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=0.9)
93 parser.add_argument("--generation_temperature", type=float, default=2.0)
95 parser.add_argument("--dirty_debug", action="store_true", default=False)
97 ######################################################################
99 parser.add_argument("--sky_height", type=int, default=6)
101 parser.add_argument("--sky_width", type=int, default=8)
103 parser.add_argument("--sky_nb_birds", type=int, default=3)
105 parser.add_argument("--sky_nb_iterations", type=int, default=2)
107 parser.add_argument("--sky_speed", type=int, default=3)
109 ######################################################################
111 args = parser.parse_args()
113 if args.min_to_validate is None:
114 args.min_to_validate = args.nb_gpts - 1
116 if args.max_to_validate is None:
117 args.max_to_validate = args.nb_gpts - 1
119 if args.result_dir is None:
120 args.result_dir = f"results_culture"
122 ######################################################################
127 "nb_train_samples": 100000,
128 "nb_test_samples": 10000,
131 for k, v in default_args.items():
132 if getattr(args, k) is None:
135 ######################################################################
137 default_model_args = {
175 if args.model in default_model_args:
176 for k, v in default_model_args[args.model].items():
177 if getattr(args, k) is None:
180 raise ValueError(f"Unknown model {args.model}")
182 ######################################################################
185 os.mkdir(args.result_dir)
186 except FileExistsError:
187 print(f"result directory {args.result_dir} already exists")
190 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
193 # torch.backends.cudnn.deterministic = True
194 # torch.backends.cudnn.benchmark = False
195 # torch.use_deterministic_algorithms(True)
196 torch.manual_seed(args.seed)
197 if torch.cuda.is_available():
198 torch.cuda.manual_seed_all(args.seed)
200 ######################################################################
204 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
206 if log_file is not None:
207 log_file.write(t + s + "\n")
214 log_string(f"argv {' '.join(sys.argv)}")
217 log_string(f"args.{n} {getattr(args, n)}")
220 ######################################################################
223 args.nb_train_samples = 2500
224 args.nb_test_samples = 100
226 if args.physical_batch_size is None:
227 args.physical_batch_size = args.batch_size
229 assert args.batch_size % args.physical_batch_size == 0
231 assert args.nb_train_samples % args.batch_size == 0
232 assert args.nb_test_samples % args.batch_size == 0
234 if args.problem == "sky":
236 height=args.sky_height,
237 width=args.sky_width,
238 nb_birds=args.sky_nb_birds,
239 nb_iterations=args.sky_nb_iterations,
240 speed=args.sky_speed,
241 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
243 nb_threads=args.nb_threads,
245 back_accuracy = False
246 elif args.problem == "grids":
247 problem = grids.Grids(
248 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
250 nb_threads=args.nb_threads,
256 quiz_machine = quiz_machine.QuizMachine(
258 nb_train_samples=args.nb_train_samples,
259 nb_test_samples=args.nb_test_samples,
260 back_accuracy=back_accuracy,
261 batch_size=args.physical_batch_size,
262 result_dir=args.result_dir,
267 ######################################################################
269 log_string(f"device {device}")
271 vocabulary_size = quiz_machine.vocabulary_size()
273 log_string(f"vocabulary_size {vocabulary_size}")
275 ######################################################################
278 ######################################################################
281 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
282 if local_device is None:
283 local_device = device
285 with torch.autograd.no_grad():
286 model.eval().to(local_device)
288 nb_test_samples, acc_test_loss = 0, 0.0
289 nb_samples_accumulated = 0
291 for input in quiz_machine.batches(model, split="test"):
292 input = input.to(local_device)
294 bs = model(mygpt.BracketedSequence(input))
297 loss = F.cross_entropy(output.transpose(1, 2), input)
299 acc_test_loss += loss.item() * input.size(0)
301 nb_test_samples += input.size(0)
303 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
305 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
307 model.main_test_accuracy = quiz_machine.produce_results(
310 result_dir=args.result_dir,
311 deterministic_synthesis=deterministic_synthesis,
315 def one_epoch(model, quiz_machine, local_device=None):
316 if local_device is None:
317 local_device = device
319 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
321 model.to(local_device).train()
323 nb_train_samples, acc_train_loss = 0, 0.0
325 for input in quiz_machine.batches(model, split="train"):
326 input = input.to(local_device)
328 if nb_train_samples % args.batch_size == 0:
329 optimizer.zero_grad()
331 output = model(mygpt.BracketedSequence(input)).x
332 loss = F.cross_entropy(output.transpose(1, 2), input)
333 acc_train_loss += loss.item() * input.size(0)
335 nb_train_samples += input.size(0)
339 if nb_train_samples % args.batch_size == 0:
342 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
344 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
346 run_tests(model, quiz_machine, deterministic_synthesis=False)
348 model.TRAINING_LOCK.release()
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))
357 # warnings.warn("TEST!!!", RuntimeWarning)
359 # return (l[:, 0] < math.log(0.99))
362 def valid_c_quizzes(recorded, criteria):
363 result = [q[criteria(lp)] for q, lp in recorded]
364 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
