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 parser.add_argument("--sky_height", type=int, default=6)
103 parser.add_argument("--sky_width", type=int, default=8)
105 parser.add_argument("--sky_nb_birds", type=int, default=3)
107 parser.add_argument("--sky_nb_iterations", type=int, default=2)
109 parser.add_argument("--sky_speed", type=int, default=3)
111 ######################################################################
113 args = parser.parse_args()
115 if args.min_to_validate is None:
116 args.min_to_validate = args.nb_gpts - 1
118 if args.max_to_validate is None:
119 args.max_to_validate = args.nb_gpts - 1
121 if args.result_dir is None:
122 args.result_dir = f"results_culture"
124 ######################################################################
129 "nb_train_samples": 100000,
130 "nb_test_samples": 10000,
133 for k, v in default_args.items():
134 if getattr(args, k) is None:
137 ######################################################################
139 default_model_args = {
177 if args.model in default_model_args:
178 for k, v in default_model_args[args.model].items():
179 if getattr(args, k) is None:
182 raise ValueError(f"Unknown model {args.model}")
184 ######################################################################
187 os.mkdir(args.result_dir)
188 except FileExistsError:
189 print(f"result directory {args.result_dir} already exists")
192 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
195 # torch.backends.cudnn.deterministic = True
196 # torch.backends.cudnn.benchmark = False
197 # torch.use_deterministic_algorithms(True)
198 torch.manual_seed(args.seed)
199 if torch.cuda.is_available():
200 torch.cuda.manual_seed_all(args.seed)
202 ######################################################################
206 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
208 if log_file is not None:
209 log_file.write(t + s + "\n")
216 log_string(f"argv {' '.join(sys.argv)}")
219 log_string(f"args.{n} {getattr(args, n)}")
222 ######################################################################
225 args.nb_train_samples = 2500
226 args.nb_test_samples = 100
228 if args.physical_batch_size is None:
229 args.physical_batch_size = args.batch_size
231 assert args.batch_size % args.physical_batch_size == 0
233 assert args.nb_train_samples % args.batch_size == 0
234 assert args.nb_test_samples % args.batch_size == 0
236 if args.problem == "sky":
238 height=args.sky_height,
239 width=args.sky_width,
240 nb_birds=args.sky_nb_birds,
241 nb_iterations=args.sky_nb_iterations,
242 speed=args.sky_speed,
243 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
245 nb_threads=args.nb_threads,
247 back_accuracy = False
248 elif args.problem == "grids":
249 problem = grids.Grids(
250 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
252 nb_threads=args.nb_threads,
258 quiz_machine = quiz_machine.QuizMachine(
260 nb_train_samples=args.nb_train_samples,
261 nb_test_samples=args.nb_test_samples,
262 back_accuracy=back_accuracy,
263 batch_size=args.physical_batch_size,
264 result_dir=args.result_dir,
269 ######################################################################
271 log_string(f"device {device}")
273 vocabulary_size = quiz_machine.vocabulary_size()
275 log_string(f"vocabulary_size {vocabulary_size}")
277 ######################################################################
280 ######################################################################
283 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
284 if local_device is None:
285 local_device = device
287 with torch.autograd.no_grad():
288 model.eval().to(local_device)
290 nb_test_samples, acc_test_loss = 0, 0.0
291 nb_samples_accumulated = 0
293 for input in quiz_machine.batches(model, split="test"):
294 input = input.to(local_device)
296 bs = model(mygpt.BracketedSequence(input))
299 loss = F.cross_entropy(output.transpose(1, 2), input)
301 acc_test_loss += loss.item() * input.size(0)
303 nb_test_samples += input.size(0)
305 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
307 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
309 model.main_test_accuracy = quiz_machine.produce_results(
312 result_dir=args.result_dir,
313 deterministic_synthesis=deterministic_synthesis,
317 def one_epoch(model, quiz_machine, local_device=None):
318 if local_device is None:
319 local_device = device
321 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
323 model.to(local_device).train()
325 nb_train_samples, acc_train_loss = 0, 0.0
327 for input in quiz_machine.batches(model, split="train"):
328 input = input.to(local_device)
330 if nb_train_samples % args.batch_size == 0:
331 optimizer.zero_grad()
333 output = model(mygpt.BracketedSequence(input)).x
334 loss = F.cross_entropy(output.transpose(1, 2), input)
335 acc_train_loss += loss.item() * input.size(0)
337 nb_train_samples += input.size(0)
341 if nb_train_samples % args.batch_size == 0:
344 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
346 log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
348 run_tests(model, quiz_machine, deterministic_synthesis=False)
351 ######################################################################
354 def standard_validity(logproba):
355 l = logproba.sort(dim=-1).values
356 return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
359 def valid_c_quizzes(recorded, criteria):
360 result = [q[criteria(lp)] for q, lp in recorded]
361 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
