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 # world quizzes vs. culture quizzes
23 ######################################################################
25 if torch.cuda.is_available():
26 device = torch.device("cuda")
27 torch.backends.cuda.matmul.allow_tf32 = True
29 device = torch.device("cpu")
31 ######################################################################
33 parser = argparse.ArgumentParser(
34 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
37 parser.add_argument("--log_filename", type=str, default="train.log")
39 parser.add_argument("--result_dir", type=str, default=None)
41 parser.add_argument("--seed", type=int, default=0)
43 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
45 ########################################
47 parser.add_argument("--nb_epochs", type=int, default=10000)
49 parser.add_argument("--batch_size", type=int, default=None)
51 parser.add_argument("--physical_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("--learning_rate", type=float, default=5e-4)
59 ########################################
61 parser.add_argument("--model", type=str, default=None)
63 parser.add_argument("--dim_model", type=int, default=None)
65 parser.add_argument("--dim_keys", type=int, default=None)
67 parser.add_argument("--dim_hidden", type=int, default=None)
69 parser.add_argument("--nb_heads", type=int, default=None)
71 parser.add_argument("--nb_blocks", type=int, default=None)
73 parser.add_argument("--dropout", type=float, default=0.1)
75 ########################################
77 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
79 parser.add_argument("--problem", type=str, default="grids")
81 parser.add_argument("--nb_threads", type=int, default=1)
83 parser.add_argument("--nb_gpus", type=int, default=1)
85 parser.add_argument("--nb_gpts", type=int, default=5)
87 parser.add_argument("--min_to_validate", type=int, default=None)
89 parser.add_argument("--max_to_validate", type=int, default=None)
91 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
93 parser.add_argument("--generation_temperature", type=float, default=2.0)
95 parser.add_argument("--deterministic_validation", action="store_true", default=False)
97 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
99 parser.add_argument("--dirty_debug", action="store_true", default=False)
101 ######################################################################
103 parser.add_argument("--sky_height", type=int, default=6)
105 parser.add_argument("--sky_width", type=int, default=8)
107 parser.add_argument("--sky_nb_birds", type=int, default=3)
109 parser.add_argument("--sky_nb_iterations", type=int, default=2)
111 parser.add_argument("--sky_speed", type=int, default=3)
113 ######################################################################
115 args = parser.parse_args()
117 if args.min_to_validate is None:
118 args.min_to_validate = args.nb_gpts - 1
120 if args.max_to_validate is None:
121 args.max_to_validate = args.nb_gpts - 1
123 if args.result_dir is None:
124 args.result_dir = f"results_culture"
126 ######################################################################
131 "nb_train_samples": 100000,
132 "nb_test_samples": 10000,
135 for k, v in default_args.items():
136 if getattr(args, k) is None:
139 ######################################################################
141 default_model_args = {
179 if args.model in default_model_args:
180 for k, v in default_model_args[args.model].items():
181 if getattr(args, k) is None:
184 raise ValueError(f"Unknown model {args.model}")
186 ######################################################################
189 os.mkdir(args.result_dir)
190 except FileExistsError:
191 print(f"result directory {args.result_dir} already exists")
194 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
197 # torch.backends.cudnn.deterministic = True
198 # torch.backends.cudnn.benchmark = False
199 # torch.use_deterministic_algorithms(True)
200 torch.manual_seed(args.seed)
201 if torch.cuda.is_available():
202 torch.cuda.manual_seed_all(args.seed)
204 ######################################################################
208 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
210 if log_file is not None:
211 log_file.write(t + s + "\n")
218 log_string(f"argv {' '.join(sys.argv)}")
221 log_string(f"args.{n} {getattr(args, n)}")
224 ######################################################################
227 args.nb_train_samples = 2500
228 args.nb_test_samples = 100
230 if args.physical_batch_size is None:
231 args.physical_batch_size = args.batch_size
233 assert args.batch_size % args.physical_batch_size == 0
235 assert args.nb_train_samples % args.batch_size == 0
236 assert args.nb_test_samples % args.batch_size == 0
238 if args.problem == "sky":
240 height=args.sky_height,
241 width=args.sky_width,
242 nb_birds=args.sky_nb_birds,
243 nb_iterations=args.sky_nb_iterations,
244 speed=args.sky_speed,
245 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
247 nb_threads=args.nb_threads,
249 back_accuracy = False
250 elif args.problem == "grids":
251 problem = grids.Grids(
252 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
254 nb_threads=args.nb_threads,
260 quiz_machine = quiz_machine.QuizMachine(
262 nb_train_samples=args.nb_train_samples,
263 nb_test_samples=args.nb_test_samples,
264 back_accuracy=back_accuracy,
