl_prog_len,
465 nb_runs=args.rpl_nb_runs,
466 no_prog=args.rpl_no_prog,
471 elif args.task == "world":
473 nb_train_samples=args.nb_train_samples,
474 nb_test_samples=args.nb_test_samples,
475 batch_size=args.batch_size,
476 vqae_nb_epochs=args.world_vqae_nb_epochs,
482 raise ValueError(f"Unknown task {args.task}")
484 ######################################################################
486 log_string(f"device {device}")
488 vocabulary_size = task.vocabulary_size()
490 log_string(f"vocabulary_size {vocabulary_size}")
492 ##############################
495 vocabulary_size=vocabulary_size,
496 dim_model=args.dim_model,
497 dim_keys=args.dim_keys,
498 dim_hidden=args.dim_hidden,
499 nb_heads=args.nb_heads,
500 nb_blocks=args.nb_blocks,
502 dropout=args.dropout,
507 nb_parameters = sum(p.numel() for p in model.parameters())
508 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
510 ######################################################################
512 nb_epochs_finished = 0
514 if args.no_checkpoint:
515 log_string(f"not trying to load checkpoint.")
519 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
520 checkpoint = torch.load(checkpoint_name)
521 nb_epochs_finished = checkpoint["nb_epochs_finished"]
522 model.load_state_dict(checkpoint["model_state"])
523 torch.set_rng_state(checkpoint["rng_state"])
524 if torch.cuda.is_available():
525 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
527 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
529 except FileNotFoundError:
530 log_string("starting from scratch.")
533 log_string("error when loading the checkpoint.")
536 ######################################################################
538 if args.task == "expr" and args.expr_input_file is not None:
539 task.produce_results(
540 n_epoch=nb_epochs_finished,
542 result_dir=args.result_dir,
544 deterministic_synthesis=args.deterministic_synthesis,
545 input_file=args.expr_input_file,
550 ######################################################################
552 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
554 # Compute the entropy of the training tokens
557 for input in task.batches(split="train"):
558 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
559 token_probas = token_count / token_count.sum()
560 entropy = -torch.xlogy(token_probas, token_probas).sum()
561 train_set_perplexity = math.exp(entropy)
563 ######################################################################
564 # A bit of paranoia never hurts
567 def subsets_as_tuples(batches, cs):
569 for batch in batches:
571 s.add(tuple([v.item() for v in x]))
578 nb_test, nb_in_train = 0, 0
579 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
581 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
582 in_train.update(test_subset.intersection(train_subset))
583 nb_in_train += len(in_train)
584 nb_test += len(test_subset)
587 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
591 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
592 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
594 ##############################
596 if args.learning_rate_schedule == "cos":
597 learning_rate_schedule = {}
598 for n_epoch in range(args.nb_epochs):
599 u = n_epoch / args.nb_epochs * math.pi
600 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
605 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
609 learning_rate_schedule = {}
610 learning_rate = args.learning_rate
611 for n_epoch in range(args.nb_epochs):
613 learning_rate = u[n_epoch]
614 learning_rate_schedule[n_epoch] = learning_rate
616 log_string(f"learning_rate_schedule {learning_rate_schedule}")
618 ##############################
622 if nb_epochs_finished >= nb_epochs:
623 task.produce_results(
624 n_epoch=nb_epochs_finished,
626 result_dir=args.result_dir,
628 deterministic_synthesis=args.deterministic_synthesis,
631 for n_epoch in range(nb_epochs_finished, nb_epochs):
632 learning_rate = learning_rate_schedule[n_epoch]
634 log_string(f"learning_rate {learning_rate}")
636 if args.optim == "sgd":
637 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
638 elif args.optim == "adam":
639 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
640 elif args.optim == "adamw":
641 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
643 raise ValueError(f"Unknown optimizer {args.optim}.")
647 nb_train_samples, acc_train_loss = 0, 0.0
649 for input in task.batches(split="train"):
650 input = input.to(device)
651 output = model(mygpt.BracketedSequence(input)).x
652 loss = F.cross_entropy(output.transpose(1, 2), input)
653 acc_train_loss += loss.item() * input.size(0)
654 nb_train_samples += input.size(0)
655 nb_samples_seen += input.size(0)
657 optimizer.zero_grad()
661 with torch.autograd.no_grad():
664 nb_test_samples, acc_test_loss = 0, 0.0
666 for input in task.batches(split="test"):
667 input = input.to(device)
669 output = model(mygpt.BracketedSequence(input)).x
670 loss = F.cross_entropy(output.transpose(1, 2), input)
671 acc_test_loss += loss.item() * input.size(0)
672 nb_test_samples += input.size(0)
674 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
675 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
678 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
681 task.produce_results(
684 result_dir=args.result_dir,
686 deterministic_synthesis=args.deterministic_synthesis,
690 "nb_epochs_finished": n_epoch + 1,
691 "model_state": model.state_dict(),
692 "rng_state": torch.get_rng_state(),
695 if torch.cuda.is_available():
696 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
698 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
699 torch.save(checkpoint, checkpoint_name)
700 log_string(f"saved checkpoint {checkpoint_name}")
702 ######################################################################