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
17 import sky, grids, quiz_machine
21 import torch.multiprocessing as mp
23 # world quizzes vs. culture quizzes
25 ######################################################################
27 if torch.cuda.is_available():
28 device = torch.device("cuda")
29 torch.backends.cuda.matmul.allow_tf32 = True
31 device = torch.device("cpu")
33 ######################################################################
35 parser = argparse.ArgumentParser(
36 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
39 parser.add_argument("--log_filename", type=str, default="train.log")
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=10000)
51 parser.add_argument("--batch_size", type=int, default=None)
53 parser.add_argument("--physical_batch_size", type=int, default=None)
55 parser.add_argument("--nb_train_samples", type=int, default=None)
57 parser.add_argument("--nb_test_samples", type=int, default=None)
59 parser.add_argument("--learning_rate", type=float, default=5e-4)
61 ########################################
63 parser.add_argument("--model", type=str, default=None)
65 parser.add_argument("--dim_model", type=int, default=None)
67 parser.add_argument("--dim_keys", type=int, default=None)
69 parser.add_argument("--dim_hidden", type=int, default=None)
71 parser.add_argument("--nb_heads", type=int, default=None)
73 parser.add_argument("--nb_blocks", type=int, default=None)
75 parser.add_argument("--dropout", type=float, default=0.1)
77 ########################################
79 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
81 parser.add_argument("--problem", type=str, default="grids")
83 parser.add_argument("--nb_threads", type=int, default=1)
85 parser.add_argument("--nb_gpus", type=int, default=1)
87 parser.add_argument("--nb_gpts", type=int, default=5)
89 parser.add_argument("--min_to_validate", type=int, default=None)
91 parser.add_argument("--max_to_validate", type=int, default=None)
93 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
95 parser.add_argument("--generation_temperature", type=float, default=2.0)
97 parser.add_argument("--dirty_debug", action="store_true", default=False)
99 ######################################################################
101 grids_tasks = ", ".join(
102 [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
109 help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
112 ######################################################################
114 parser.add_argument("--sky_height", type=int, default=6)
116 parser.add_argument("--sky_width", type=int, default=8)
118 parser.add_argument("--sky_nb_birds", type=int, default=3)
120 parser.add_argument("--sky_nb_iterations", type=int, default=2)
122 parser.add_argument("--sky_speed", type=int, default=3)
124 ######################################################################
126 args = parser.parse_args()
128 if args.min_to_validate is None:
129 args.min_to_validate = args.nb_gpts - 1
131 if args.max_to_validate is None:
132 args.max_to_validate = args.nb_gpts - 1
134 if args.result_dir is None:
135 args.result_dir = f"results_culture"
137 ######################################################################
142 "nb_train_samples": 100000,
143 "nb_test_samples": 10000,
146 for k, v in default_args.items():
147 if getattr(args, k) is None:
150 ######################################################################
152 default_model_args = {
190 if args.model in default_model_args:
191 for k, v in default_model_args[args.model].items():
192 if getattr(args, k) is None:
195 raise ValueError(f"Unknown model {args.model}")
197 ######################################################################
200 os.mkdir(args.result_dir)
201 except FileExistsError:
202 print(f"result directory {args.result_dir} already exists")
205 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
208 # torch.backends.cudnn.deterministic = True
209 # torch.backends.cudnn.benchmark = False
210 # torch.use_deterministic_algorithms(True)
211 torch.manual_seed(args.seed)
212 if torch.cuda.is_available():
213 torch.cuda.manual_seed_all(args.seed)
215 ######################################################################
219 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
221 if log_file is not None:
222 log_file.write(t + s + "\n")
229 log_string(f"argv {' '.join(sys.argv)}")
232 log_string(f"args.{n} {getattr(args, n)}")
235 ######################################################################
238 args.nb_train_samples = 2500
239 args.nb_test_samples = 100
241 if args.physical_batch_size is None:
242 args.physical_batch_size = args.batch_size
244 assert args.batch_size % args.physical_batch_size == 0
246 assert args.nb_train_samples % args.batch_size == 0
247 assert args.nb_test_samples % args.batch_size == 0
249 if args.problem == "sky":
251 height=args.sky_height,
252 width=args.sky_width,
253 nb_birds=args.sky_nb_birds,
254 nb_iterations=args.sky_nb_iterations,
255 speed=args.sky_speed,
256 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
258 nb_threads=args.nb_threads,
260 back_accuracy = False
261 elif args.problem == "grids":
262 problem = grids.Grids(
263 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
265 nb_threads=args.nb_threads,
266 tasks=args.grids_tasks,
272 problem.save_some_examples(args.result_dir)
274 quiz_machine = quiz_machine.QuizMachine(
276 nb_train_samples=args.nb_train_samples,
277 nb_test_samples=args.nb_test_samples,
278 back_accuracy=back_accuracy,
279 batch_size=args.physical_batch_size,
280 result_dir=args.result_dir,
285 ######################################################################
287 log_string(f"device {device}")
289 vocabulary_size = quiz_machine.vocabulary_size()
291 log_string(f"vocabulary_size {vocabulary_size}")
293 ######################################################################
296 ######################################################################
299 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
300 if local_device is None:
301 local_device = device
303 with torch.autograd.no_grad():
304 model.eval().to(local_device)
306 nb_test_samples, acc_test_loss = 0, 0.0
307 nb_samples_accumulated = 0
309 for input in quiz_machine.batches(model, split="test"):
310 input = input.to(local_device)
312 bs = model(mygpt.BracketedSequence(input))
315 loss = F.cross_entropy(output.transpose(1, 2), input)
317 acc_test_loss += loss.item() * input.size(0)
319 nb_test_samples += input.size(0)
321 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
323 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
325 model.main_test_accuracy = quiz_machine.produce_results(
328 result_dir=args.result_dir,
329 deterministic_synthesis=deterministic_synthesis,
