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 ######################################################################
78 args = parser.parse_args()
80 if args.result_dir is None:
81 args.result_dir = f"results_{args.task}"
83 ######################################################################
89 "nb_train_samples": 250000,
90 "nb_test_samples": 10000,
94 if args.task in default_task_args:
95 for k, v in default_task_args[args.task].items():
96 if getattr(args, k) is None:
99 ######################################################################
101 default_model_args = {
139 if args.model in default_model_args:
140 for k, v in default_model_args[args.model].items():
141 if getattr(args, k) is None:
144 raise ValueError(f"Unknown model {args.model}")
146 ######################################################################
149 os.mkdir(args.result_dir)
150 except FileExistsError:
151 print(f"result directory {args.result_dir} already exists")
154 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
157 # torch.backends.cudnn.deterministic = True
158 # torch.backends.cudnn.benchmark = False
159 # torch.use_deterministic_algorithms(True)
160 torch.manual_seed(args.seed)
161 if torch.cuda.is_available():
162 torch.cuda.manual_seed_all(args.seed)
164 ######################################################################
168 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
170 if log_file is not None:
171 log_file.write(t + s + "\n")
178 log_string(f"argv {' '.join(sys.argv)}")
181 log_string(f"args.{n} {getattr(args, n)}")
184 ######################################################################
187 if args.physical_batch_size is None:
188 args.physical_batch_size = args.batch_size
190 assert args.batch_size % args.physical_batch_size == 0
192 assert args.nb_train_samples % args.batch_size == 0
193 assert args.nb_test_samples % args.batch_size == 0
195 if args.task == "file":
197 args.filetask_train_file is not None and args.filetask_test_file is not None
198 ), "You have to specify the task train and test files"
199 task = tasks.TaskFromFile(
200 args.filetask_train_file,
201 args.filetask_test_file,
202 nb_train_samples=args.nb_train_samples,
203 nb_test_samples=args.nb_test_samples,
204 batch_size=args.physical_batch_size,
208 args.max_percents_of_test_in_train = 0
210 elif args.task == "byheart":
211 task = tasks.SandBox(
212 problem=problems.ProblemByHeart(separation=args.byheart_separation),
213 nb_train_samples=args.nb_train_samples,
214 nb_test_samples=args.nb_test_samples,
215 batch_size=args.physical_batch_size,
219 args.max_percents_of_test_in_train = -1
221 elif args.task == "world":
223 nb_train_samples=args.nb_train_samples,
224 nb_test_samples=args.nb_test_samples,
225 batch_size=args.physical_batch_size,
226 result_dir=args.result_dir,
230 args.max_percents_of_test_in_train = -1
232 elif args.task == "learnop":
233 task = tasks.SandBox(
234 problem=problems.ProblemLearnOperator(),
235 nb_train_samples=args.nb_train_samples,
236 nb_test_samples=args.nb_test_samples,
237 batch_size=args.physical_batch_size,
243 elif args.task == "guessop":
244 task = tasks.SandBox(
245 problem=problems.ProblemGuessOperator(),
246 nb_train_samples=args.nb_train_samples,
247 nb_test_samples=args.nb_test_samples,
248 batch_size=args.physical_batch_size,
254 elif args.task == "twotargets":
255 task = tasks.SandBox(
256 problem=problems.ProblemTwoTargets(),
257 nb_train_samples=args.nb_train_samples,
258 nb_test_samples=args.nb_test_samples,
259 batch_size=args.physical_batch_size,
264 elif args.task == "memory":
265 task = tasks.SandBox(
266 problem=problems.ProblemMemory(),
267 nb_train_samples=args.nb_train_samples,
268 nb_test_samples=args.nb_test_samples,
269 batch_size=args.physical_batch_size,
274 elif args.task == "mixing":
275 task = tasks.SandBox(
276 problem=problems.ProblemMixing(
277 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
