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 ######################################################################
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,
32 parser.add_argument("--task", type=str, default="world", help="world")
34 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
36 parser.add_argument("--result_dir", type=str, default=None)
38 parser.add_argument("--seed", type=int, default=0)
40 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
42 ########################################
44 parser.add_argument("--nb_epochs", type=int, default=10000)
46 parser.add_argument("--batch_size", type=int, default=None)
48 parser.add_argument("--physical_batch_size", type=int, default=None)
50 parser.add_argument("--nb_train_samples", type=int, default=None)
52 parser.add_argument("--nb_test_samples", type=int, default=None)
54 parser.add_argument("--learning_rate", type=float, default=1e-4)
56 ########################################
58 parser.add_argument("--model", type=str, default=None)
60 parser.add_argument("--dim_model", type=int, default=None)
62 parser.add_argument("--dim_keys", type=int, default=None)
64 parser.add_argument("--dim_hidden", type=int, default=None)
66 parser.add_argument("--nb_heads", type=int, default=None)
68 parser.add_argument("--nb_blocks", type=int, default=None)
70 parser.add_argument("--dropout", type=float, default=0.1)
72 ########################################
74 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
76 parser.add_argument("--nb_gpts", type=int, default=5)
78 parser.add_argument("--check", action="store_true", default=False)
80 ######################################################################
82 args = parser.parse_args()
84 if args.result_dir is None:
85 args.result_dir = f"results_{args.task}"
87 ######################################################################
93 "nb_train_samples": 250000,
94 "nb_test_samples": 10000,
98 if args.task in default_task_args:
99 for k, v in default_task_args[args.task].items():
100 if getattr(args, k) is None:
103 ######################################################################
105 default_model_args = {
143 if args.model in default_model_args:
144 for k, v in default_model_args[args.model].items():
145 if getattr(args, k) is None:
148 raise ValueError(f"Unknown model {args.model}")
150 ######################################################################
153 os.mkdir(args.result_dir)
154 except FileExistsError:
155 print(f"result directory {args.result_dir} already exists")
158 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
161 # torch.backends.cudnn.deterministic = True
162 # torch.backends.cudnn.benchmark = False
163 # torch.use_deterministic_algorithms(True)
164 torch.manual_seed(args.seed)
165 if torch.cuda.is_available():
166 torch.cuda.manual_seed_all(args.seed)
168 ######################################################################
172 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
174 if log_file is not None:
175 log_file.write(t + s + "\n")
182 log_string(f"argv {' '.join(sys.argv)}")
185 log_string(f"args.{n} {getattr(args, n)}")
188 ######################################################################
191 args.nb_train_samples = 500
192 args.nb_test_samples = 100
194 if args.physical_batch_size is None:
195 args.physical_batch_size = args.batch_size
197 assert args.batch_size % args.physical_batch_size == 0
199 assert args.nb_train_samples % args.batch_size == 0
200 assert args.nb_test_samples % args.batch_size == 0
202 if args.task == "file":
204 args.filetask_train_file is not None and args.filetask_test_file is not None
205 ), "You have to specify the task train and test files"
206 task = tasks.TaskFromFile(
207 args.filetask_train_file,
208 args.filetask_test_file,
209 nb_train_samples=args.nb_train_samples,
210 nb_test_samples=args.nb_test_samples,
211 batch_size=args.physical_batch_size,
215 args.max_percents_of_test_in_train = 0
217 elif args.task == "byheart":
218 task = tasks.SandBox(
219 problem=problems.ProblemByHeart(separation=args.byheart_separation),
220 nb_train_samples=args.nb_train_samples,
221 nb_test_samples=args.nb_test_samples,
222 batch_size=args.physical_batch_size,
226 args.max_percents_of_test_in_train = -1
228 elif args.task == "world":
230 nb_train_samples=args.nb_train_samples,
231 nb_test_samples=args.nb_test_samples,
232 batch_size=args.physical_batch_size,
233 result_dir=args.result_dir,
237 args.max_percents_of_test_in_train = -1
239 elif args.task == "learnop":
240 task = tasks.SandBox(
