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="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world",
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("--sandbox_level", type=int, default=0)
107 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
109 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
111 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
113 ##############################
116 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
118 parser.add_argument("--picoclvr_height", type=int, default=12)
120 parser.add_argument("--picoclvr_width", type=int, default=16)
122 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
124 ##############################
127 parser.add_argument("--maze_height", type=int, default=23)
129 parser.add_argument("--maze_width", type=int, default=39)
131 parser.add_argument("--maze_nb_walls", type=int, default=45)
133 ##############################
136 parser.add_argument("--snake_height", type=int, default=6)
138 parser.add_argument("--snake_width", type=int, default=8)
140 parser.add_argument("--snake_nb_colors", type=int, default=5)
142 parser.add_argument("--snake_length", type=int, default=200)
144 ##############################
147 parser.add_argument("--stack_nb_steps", type=int, default=100)
149 parser.add_argument("--stack_nb_stacks", type=int, default=3)
151 parser.add_argument("--stack_nb_digits", type=int, default=3)
153 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
155 ##############################
158 parser.add_argument("--expr_nb_variables", type=int, default=5)
160 parser.add_argument("--expr_sequence_length", type=int, default=40)
162 parser.add_argument("--expr_operand_max", type=int, default=9)
164 parser.add_argument("--expr_result_max", type=int, default=99)
166 parser.add_argument("--expr_input_file", type=str, default=None)
168 ##############################
171 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
173 ######################################################################
175 args = parser.parse_args()
177 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
179 if args.result_dir is None:
180 args.result_dir = f"results_{args.task}"
182 ######################################################################
184 default_task_args = {
188 "nb_train_samples": 100000,
189 "nb_test_samples": 10000,
194 "nb_train_samples": 250000,
195 "nb_test_samples": 10000,
200 "nb_train_samples": 250000,
201 "nb_test_samples": 10000,
206 "nb_train_samples": 250000,
207 "nb_test_samples": 10000,
212 "nb_train_samples": 250000,
213 "nb_test_samples": 10000,
218 "nb_train_samples": 100000,
219 "nb_test_samples": 1000,
224 "nb_train_samples": 1000000,
225 "nb_test_samples": 10000,
230 "nb_train_samples": 100000,
231 "nb_test_samples": 10000,
236 "nb_train_samples": 25000,
237 "nb_test_samples": 1000,
241 if args.task in default_task_args:
242 for k, v in default_task_args[args.task].items():
243 if getattr(args, k) is None:
246 ######################################################################
248 default_model_args = {
279 if args.model in default_model_args:
280 for k, v in default_model_args[args.model].items():
281 if getattr(args, k) is None:
284 raise ValueError(f"Unknown model {args.model}")
286 ######################################################################
289 os.mkdir(args.result_dir)
290 except FileExistsError:
291 if not args.overwrite_results:
292 print(f"result directory {args.result_dir} already exists")
295 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
298 # torch.backends.cudnn.deterministic = True
299 # torch.backends.cudnn.benchmark = False
300 # torch.use_deterministic_algorithms(True)
301 torch.manual_seed(args.seed)
302 if torch.cuda.is_available():
303 torch.cuda.manual_seed_all(args.seed)
305 ######################################################################
309 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
311 if log_file is not None:
312 log_file.write(t + s + "\n")
320 log_string(f"args.{n} {getattr(args, n)}")
323 ######################################################################
326 def picoclvr_pruner_horizontal_green(p):
327 return not ("green" in p and ("left" in p or "right" in p))