367 ######################################################################
370 def create_c_quizzes(
376 quizzes_and_logproba_records = []
378 nb_to_create = nb_for_train + nb_for_test
380 # ------------------------------------------------------------
382 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
384 with open(file_name, "w") as logp_file:
386 valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
389 # Select a model at random to generate the new quizzes
391 model_for_generation = models[torch.randint(len(models), (1,))]
393 c_quizzes = quiz_machine.generate_quizzes(
395 model_for_generation=model_for_generation,
396 temperature=args.generation_temperature,
399 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
401 if c_quizzes.size(0) > 0:
402 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
404 s = " ".join([str(x.item()) for x in l])
405 logp_file.write(s + "\n")
406 quizzes_and_logproba_records.append((c_quizzes, logproba))
408 nb_validated = valid_c_quizzes(
409 quizzes_and_logproba_records, standard_validity
413 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
416 # store the new c_quizzes which have been validated
418 new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
420 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
422 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
423 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
425 # save a bunch of images to investigate what quizzes with a
426 # certain nb of correct predictions look like
428 q = new_c_quizzes[:72]
431 quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
434 ######################################################################
438 for k in range(args.nb_gpts):
439 log_string(f"creating model {k} and its w_quizzes")
441 vocabulary_size=vocabulary_size,
442 dim_model=args.dim_model,
443 dim_keys=args.dim_keys,
444 dim_hidden=args.dim_hidden,
445 nb_heads=args.nb_heads,
446 nb_blocks=args.nb_blocks,
448 dropout=args.dropout,
451 model.main_test_accuracy = 0.0
453 model.TRAINING_LOCK = threading.Lock()
455 model.train_w_quizzes = quiz_machine.generate_token_sequences(
456 args.nb_train_samples
458 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
459 model.test_w_quizzes = quiz_machine.generate_token_sequences(
462 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
467 nb_parameters = sum(p.numel() for p in models[0].parameters())
468 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
470 ######################################################################
472 # Compute the entropy of the training tokens
475 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
476 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
479 token_probas = token_count / token_count.sum()
480 entropy = -torch.xlogy(token_probas, token_probas).sum()
481 train_set_perplexity = math.exp(entropy)
483 ######################################################################
484 # A bit of paranoia never hurts
486 if args.max_percents_of_test_in_train >= 0:
488 def subsets_as_tuples(batches, cs):
490 for batch in batches:
492 s.add(tuple([v.item() for v in x]))
498 nb_test, nb_in_train = 0, 0
499 for test_subset in subsets_as_tuples(
500 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
503 for train_subset in subsets_as_tuples(
504 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
506 in_train.update(test_subset.intersection(train_subset))
507 nb_in_train += len(in_train)
508 nb_test += len(test_subset)
511 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
515 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
516 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
518 ######################################################################
520 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
521 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
524 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}"
527 ######################################################################
530 args.accuracy_to_make_c_quizzes = 0.0
532 nb_new_c_quizzes_for_train = 100
533 nb_new_c_quizzes_for_test = 10
535 def standard_validity(logproba):
536 l = logproba.sort(dim=-1).values
537 return l[:, 0] < math.log(0.99)
540 ######################################################################
542 for n_epoch in range(args.nb_epochs):
543 log_string(f"--- epoch {n_epoch} ----------------------------------------")
545 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
546 log_string(f"current_test_accuracies {cta}")
548 ##################################################
549 # Select, improve, and eval the worst models
551 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
553 weakest_models = ranked_models[: args.nb_gpus]
555 for gpu_id, model in enumerate(weakest_models):
556 model.TRAINING_LOCK.acquire()
559 f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
563 target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
566 for model in weakest_models:
567 model.TRAINING_LOCK.acquire()
568 model.TRAINING_LOCK.release()
570 ##################################################
571 # Renew the train sets
574 f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
577 for model in weakest_models:
578 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
580 ##################################################
581 # If all the models are good enough, generate new quizzes and
582 # re-compute the test errors
584 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
588 nb_for_train=nb_new_c_quizzes_for_train,
589 nb_for_test=nb_new_c_quizzes_for_test,
592 ######################################################################