241 problem=problems.ProblemLearnOperator(),
242 nb_train_samples=args.nb_train_samples,
243 nb_test_samples=args.nb_test_samples,
244 batch_size=args.physical_batch_size,
250 elif args.task == "guessop":
251 task = tasks.SandBox(
252 problem=problems.ProblemGuessOperator(),
253 nb_train_samples=args.nb_train_samples,
254 nb_test_samples=args.nb_test_samples,
255 batch_size=args.physical_batch_size,
261 elif args.task == "twotargets":
262 task = tasks.SandBox(
263 problem=problems.ProblemTwoTargets(),
264 nb_train_samples=args.nb_train_samples,
265 nb_test_samples=args.nb_test_samples,
266 batch_size=args.physical_batch_size,
271 elif args.task == "memory":
272 task = tasks.SandBox(
273 problem=problems.ProblemMemory(),
274 nb_train_samples=args.nb_train_samples,
275 nb_test_samples=args.nb_test_samples,
276 batch_size=args.physical_batch_size,
281 elif args.task == "mixing":
282 task = tasks.SandBox(
283 problem=problems.ProblemMixing(
284 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
286 nb_train_samples=args.nb_train_samples,
287 nb_test_samples=args.nb_test_samples,
288 batch_size=args.physical_batch_size,
293 elif args.task == "addition":
294 task = tasks.SandBox(
295 problem=problems.ProblemAddition(),
296 nb_train_samples=args.nb_train_samples,
297 nb_test_samples=args.nb_test_samples,
298 batch_size=args.physical_batch_size,
303 elif args.task == "picoclvr":
304 task = tasks.PicoCLVR(
305 nb_train_samples=args.nb_train_samples,
306 nb_test_samples=args.nb_test_samples,
307 batch_size=args.physical_batch_size,
308 height=args.picoclvr_height,
309 width=args.picoclvr_width,
310 nb_colors=args.picoclvr_nb_colors,
313 pruner_train=picoclvr_pruner_train,
314 pruner_eval=picoclvr_pruner_eval,
317 elif args.task == "mnist":
319 nb_train_samples=args.nb_train_samples,
320 nb_test_samples=args.nb_test_samples,
321 batch_size=args.physical_batch_size,
325 elif args.task == "maze":
327 nb_train_samples=args.nb_train_samples,
328 nb_test_samples=args.nb_test_samples,
329 batch_size=args.physical_batch_size,
330 height=args.maze_height,
331 width=args.maze_width,
332 nb_walls=args.maze_nb_walls,
336 elif args.task == "snake":
338 nb_train_samples=args.nb_train_samples,
339 nb_test_samples=args.nb_test_samples,
340 batch_size=args.physical_batch_size,
341 height=args.snake_height,
342 width=args.snake_width,
343 nb_colors=args.snake_nb_colors,
344 length=args.snake_length,
345 prompt_length=args.snake_length // 2,
349 elif args.task == "stack":
351 nb_train_samples=args.nb_train_samples,
352 nb_test_samples=args.nb_test_samples,
353 batch_size=args.physical_batch_size,
355 nb_steps=args.stack_nb_steps,
356 nb_stacks=args.stack_nb_stacks,
357 nb_digits=args.stack_nb_digits,
358 fraction_values_for_train=args.stack_fraction_values_for_train,
362 elif args.task == "expr":
364 nb_train_samples=args.nb_train_samples,
365 nb_test_samples=args.nb_test_samples,
366 nb_variables=args.expr_nb_variables,
367 sequence_length=args.expr_sequence_length,
368 operand_max=args.expr_operand_max,
369 result_max=args.expr_result_max,
370 batch_size=args.physical_batch_size,
374 elif args.task == "rpl":
376 nb_train_samples=args.nb_train_samples,
377 nb_test_samples=args.nb_test_samples,
378 batch_size=args.physical_batch_size,
379 nb_starting_values=args.rpl_nb_starting_values,
380 max_input=args.rpl_max_input,
381 prog_len=args.rpl_prog_len,
382 nb_runs=args.rpl_nb_runs,
383 no_prog=args.rpl_no_prog,
388 elif args.task == "grid":
390 nb_train_samples=args.nb_train_samples,
391 nb_test_samples=args.nb_test_samples,
392 batch_size=args.physical_batch_size,
394 fraction_play=args.grid_fraction_play,
399 elif args.task == "qmlp":
401 nb_train_samples=args.nb_train_samples,
402 nb_test_samples=args.nb_test_samples,
403 batch_size=args.physical_batch_size,
404 result_dir=args.result_dir,
409 elif args.task == "greed":
411 nb_train_samples=args.nb_train_samples,
412 nb_test_samples=args.nb_test_samples,
413 batch_size=args.physical_batch_size,
414 height=args.greed_height,
415 width=args.greed_width,
417 nb_walls=args.greed_nb_walls,
418 nb_coins=args.greed_nb_coins,
424 raise ValueError(f"Unknown task {args.task}")
426 ######################################################################
428 log_string(f"device {device}")
430 vocabulary_size = task.vocabulary_size()
432 log_string(f"vocabulary_size {vocabulary_size}")
434 ######################################################################
436 # Compute the entropy of the training tokens
439 for input in task.batches(split="train", desc="train-entropy"):