265 batch_size=args.physical_batch_size,
266 result_dir=args.result_dir,
271 ######################################################################
273 log_string(f"device {device}")
275 vocabulary_size = quiz_machine.vocabulary_size()
277 log_string(f"vocabulary_size {vocabulary_size}")
279 ######################################################################
282 ######################################################################
285 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
286 if local_device is None:
287 local_device = device
289 with torch.autograd.no_grad():
290 model.eval().to(local_device)
292 nb_test_samples, acc_test_loss = 0, 0.0
293 nb_samples_accumulated = 0
295 for input in quiz_machine.batches(model, split="test"):
296 input = input.to(local_device)
298 bs = model(mygpt.BracketedSequence(input))
301 loss = F.cross_entropy(output.transpose(1, 2), input)
303 acc_test_loss += loss.item() * input.size(0)
305 nb_test_samples += input.size(0)
307 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
309 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
311 model.main_test_accuracy = quiz_machine.produce_results(
314 result_dir=args.result_dir,
315 deterministic_synthesis=deterministic_synthesis,
319 def one_epoch(model, quiz_machine, local_device=None):
320 if local_device is None:
321 local_device = device
323 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
325 model.to(local_device).train()
327 nb_train_samples, acc_train_loss = 0, 0.0
329 for input in quiz_machine.batches(model, split="train"):
330 input = input.to(local_device)
332 if nb_train_samples % args.batch_size == 0:
333 optimizer.zero_grad()
335 output = model(mygpt.BracketedSequence(input)).x
336 loss = F.cross_entropy(output.transpose(1, 2), input)
337 acc_train_loss += loss.item() * input.size(0)
339 nb_train_samples += input.size(0)
343 if nb_train_samples % args.batch_size == 0:
346 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
348 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
350 run_tests(model, quiz_machine, deterministic_synthesis=False)
352 model.TRAINING_LOCK.release()
355 ######################################################################
358 def standard_validity(logproba):
359 l = logproba.sort(dim=-1).values
360 return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
361 # warnings.warn("TEST!!!", RuntimeWarning)
363 # return (l[:, 0] < math.log(0.99))
366 def valid_c_quizzes(recorded, criteria):
367 result = [q[criteria(lp)] for q, lp in recorded]
368 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
371 ######################################################################
374 def create_c_quizzes(
380 quizzes_and_logproba_records = []
382 nb_to_create = nb_for_train + nb_for_test
384 # ------------------------------------------------------------
386 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
388 with open(file_name, "w") as logp_file:
390 valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
393 # Select a model at random to generate the new quizzes
395 model_for_generation = models[torch.randint(len(models), (1,))]
397 c_quizzes = quiz_machine.generate_quizzes(
399 model_for_generation=model_for_generation,
400 temperature=args.generation_temperature,
403 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
405 if c_quizzes.size(0) > 0:
406 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
408 s = " ".join([str(x.item()) for x in l])
409 logp_file.write(s + "\n")
410 quizzes_and_logproba_records.append((c_quizzes, logproba))
412 nb_validated = valid_c_quizzes(
413 quizzes_and_logproba_records, standard_validity
417 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
420 # store the new c_quizzes which have been validated
422 new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
424 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
426 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
427 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
429 # save a bunch of images to investigate what quizzes with a
430 # certain nb of correct predictions look like
432 q = new_c_quizzes[:72]
435 quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
438 ######################################################################
441 def create_c_quizzes_(
447 quizzes_and_nb_correct_records = []
449 nb_to_create = nb_for_train + nb_for_test
451 # ------------------------------------------------------------
453 standard_validity = lambda nb_correct: (nb_correct >= args.min_to_validate) & (
454 nb_correct <= args.max_to_validate
457 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
459 with open(file_name, "w") as logp_file:
461 valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
464 # Select a model at random to generate the new quizzes
466 model_for_generation = models[torch.randint(len(models), (1,))]
468 c_quizzes = quiz_machine.generate_quizzes(
470 model_for_generation=model_for_generation,
471 temperature=args.generation_temperature,
474 # if args.prediction_correctness:
477 # logproba = quiz_machine.new(quiz_machine.size(0), len(models))
478 # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)):
479 # for model in models:
480 # l[...] = F.cross_entropy(model(q))
482 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
484 if c_quizzes.size(0) > 0:
485 nb_correct, seq_logproba = quiz_machine.compute_correctness(