279 nb_train_samples=args.nb_train_samples,
280 nb_test_samples=args.nb_test_samples,
281 batch_size=args.physical_batch_size,
286 elif args.task == "addition":
287 task = tasks.SandBox(
288 problem=problems.ProblemAddition(),
289 nb_train_samples=args.nb_train_samples,
290 nb_test_samples=args.nb_test_samples,
291 batch_size=args.physical_batch_size,
296 elif args.task == "picoclvr":
297 task = tasks.PicoCLVR(
298 nb_train_samples=args.nb_train_samples,
299 nb_test_samples=args.nb_test_samples,
300 batch_size=args.physical_batch_size,
301 height=args.picoclvr_height,
302 width=args.picoclvr_width,
303 nb_colors=args.picoclvr_nb_colors,
306 pruner_train=picoclvr_pruner_train,
307 pruner_eval=picoclvr_pruner_eval,
310 elif args.task == "mnist":
312 nb_train_samples=args.nb_train_samples,
313 nb_test_samples=args.nb_test_samples,
314 batch_size=args.physical_batch_size,
318 elif args.task == "maze":
320 nb_train_samples=args.nb_train_samples,
321 nb_test_samples=args.nb_test_samples,
322 batch_size=args.physical_batch_size,
323 height=args.maze_height,
324 width=args.maze_width,
325 nb_walls=args.maze_nb_walls,
329 elif args.task == "snake":
331 nb_train_samples=args.nb_train_samples,
332 nb_test_samples=args.nb_test_samples,
333 batch_size=args.physical_batch_size,
334 height=args.snake_height,
335 width=args.snake_width,
336 nb_colors=args.snake_nb_colors,
337 length=args.snake_length,
338 prompt_length=args.snake_length // 2,
342 elif args.task == "stack":
344 nb_train_samples=args.nb_train_samples,
345 nb_test_samples=args.nb_test_samples,
346 batch_size=args.physical_batch_size,
348 nb_steps=args.stack_nb_steps,
349 nb_stacks=args.stack_nb_stacks,
350 nb_digits=args.stack_nb_digits,
351 fraction_values_for_train=args.stack_fraction_values_for_train,
355 elif args.task == "expr":
357 nb_train_samples=args.nb_train_samples,
358 nb_test_samples=args.nb_test_samples,
359 nb_variables=args.expr_nb_variables,
360 sequence_length=args.expr_sequence_length,
361 operand_max=args.expr_operand_max,
362 result_max=args.expr_result_max,
363 batch_size=args.physical_batch_size,
367 elif args.task == "rpl":
369 nb_train_samples=args.nb_train_samples,
370 nb_test_samples=args.nb_test_samples,
371 batch_size=args.physical_batch_size,
372 nb_starting_values=args.rpl_nb_starting_values,
373 max_input=args.rpl_max_input,
374 prog_len=args.rpl_prog_len,
375 nb_runs=args.rpl_nb_runs,
376 no_prog=args.rpl_no_prog,
381 elif args.task == "grid":
383 nb_train_samples=args.nb_train_samples,
384 nb_test_samples=args.nb_test_samples,
385 batch_size=args.physical_batch_size,
387 fraction_play=args.grid_fraction_play,
392 elif args.task == "qmlp":
394 nb_train_samples=args.nb_train_samples,
395 nb_test_samples=args.nb_test_samples,
396 batch_size=args.physical_batch_size,
397 result_dir=args.result_dir,
402 elif args.task == "greed":
404 nb_train_samples=args.nb_train_samples,
405 nb_test_samples=args.nb_test_samples,
406 batch_size=args.physical_batch_size,
407 height=args.greed_height,
408 width=args.greed_width,
410 nb_walls=args.greed_nb_walls,
411 nb_coins=args.greed_nb_coins,
417 raise ValueError(f"Unknown task {args.task}")
419 ######################################################################
421 log_string(f"device {device}")
423 vocabulary_size = task.vocabulary_size()
425 log_string(f"vocabulary_size {vocabulary_size}")
427 ######################################################################
429 # Compute the entropy of the training tokens
432 for input in task.batches(split="train", desc="train-entropy"):
433 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
434 token_probas = token_count / token_count.sum()
435 entropy = -torch.xlogy(token_probas, token_probas).sum()
436 train_set_perplexity = math.exp(entropy)
438 ######################################################################
439 # A bit of paranoia never hurts