330 picoclvr_pruner_train = (
331 picoclvr_pruner_horizontal_green
332 if args.picocvlr_prune_properties in {"train+eval"}
336 picoclvr_pruner_eval = (
337 (lambda p: not picoclvr_pruner_horizontal_green(p))
338 if args.picocvlr_prune_properties in {"train+eval", "eval"}
342 ######################################################################
344 if args.task == "sandbox":
345 if args.sandbox_level == 0:
346 problem = problems.ProblemLevel0(
347 nb_sentences=args.sandbox_levels_nb_items,
348 len_prompt=args.sandbox_levels_len_source,
349 len_result=args.sandbox_levels_len_result,
351 elif args.sandbox_level == 1:
352 problem = problems.ProblemLevel1(
353 nb_operators=args.sandbox_levels_nb_items,
354 len_source=args.sandbox_levels_len_source,
355 len_result=args.sandbox_levels_len_result,
357 elif args.sandbox_level == 2:
358 problem = problems.ProblemLevel2(
359 len_source=args.sandbox_levels_len_source,
360 len_result=args.sandbox_levels_len_result,
363 raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
365 task = tasks.SandBox(
367 # problems.ProblemAddition(zero_padded=False, inverted_result=False),
368 problems.ProblemLenId(len_max=args.sandbox_levels_len_source),
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 == "picoclvr":
377 task = tasks.PicoCLVR(
378 nb_train_samples=args.nb_train_samples,
379 nb_test_samples=args.nb_test_samples,
380 batch_size=args.batch_size,
381 height=args.picoclvr_height,
382 width=args.picoclvr_width,
383 nb_colors=args.picoclvr_nb_colors,
386 pruner_train=picoclvr_pruner_train,
387 pruner_eval=picoclvr_pruner_eval,
390 elif args.task == "mnist":
392 nb_train_samples=args.nb_train_samples,
393 nb_test_samples=args.nb_test_samples,
394 batch_size=args.batch_size,
398 elif args.task == "maze":
400 nb_train_samples=args.nb_train_samples,
401 nb_test_samples=args.nb_test_samples,
402 batch_size=args.batch_size,
403 height=args.maze_height,
404 width=args.maze_width,
405 nb_walls=args.maze_nb_walls,
409 elif args.task == "snake":
411 nb_train_samples=args.nb_train_samples,
412 nb_test_samples=args.nb_test_samples,
413 batch_size=args.batch_size,
414 height=args.snake_height,
415 width=args.snake_width,
416 nb_colors=args.snake_nb_colors,
417 length=args.snake_length,
418 prompt_length=args.snake_length // 2,
422 elif args.task == "stack":
424 nb_train_samples=args.nb_train_samples,
425 nb_test_samples=args.nb_test_samples,
426 batch_size=args.batch_size,
428 nb_steps=args.stack_nb_steps,
429 nb_stacks=args.stack_nb_stacks,
430 nb_digits=args.stack_nb_digits,
431 fraction_values_for_train=args.stack_fraction_values_for_train,
435 elif args.task == "expr":
437 nb_train_samples=args.nb_train_samples,
438 nb_test_samples=args.nb_test_samples,
439 nb_variables=args.expr_nb_variables,
440 sequence_length=args.expr_sequence_length,
441 operand_max=args.expr_operand_max,
442 result_max=args.expr_result_max,
443 batch_size=args.batch_size,
447 elif args.task == "rpl":
449 nb_train_samples=args.nb_train_samples,
450 nb_test_samples=args.nb_test_samples,
451 batch_size=args.batch_size,
452 nb_starting_values=args.rpl_nb_starting_values,
453 max_input=args.rpl_max_input,
454 prog_len=args.rpl_prog_len,
455 nb_runs=args.rpl_nb_runs,
456 no_prog=args.rpl_no_prog,
461 elif args.task == "world":
463 nb_train_samples=args.nb_train_samples,
464 nb_test_samples=args.nb_test_samples,
465 batch_size=args.batch_size,
466 vqae_nb_epochs=args.world_vqae_nb_epochs,
472 raise ValueError(f"Unknown task {args.task}")
474 ######################################################################
476 log_string(f"device {device}")
478 vocabulary_size = task.vocabulary_size()
480 log_string(f"vocabulary_size {vocabulary_size}")
482 ##############################
485 vocabulary_size=vocabulary_size,
486 dim_model=args.dim_model,
487 dim_keys=args.dim_keys,
488 dim_hidden=args.dim_hidden,
489 nb_heads=args.nb_heads,
490 nb_blocks=args.nb_blocks,
492 dropout=args.dropout,
497 nb_parameters = sum(p.numel() for p in model.parameters())
498 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
500 ######################################################################
502 nb_epochs_finished = 0
504 if args.no_checkpoint:
505 log_string(f"not trying to load checkpoint.")