364 ######################################################################
367 def create_c_quizzes(
373 quizzes_and_logproba_records = []
375 nb_to_create = nb_for_train + nb_for_test
377 # ------------------------------------------------------------
379 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
381 with open(file_name, "w") as logp_file:
383 valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
386 # Select a model at random to generate the new quizzes
388 model_for_generation = models[torch.randint(len(models), (1,))]
390 c_quizzes = quiz_machine.generate_quizzes(
392 model_for_generation=model_for_generation,
393 temperature=args.generation_temperature,
396 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
398 if c_quizzes.size(0) > 0:
399 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
401 s = " ".join([str(x.item()) for x in l])
402 logp_file.write(s + "\n")
403 quizzes_and_logproba_records.append((c_quizzes, logproba))
405 nb_validated = valid_c_quizzes(
406 quizzes_and_logproba_records, standard_validity
410 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
413 # store the new c_quizzes which have been validated
415 new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
417 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
419 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
420 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
422 # save a bunch of images to investigate what quizzes with a
423 # certain nb of correct predictions look like
425 q = new_c_quizzes[:72]
428 quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
431 ######################################################################
435 for k in range(args.nb_gpts):
436 log_string(f"creating model {k} and its w_quizzes")
438 vocabulary_size=vocabulary_size,
439 dim_model=args.dim_model,
440 dim_keys=args.dim_keys,
441 dim_hidden=args.dim_hidden,
442 nb_heads=args.nb_heads,
443 nb_blocks=args.nb_blocks,
445 dropout=args.dropout,
448 model.main_test_accuracy = 0.0
451 model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
452 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
453 model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
454 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
459 nb_parameters = sum(p.numel() for p in models[0].parameters())
460 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
462 ######################################################################
464 # Compute the entropy of the training tokens
467 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
468 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
471 token_probas = token_count / token_count.sum()
472 entropy = -torch.xlogy(token_probas, token_probas).sum()
473 train_set_perplexity = math.exp(entropy)
475 ######################################################################
476 # A bit of paranoia never hurts
478 if args.max_percents_of_test_in_train >= 0:
480 def subsets_as_tuples(batches, cs):
482 for batch in batches:
484 s.add(tuple([v.item() for v in x]))
490 nb_test, nb_in_train = 0, 0
491 for test_subset in subsets_as_tuples(
492 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
495 for train_subset in subsets_as_tuples(
496 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
498 in_train.update(test_subset.intersection(train_subset))
499 nb_in_train += len(in_train)
500 nb_test += len(test_subset)
503 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
507 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
508 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
510 ######################################################################
512 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
513 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
516 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}"
519 ######################################################################
522 args.accuracy_to_make_c_quizzes = 0.0
524 nb_new_c_quizzes_for_train = 100
525 nb_new_c_quizzes_for_test = 10
527 def standard_validity(logproba):
528 l = logproba.sort(dim=-1).values
529 return l[:, 0] < math.log(0.5)
532 ######################################################################
534 for n_epoch in range(args.nb_epochs):
535 log_string(f"--- epoch {n_epoch} ----------------------------------------")
537 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
538 log_string(f"current_test_accuracies {cta}")
540 ##################################################
541 # Select, improve, and eval the worst model
543 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
545 weakest_models = ranked_models[: args.nb_gpus]
549 for gpu_id, model in enumerate(weakest_models):
550 log_string(f"training model {model.id}")
552 t = threading.Thread(
553 target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
563 ##################################################
564 # Replace a fraction of the w_quizzes with fresh ones
567 f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
570 # Renew entirely the train set
572 for model in weakest_models:
573 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
575 ##################################################
576 # If all the models are good enough, generate new quizzes and
577 # re-compute the test errors
579 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
583 nb_for_train=nb_new_c_quizzes_for_train,
584 nb_for_test=nb_new_c_quizzes_for_test,
587 ######################################################################