441 if args.max_percents_of_test_in_train >= 0:
443 def subsets_as_tuples(batches, cs):
445 for batch in batches:
447 s.add(tuple([v.item() for v in x]))
453 nb_test, nb_in_train = 0, 0
454 for test_subset in subsets_as_tuples(
455 task.batches(split="test", desc="test-check"), 25000
458 for train_subset in subsets_as_tuples(
459 task.batches(split="train", desc="train-check"), 25000
461 in_train.update(test_subset.intersection(train_subset))
462 nb_in_train += len(in_train)
463 nb_test += len(test_subset)
466 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
470 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
471 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
473 ##############################
476 def one_epoch(model, task):
477 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
481 nb_train_samples, acc_train_loss = 0, 0.0
483 for input in task.batches(split="train"):
484 input = input.to(device)
486 if nb_train_samples % args.batch_size == 0:
487 optimizer.zero_grad()
489 output = model(mygpt.BracketedSequence(input)).x
490 loss = F.cross_entropy(output.transpose(1, 2), input)
491 acc_train_loss += loss.item() * input.size(0)
493 nb_train_samples += input.size(0)
497 if nb_train_samples % args.batch_size == 0:
500 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
502 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
505 ######################################################################
508 def run_tests(model, task, deterministic_synthesis):
509 with torch.autograd.no_grad():
512 nb_test_samples, acc_test_loss = 0, 0.0
513 nb_samples_accumulated = 0
515 for input in task.batches(split="test"):
516 input = input.to(device)
518 bs = model(mygpt.BracketedSequence(input))
521 loss = F.cross_entropy(output.transpose(1, 2), input)
523 acc_test_loss += loss.item() * input.size(0)
525 nb_test_samples += input.size(0)
527 main_test_accuracy = task.produce_results(
530 result_dir=args.result_dir,
532 deterministic_synthesis=deterministic_synthesis,
535 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
537 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
539 model.main_test_accuracy = main_test_accuracy
542 ######################################################################
554 while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
555 new_quizzes, nb_correct = task.create_new_quizzes(
557 result_dir=args.result_dir,
559 nb=4 * (nb_for_train + nb_for_test),
561 other_models=other_models,
564 to_keep = new_quizzes[nb_correct == len(other_models) - 1]
565 log_string(f"keep {to_keep.size(0)} quizzes")
568 new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
570 task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
571 task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
576 f"world_new_{n_epoch:04d}_{model.id:02d}.png",
581 ######################################################################
587 vocabulary_size=vocabulary_size,
588 dim_model=args.dim_model,
589 dim_keys=args.dim_keys,
590 dim_hidden=args.dim_hidden,
591 nb_heads=args.nb_heads,
592 nb_blocks=args.nb_blocks,
594 dropout=args.dropout,
597 model.main_test_accuracy = 0.0
603 nb_parameters = sum(p.numel() for p in models[0].parameters())
604 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
606 ######################################################################
608 accuracy_to_make_quizzes = 0.975
610 for n_epoch in range(args.nb_epochs):
611 # select the model with lowest accuracy
612 models.sort(key=lambda model: model.main_test_accuracy)
616 f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
620 one_epoch(model, task)
623 f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
627 run_tests(model, task, deterministic_synthesis=False)
629 if model.main_test_accuracy >= accuracy_to_make_quizzes:
630 other_models = models.copy()
631 other_models.remove(model)
642 ######################################################################