509 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
510 checkpoint = torch.load(checkpoint_name)
511 nb_epochs_finished = checkpoint["nb_epochs_finished"]
512 model.load_state_dict(checkpoint["model_state"])
513 torch.set_rng_state(checkpoint["rng_state"])
514 if torch.cuda.is_available():
515 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
517 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
519 except FileNotFoundError:
520 log_string("starting from scratch.")
523 log_string("error when loading the checkpoint.")
526 ######################################################################
528 if args.task == "expr" and args.expr_input_file is not None:
529 task.produce_results(
530 n_epoch=nb_epochs_finished,
532 result_dir=args.result_dir,
534 deterministic_synthesis=args.deterministic_synthesis,
535 input_file=args.expr_input_file,
540 ######################################################################
542 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
544 # Compute the entropy of the training tokens
547 for input in task.batches(split="train"):
548 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
549 token_probas = token_count / token_count.sum()
550 entropy = -torch.xlogy(token_probas, token_probas).sum()
551 train_set_perplexity = math.exp(entropy)
553 ######################################################################
554 # A bit of paranoia never hurts
557 def subsets_as_tuples(batches, cs):
559 for batch in batches:
561 s.add(tuple([v.item() for v in x]))
568 nb_test, nb_in_train = 0, 0
569 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
571 for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
572 in_train.update(test_subset.intersection(train_subset))
573 nb_in_train += len(in_train)
574 nb_test += len(test_subset)
577 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
581 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
582 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
584 ##############################
586 if args.learning_rate_schedule == "cos":
587 learning_rate_schedule = {}
588 for n_epoch in range(args.nb_epochs):
589 u = n_epoch / args.nb_epochs * math.pi
590 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
595 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
599 learning_rate_schedule = {}
600 learning_rate = args.learning_rate
601 for n_epoch in range(args.nb_epochs):
603 learning_rate = u[n_epoch]
604 learning_rate_schedule[n_epoch] = learning_rate
606 log_string(f"learning_rate_schedule {learning_rate_schedule}")
608 ##############################
612 if nb_epochs_finished >= nb_epochs:
613 task.produce_results(
614 n_epoch=nb_epochs_finished,
616 result_dir=args.result_dir,
618 deterministic_synthesis=args.deterministic_synthesis,
621 for n_epoch in range(nb_epochs_finished, nb_epochs):
622 learning_rate = learning_rate_schedule[n_epoch]
624 log_string(f"learning_rate {learning_rate}")
626 if args.optim == "sgd":
627 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
628 elif args.optim == "adam":
629 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
630 elif args.optim == "adamw":
631 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
633 raise ValueError(f"Unknown optimizer {args.optim}.")
637 nb_train_samples, acc_train_loss = 0, 0.0
639 for input in task.batches(split="train"):
640 input = input.to(device)
641 output = model(mygpt.BracketedSequence(input)).x
642 loss = F.cross_entropy(output.transpose(1, 2), input)
643 acc_train_loss += loss.item() * input.size(0)
644 nb_train_samples += input.size(0)
645 nb_samples_seen += input.size(0)
647 optimizer.zero_grad()
651 with torch.autograd.no_grad():
654 nb_test_samples, acc_test_loss = 0, 0.0
656 for input in task.batches(split="test"):
657 input = input.to(device)
659 output = model(mygpt.BracketedSequence(input)).x
660 loss = F.cross_entropy(output.transpose(1, 2), input)
661 acc_test_loss += loss.item() * input.size(0)
662 nb_test_samples += input.size(0)
664 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
665 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
668 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
671 task.produce_results(
674 result_dir=args.result_dir,
676 deterministic_synthesis=args.deterministic_synthesis,
680 "nb_epochs_finished": n_epoch + 1,
681 "model_state": model.state_dict(),
682 "rng_state": torch.get_rng_state(),
685 if torch.cuda.is_available():
686 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
688 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
689 torch.save(checkpoint, checkpoint_name)
690 log_string(f"saved checkpoint {checkpoint_name}")
692 ######################################################################