440 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
441 token_probas = token_count / token_count.sum()
442 entropy = -torch.xlogy(token_probas, token_probas).sum()
443 train_set_perplexity = math.exp(entropy)
445 ######################################################################
446 # A bit of paranoia never hurts
448 if args.max_percents_of_test_in_train >= 0:
450 def subsets_as_tuples(batches, cs):
452 for batch in batches:
454 s.add(tuple([v.item() for v in x]))
460 nb_test, nb_in_train = 0, 0
461 for test_subset in subsets_as_tuples(
462 task.batches(split="test", desc="test-check"), 25000
465 for train_subset in subsets_as_tuples(
466 task.batches(split="train", desc="train-check"), 25000
468 in_train.update(test_subset.intersection(train_subset))
469 nb_in_train += len(in_train)
470 nb_test += len(test_subset)
473 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
477 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
478 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
480 ##############################
483 def one_epoch(model, task):
484 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
488 nb_train_samples, acc_train_loss = 0, 0.0
490 for input in task.batches(split="train"):
491 input = input.to(device)
493 if nb_train_samples % args.batch_size == 0:
494 optimizer.zero_grad()
496 output = model(mygpt.BracketedSequence(input)).x
497 loss = F.cross_entropy(output.transpose(1, 2), input)
498 acc_train_loss += loss.item() * input.size(0)
500 nb_train_samples += input.size(0)
504 if nb_train_samples % args.batch_size == 0:
507 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
509 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
512 ######################################################################
515 def run_tests(model, task, deterministic_synthesis):
516 with torch.autograd.no_grad():
519 nb_test_samples, acc_test_loss = 0, 0.0
520 nb_samples_accumulated = 0
522 for input in task.batches(split="test"):
523 input = input.to(device)
525 bs = model(mygpt.BracketedSequence(input))
528 loss = F.cross_entropy(output.transpose(1, 2), input)
530 acc_test_loss += loss.item() * input.size(0)
532 nb_test_samples += input.size(0)
534 main_test_accuracy = task.produce_results(
537 result_dir=args.result_dir,
539 deterministic_synthesis=deterministic_synthesis,
542 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
544 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
546 model.main_test_accuracy = main_test_accuracy
549 ######################################################################
561 while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
562 new_quizzes, nb_correct = task.create_new_quizzes(
564 result_dir=args.result_dir,
566 nb=4 * (nb_for_train + nb_for_test),
568 other_models=other_models,
573 to_keep = new_quizzes[nb_correct == len(other_models) - 1]
574 log_string(f"keep {to_keep.size(0)} quizzes")
577 new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
579 task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
580 task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
585 f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
590 ######################################################################
594 for k in range(args.nb_gpts):
596 vocabulary_size=vocabulary_size,
597 dim_model=args.dim_model,
598 dim_keys=args.dim_keys,
599 dim_hidden=args.dim_hidden,
600 nb_heads=args.nb_heads,
601 nb_blocks=args.nb_blocks,
603 dropout=args.dropout,
606 model.main_test_accuracy = 0.0
612 nb_parameters = sum(p.numel() for p in models[0].parameters())
613 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
615 ######################################################################
617 accuracy_to_make_quizzes = 0.975
618 nb_new_quizzes_for_train = 1000
619 nb_new_quizzes_for_test = 100
622 accuracy_to_make_quizzes = 0.0
623 nb_new_quizzes_for_train = 10
624 nb_new_quizzes_for_test = 10
626 for n_epoch in range(args.nb_epochs):
627 # select the model with lowest accuracy
628 models.sort(key=lambda model: model.main_test_accuracy)
632 f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
636 one_epoch(model, task)
639 f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
643 run_tests(model, task, deterministic_synthesis=False)
645 if model.main_test_accuracy >= accuracy_to_make_quizzes:
646 other_models = models.copy()
647 other_models.remove(model)
653 nb_for_train=nb_new_quizzes_for_train,
654 nb_for_test=nb_new_quizzes_for_test,
658 ######################################################################