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
19 import torch.multiprocessing as mp
21 # mp.set_start_method('spawn')
23 # world quizzes vs. culture quizzes
25 ######################################################################
29 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
31 if log_file is not None:
32 log_file.write(t + s + "\n")
39 ######################################################################
42 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
43 if local_device is None:
46 with torch.autograd.no_grad():
47 model.eval().to(local_device)
49 nb_test_samples, acc_test_loss = 0, 0.0
50 nb_samples_accumulated = 0
52 for input in quiz_machine.batches(model, split="test"):
53 input = input.to(local_device)
55 bs = model(mygpt.BracketedSequence(input))
58 loss = F.cross_entropy(output.transpose(1, 2), input)
60 acc_test_loss += loss.item() * input.size(0)
62 nb_test_samples += input.size(0)
64 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
66 log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
68 model.main_test_accuracy = quiz_machine.produce_results(
71 result_dir=args.result_dir,
72 deterministic_synthesis=deterministic_synthesis,
76 def one_epoch(model, quiz_machine, local_device=None):
77 if local_device is None:
80 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
82 model.to(local_device).train()
84 nb_train_samples, acc_train_loss = 0, 0.0
86 for input in quiz_machine.batches(model, split="train"):
87 input = input.to(local_device)
89 if nb_train_samples % args.batch_size == 0:
92 output = model(mygpt.BracketedSequence(input)).x
93 loss = F.cross_entropy(output.transpose(1, 2), input)
94 acc_train_loss += loss.item() * input.size(0)
96 nb_train_samples += input.size(0)
100 if nb_train_samples % args.batch_size == 0:
103 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
105 log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
107 run_tests(model, quiz_machine, deterministic_synthesis=False)
110 ######################################################################
113 def standard_validity(logproba):
114 l = logproba.sort(dim=-1).values
115 return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
116 # warnings.warn("TEST!!!", RuntimeWarning)
118 # return (l[:, 0] < math.log(0.99))
121 def valid_c_quizzes(recorded, criteria):
122 result = [q[criteria(lp)] for q, lp in recorded]
123 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
126 ######################################################################
129 def create_c_quizzes(
135 quizzes_and_logproba_records = []
137 nb_to_create = nb_for_train + nb_for_test
139 # ------------------------------------------------------------
141 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
143 with open(file_name, "w") as logp_file:
145 valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
148 # Select a model at random to generate the new quizzes
150 model_for_generation = models[torch.randint(len(models), (1,))]
152 c_quizzes = quiz_machine.generate_quizzes(
154 model_for_generation=model_for_generation,
155 temperature=args.generation_temperature,
158 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
160 if c_quizzes.size(0) > 0:
161 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
163 s = " ".join([str(x.item()) for x in l])
164 logp_file.write(s + "\n")
165 quizzes_and_logproba_records.append((c_quizzes, logproba))
167 nb_validated = valid_c_quizzes(
168 quizzes_and_logproba_records, standard_validity
172 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
175 # store the new c_quizzes which have been validated
177 new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
179 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
181 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
182 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
184 # save a bunch of images to investigate what quizzes with a
185 # certain nb of correct predictions look like
187 q = new_c_quizzes[:72]
190 quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
193 ######################################################################
195 if torch.cuda.is_available():
196 device = torch.device("cuda")
197 torch.backends.cuda.matmul.allow_tf32 = True
199 device = torch.device("cpu")
201 ######################################################################
203 parser = argparse.ArgumentParser(
204 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
207 parser.add_argument("--log_filename", type=str, default="train.log")
209 parser.add_argument("--result_dir", type=str, default=None)
211 parser.add_argument("--seed", type=int, default=0)
213 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
215 ########################################
217 parser.add_argument("--nb_epochs", type=int, default=10000)
219 parser.add_argument("--batch_size", type=int, default=None)
221 parser.add_argument("--physical_batch_size", type=int, default=None)
223 parser.add_argument("--nb_train_samples", type=int, default=None)
225 parser.add_argument("--nb_test_samples", type=int, default=None)
227 parser.add_argument("--learning_rate", type=float, default=5e-4)
229 ########################################
231 parser.add_argument("--model", type=str, default=None)
233 parser.add_argument("--dim_model", type=int, default=None)
235 parser.add_argument("--dim_keys", type=int, default=None)
237 parser.add_argument("--dim_hidden", type=int, default=None)
239 parser.add_argument("--nb_heads", type=int, default=None)
241 parser.add_argument("--nb_blocks", type=int, default=None)
243 parser.add_argument("--dropout", type=float, default=0.1)
245 ########################################
247 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
249 parser.add_argument("--problem", type=str, default="grids")
251 parser.add_argument("--nb_threads", type=int, default=1)
253 parser.add_argument("--nb_gpus", type=int, default=1)
255 parser.add_argument("--nb_gpts", type=int, default=5)
257 parser.add_argument("--min_to_validate", type=int, default=None)
259 parser.add_argument("--max_to_validate", type=int, default=None)
261 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.9)
263 parser.add_argument("--generation_temperature", type=float, default=2.0)
265 parser.add_argument("--dirty_debug", action="store_true", default=False)
267 ######################################################################
269 parser.add_argument("--sky_height", type=int, default=6)
271 parser.add_argument("--sky_width", type=int, default=8)
273 parser.add_argument("--sky_nb_birds", type=int, default=3)
275 parser.add_argument("--sky_nb_iterations", type=int, default=2)
277 parser.add_argument("--sky_speed", type=int, default=3)