488 bidirectional_validation=args.bidirectional_validation,
489 deterministic_validation=args.deterministic_validation,
492 for n, l in zip(nb_correct, seq_logproba):
493 s = " ".join([str(x.item()) for x in l])
494 logp_file.write(f"{n} {s}\n")
497 nb_correct = torch.randint(
498 len(models) + 1, nb_correct.size(), device=c_quizzes.device
501 quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
503 nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
504 nv = " ".join([str(x.item()) for x in nv])
506 nb_validated = valid_c_quizzes(
507 quizzes_and_nb_correct_records, standard_validity
511 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
514 # store the new c_quizzes which have been validated
516 new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
518 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
520 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
521 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
523 # save a bunch of images to investigate what quizzes with a
524 # certain nb of correct predictions look like
526 for n in range(len(models) + 1):
529 if n >= args.min_to_validate and n <= args.max_to_validate
534 quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
537 quiz_machine.reverse_random_half_in_place(q)
540 quiz_machine.save_quizzes(
541 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
545 ######################################################################
549 for k in range(args.nb_gpts):
550 log_string(f"creating model {k} and its w_quizzes")
552 vocabulary_size=vocabulary_size,
553 dim_model=args.dim_model,
554 dim_keys=args.dim_keys,
555 dim_hidden=args.dim_hidden,
556 nb_heads=args.nb_heads,
557 nb_blocks=args.nb_blocks,
559 dropout=args.dropout,
562 model.main_test_accuracy = 0.0
564 model.TRAINING_LOCK = threading.Lock()
566 model.train_w_quizzes = quiz_machine.generate_token_sequences(
567 args.nb_train_samples
569 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
570 model.test_w_quizzes = quiz_machine.generate_token_sequences(
573 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
578 nb_parameters = sum(p.numel() for p in models[0].parameters())
579 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
581 ######################################################################
583 # Compute the entropy of the training tokens
586 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
587 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
590 token_probas = token_count / token_count.sum()
591 entropy = -torch.xlogy(token_probas, token_probas).sum()
592 train_set_perplexity = math.exp(entropy)
594 ######################################################################
595 # A bit of paranoia never hurts
597 if args.max_percents_of_test_in_train >= 0:
599 def subsets_as_tuples(batches, cs):
601 for batch in batches:
603 s.add(tuple([v.item() for v in x]))
609 nb_test, nb_in_train = 0, 0
610 for test_subset in subsets_as_tuples(
611 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
614 for train_subset in subsets_as_tuples(
615 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
617 in_train.update(test_subset.intersection(train_subset))
618 nb_in_train += len(in_train)
619 nb_test += len(test_subset)
622 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
626 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
627 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
629 ######################################################################
631 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
632 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
635 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}"
638 ######################################################################
641 args.accuracy_to_make_c_quizzes = 0.0
643 nb_new_c_quizzes_for_train = 100
644 nb_new_c_quizzes_for_test = 10
646 ######################################################################
648 for n_epoch in range(args.nb_epochs):
649 log_string(f"--- epoch {n_epoch} ----------------------------------------")
651 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
652 log_string(f"current_test_accuracies {cta}")
654 ##################################################
655 # Select, improve, and eval the worst model
657 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
659 weakest_models = ranked_models[: args.nb_gpus]
661 for gpu_id, model in enumerate(weakest_models):
662 model.TRAINING_LOCK.acquire()
665 f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
669 target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
672 for model in weakest_models:
673 model.TRAINING_LOCK.acquire()
674 model.TRAINING_LOCK.release()
676 ##################################################
677 # Replace a fraction of the w_quizzes with fresh ones
680 f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
683 # Renew entirely the train set
685 for model in weakest_models:
686 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
688 ##################################################
689 # If all the models are good enough, generate new quizzes and
690 # re-compute the test errors
692 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
696 nb_for_train=nb_new_c_quizzes_for_train,
697 nb_for_test=nb_new_c_quizzes_for_test,
700 ######################################################################