333 def one_epoch(model, quiz_machine, local_device=None):
334 if local_device is None:
335 local_device = device
337 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
339 model.to(local_device).train()
341 nb_train_samples, acc_train_loss = 0, 0.0
343 for input in quiz_machine.batches(model, split="train"):
344 input = input.to(local_device)
346 if nb_train_samples % args.batch_size == 0:
347 optimizer.zero_grad()
349 output = model(mygpt.BracketedSequence(input)).x
350 loss = F.cross_entropy(output.transpose(1, 2), input)
351 acc_train_loss += loss.item() * input.size(0)
353 nb_train_samples += input.size(0)
357 if nb_train_samples % args.batch_size == 0:
360 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
362 log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
364 run_tests(model, quiz_machine, deterministic_synthesis=False)
367 ######################################################################
370 def standard_validity(logproba):
371 l = logproba.sort(dim=-1).values
372 return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
375 def valid_c_quizzes(recorded, criteria):
376 result = [q[criteria(lp)] for q, lp in recorded]
377 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
380 ######################################################################
383 def create_c_quizzes(
389 quizzes_and_logproba_records = []
391 nb_to_create = nb_for_train + nb_for_test
393 # ------------------------------------------------------------
395 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
397 with open(file_name, "w") as logp_file:
399 valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
402 # Select a model at random to generate the new quizzes
404 model_for_generation = models[torch.randint(len(models), (1,))]
406 c_quizzes = quiz_machine.generate_quizzes(
408 model_for_generation=model_for_generation,
409 temperature=args.generation_temperature,
412 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
414 if c_quizzes.size(0) > 0:
415 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
417 s = " ".join([str(x.item()) for x in l])
418 logp_file.write(s + "\n")
419 quizzes_and_logproba_records.append((c_quizzes, logproba))
421 nb_validated = valid_c_quizzes(
422 quizzes_and_logproba_records, standard_validity
426 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
429 # store the new c_quizzes which have been validated
431 new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
433 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
435 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
436 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
438 # save a bunch of images to investigate what quizzes with a
439 # certain nb of correct predictions look like
441 q = new_c_quizzes[:72]
444 quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
447 ######################################################################
451 for k in range(args.nb_gpts):
452 log_string(f"creating model {k} and its w_quizzes")
454 vocabulary_size=vocabulary_size,
455 dim_model=args.dim_model,
456 dim_keys=args.dim_keys,
457 dim_hidden=args.dim_hidden,
458 nb_heads=args.nb_heads,
459 nb_blocks=args.nb_blocks,
461 dropout=args.dropout,
464 model.main_test_accuracy = 0.0
467 model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
468 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
469 model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
470 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
475 nb_parameters = sum(p.numel() for p in models[0].parameters())
476 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
478 ######################################################################
480 # Compute the entropy of the training tokens
483 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
484 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
487 token_probas = token_count / token_count.sum()
488 entropy = -torch.xlogy(token_probas, token_probas).sum()
489 train_set_perplexity = math.exp(entropy)
491 ######################################################################
492 # A bit of paranoia never hurts
494 if args.max_percents_of_test_in_train >= 0:
496 def subsets_as_tuples(batches, cs):
498 for batch in batches:
500 s.add(tuple([v.item() for v in x]))
506 nb_test, nb_in_train = 0, 0
507 for test_subset in subsets_as_tuples(
508 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
511 for train_subset in subsets_as_tuples(
512 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
514 in_train.update(test_subset.intersection(train_subset))
515 nb_in_train += len(in_train)
516 nb_test += len(test_subset)
519 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
523 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
524 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
526 ######################################################################
528 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
529 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
532 f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}"
535 ######################################################################
538 args.accuracy_to_make_c_quizzes = 0.0
540 nb_new_c_quizzes_for_train = 100
541 nb_new_c_quizzes_for_test = 10
543 def standard_validity(logproba):
544 l = logproba.sort(dim=-1).values
545 return l[:, 0] < math.log(0.5)
548 ######################################################################
550 for n_epoch in range(args.nb_epochs):
551 log_string(f"--- epoch {n_epoch} ----------------------------------------")
553 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
554 log_string(f"current_test_accuracies {cta}")
556 ##################################################
557 # Select, improve, and eval the worst model
559 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
561 weakest_models = ranked_models[: args.nb_gpus]
565 for gpu_id, model in enumerate(weakest_models):
566 log_string(f"training model {model.id}")
568 t = threading.Thread(
569 target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
579 ##################################################
580 # Replace a fraction of the w_quizzes with fresh ones
583 f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
586 # Renew entirely the train set
588 for model in weakest_models:
589 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
591 ##################################################
592 # If all the models are good enough, generate new quizzes and
593 # re-compute the test errors
595 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
599 nb_for_train=nb_new_c_quizzes_for_train,
600 nb_for_test=nb_new_c_quizzes_for_test,
603 ######################################################################