279 ######################################################################
281 args = parser.parse_args()
283 if args.min_to_validate is None:
284 args.min_to_validate = args.nb_gpts - 1
286 if args.max_to_validate is None:
287 args.max_to_validate = args.nb_gpts - 1
289 if args.result_dir is None:
290 args.result_dir = f"results_culture"
292 ######################################################################
297 "nb_train_samples": 100000,
298 "nb_test_samples": 10000,
301 for k, v in default_args.items():
302 if getattr(args, k) is None:
305 ######################################################################
307 default_model_args = {
345 if args.model in default_model_args:
346 for k, v in default_model_args[args.model].items():
347 if getattr(args, k) is None:
350 raise ValueError(f"Unknown model {args.model}")
352 ######################################################################
355 os.mkdir(args.result_dir)
356 except FileExistsError:
357 print(f"result directory {args.result_dir} already exists")
360 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
362 log_string(f"argv {' '.join(sys.argv)}")
365 log_string(f"args.{n} {getattr(args, n)}")
368 # torch.backends.cudnn.deterministic = True
369 # torch.backends.cudnn.benchmark = False
370 # torch.use_deterministic_algorithms(True)
371 torch.manual_seed(args.seed)
372 if torch.cuda.is_available():
373 torch.cuda.manual_seed_all(args.seed)
375 ######################################################################
378 args.nb_train_samples = 2500
379 args.nb_test_samples = 100
381 if args.physical_batch_size is None:
382 args.physical_batch_size = args.batch_size
384 assert args.batch_size % args.physical_batch_size == 0
386 assert args.nb_train_samples % args.batch_size == 0
387 assert args.nb_test_samples % args.batch_size == 0
389 if args.problem == "sky":
391 height=args.sky_height,
392 width=args.sky_width,
393 nb_birds=args.sky_nb_birds,
394 nb_iterations=args.sky_nb_iterations,
395 speed=args.sky_speed,
396 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
398 nb_threads=args.nb_threads,
400 back_accuracy = False
401 elif args.problem == "grids":
402 problem = grids.Grids(
403 max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
405 nb_threads=args.nb_threads,
411 problem.save_some_examples(args.result_dir)
413 quiz_machine = quiz_machine.QuizMachine(
415 nb_train_samples=args.nb_train_samples,
416 nb_test_samples=args.nb_test_samples,
417 back_accuracy=back_accuracy,
418 batch_size=args.physical_batch_size,
419 result_dir=args.result_dir,
424 ######################################################################
426 log_string(f"device {device}")
428 vocabulary_size = quiz_machine.vocabulary_size()
430 log_string(f"vocabulary_size {vocabulary_size}")
432 ######################################################################
436 for k in range(args.nb_gpts):
437 log_string(f"creating model {k} and its w_quizzes")
439 vocabulary_size=vocabulary_size,
440 dim_model=args.dim_model,
441 dim_keys=args.dim_keys,
442 dim_hidden=args.dim_hidden,
443 nb_heads=args.nb_heads,
444 nb_blocks=args.nb_blocks,
446 dropout=args.dropout,
449 model.main_test_accuracy = 0.0
452 model.train_w_quizzes = quiz_machine.generate_token_sequences(
453 args.nb_train_samples
455 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
456 model.test_w_quizzes = quiz_machine.generate_token_sequences(
459 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
464 nb_parameters = sum(p.numel() for p in models[0].parameters())
465 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
467 ######################################################################
469 # Compute the entropy of the training tokens
472 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
473 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
476 token_probas = token_count / token_count.sum()
477 entropy = -torch.xlogy(token_probas, token_probas).sum()
478 train_set_perplexity = math.exp(entropy)
480 ######################################################################
481 # A bit of paranoia never hurts
483 if args.max_percents_of_test_in_train >= 0:
485 def subsets_as_tuples(batches, cs):
487 for batch in batches:
489 s.add(tuple([v.item() for v in x]))
495 nb_test, nb_in_train = 0, 0
496 for test_subset in subsets_as_tuples(
497 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
500 for train_subset in subsets_as_tuples(
501 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
503 in_train.update(test_subset.intersection(train_subset))
504 nb_in_train += len(in_train)
505 nb_test += len(test_subset)
508 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
512 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
513 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
515 ######################################################################
517 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
518 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
521 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}"
524 ######################################################################
527 args.accuracy_to_make_c_quizzes = 0.0
529 nb_new_c_quizzes_for_train = 100
530 nb_new_c_quizzes_for_test = 10
532 def standard_validity(logproba):
533 l = logproba.sort(dim=-1).values
534 return l[:, 0] < math.log(0.99)
537 ######################################################################
539 for n_epoch in range(args.nb_epochs):
540 log_string(f"--- epoch {n_epoch} ----------------------------------------")
542 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
543 log_string(f"current_test_accuracies {cta}")
545 ##################################################
546 # Select, improve, and eval the worst models
548 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
550 weakest_models = ranked_models[: args.nb_gpus]
554 for gpu_id, model in enumerate(weakest_models):
556 f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
559 process = mp.Process(
560 target=one_epoch, args=(model, quiz_machine, f"cuda:{gpu_id}")
563 processes.append(process)
565 for process in processes:
568 for process in processes:
571 ##################################################
572 # Renew the train sets
575 f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
578 for model in weakest_models:
579 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
581 ##################################################
582 # If all the models are good enough, generate new quizzes and
583 # re-compute the test errors
585 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
589 nb_for_train=nb_new_c_quizzes_for_train,
590 nb_for_test=nb_new_c_quizzes_for_test,
593 ######################################################################