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 ######################################################################
25 parser = argparse.ArgumentParser(
26 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
29 parser.add_argument("--log_filename", type=str, default="train.log")
31 parser.add_argument("--result_dir", type=str, default=None)
33 parser.add_argument("--seed", type=int, default=0)
35 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
37 ########################################
39 parser.add_argument("--nb_epochs", type=int, default=10000)
41 parser.add_argument("--batch_size", type=int, default=None)
43 parser.add_argument("--physical_batch_size", type=int, default=None)
45 parser.add_argument("--nb_train_samples", type=int, default=None)
47 parser.add_argument("--nb_test_samples", type=int, default=None)
49 parser.add_argument("--learning_rate", type=float, default=5e-4)
51 ########################################
53 parser.add_argument("--model", type=str, default=None)
55 parser.add_argument("--dim_model", type=int, default=None)
57 parser.add_argument("--dim_keys", type=int, default=None)
59 parser.add_argument("--dim_hidden", type=int, default=None)
61 parser.add_argument("--nb_heads", type=int, default=None)
63 parser.add_argument("--nb_blocks", type=int, default=None)
65 parser.add_argument("--dropout", type=float, default=0.1)
67 ########################################
69 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
71 parser.add_argument("--problem", type=str, default="grids")
73 parser.add_argument("--nb_threads", type=int, default=1)
75 parser.add_argument("--gpus", type=str, default="all")
77 parser.add_argument("--nb_gpts", type=int, default=5)
79 parser.add_argument("--min_to_validate", type=int, default=None)
81 parser.add_argument("--max_to_validate", type=int, default=None)
83 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
85 parser.add_argument("--proba_understands", type=float, default=0.99)
87 parser.add_argument("--proba_not_understands", type=float, default=0.5)
89 parser.add_argument("--generation_temperature", type=float, default=2.0)
91 parser.add_argument("--dirty_debug", action="store_true", default=False)
93 ######################################################################
95 grids_tasks = ", ".join(
96 [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
103 help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
106 ######################################################################
108 parser.add_argument("--sky_height", type=int, default=6)
110 parser.add_argument("--sky_width", type=int, default=8)
112 parser.add_argument("--sky_nb_birds", type=int, default=3)
114 parser.add_argument("--sky_nb_iterations", type=int, default=2)
116 parser.add_argument("--sky_speed", type=int, default=3)
118 ######################################################################
120 args = parser.parse_args()
122 if args.min_to_validate is None:
123 args.min_to_validate = args.nb_gpts - 1
125 if args.max_to_validate is None:
126 args.max_to_validate = args.nb_gpts - 1
128 if args.result_dir is None:
129 args.result_dir = f"results_culture"
131 ######################################################################
136 "nb_train_samples": 100000,
137 "nb_test_samples": 10000,
140 for k, v in default_args.items():
141 if getattr(args, k) is None:
144 ######################################################################
146 default_model_args = {
184 if args.model in default_model_args:
185 for k, v in default_model_args[args.model].items():
186 if getattr(args, k) is None:
189 raise ValueError(f"Unknown model {args.model}")
191 ######################################################################
194 os.mkdir(args.result_dir)
195 except FileExistsError:
196 print(f"result directory {args.result_dir} already exists")
199 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
202 # torch.backends.cudnn.deterministic = True
203 # torch.backends.cudnn.benchmark = False
204 # torch.use_deterministic_algorithms(True)
205 torch.manual_seed(args.seed)
206 if torch.cuda.is_available():
207 torch.cuda.manual_seed_all(args.seed)
209 ######################################################################
213 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
215 if log_file is not None:
216 log_file.write(t + s + "\n")
223 log_string(f"argv {' '.join(sys.argv)}")
226 log_string(f"args.{n} {getattr(args, n)}")
229 ######################################################################
231 if args.gpus == "all":
232 gpus_idx = range(torch.cuda.device_count())
234 gpus_idx = [int(k) for k in args.gpus.split(",")]
236 gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
238 if torch.cuda.is_available():
239 main_device = gpus[0]
241 assert len(gpus) == 0
242 main_device = torch.device("cpu")
245 args.nb_train_samples = 2500
246 args.nb_test_samples = 100
248 if args.physical_batch_size is None:
249 args.physical_batch_size = args.batch_size
251 assert args.batch_size % args.physical_batch_size == 0
253 assert args.nb_train_samples % args.batch_size == 0
254 assert args.nb_test_samples % args.batch_size == 0
256 if args.problem == "sky":
258 height=args.sky_height,
259 width=args.sky_width,
260 nb_birds=args.sky_nb_birds,
261 nb_iterations=args.sky_nb_iterations,
262 speed=args.sky_speed,
263 max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
265 nb_threads=args.nb_threads,
267 back_accuracy = False
268 elif args.problem == "grids":
269 problem = grids.Grids(
270 max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
272 nb_threads=args.nb_threads,
273 tasks=args.grids_tasks,
279 problem.save_some_examples(args.result_dir)
281 quiz_machine = quiz_machine.QuizMachine(
283 nb_train_samples=args.nb_train_samples,
284 nb_test_samples=args.nb_test_samples,
285 back_accuracy=back_accuracy,
286 batch_size=args.physical_batch_size,
287 result_dir=args.result_dir,
292 ######################################################################
294 log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
296 vocabulary_size = quiz_machine.vocabulary_size()
298 log_string(f"vocabulary_size {vocabulary_size}")
300 ######################################################################
303 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
304 with torch.autograd.no_grad():
305 model.eval().to(local_device)
307 nb_test_samples, acc_test_loss = 0, 0.0
308 nb_samples_accumulated = 0
310 for input in quiz_machine.batches(model, split="test"):
311 input = input.to(local_device)
313 bs = model(mygpt.BracketedSequence(input))
316 loss = F.cross_entropy(output.transpose(1, 2), input)
318 acc_test_loss += loss.item() * input.size(0)
320 nb_test_samples += input.size(0)
322 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
324 log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
326 model.main_test_accuracy = quiz_machine.produce_results(
329 result_dir=args.result_dir,
330 deterministic_synthesis=deterministic_synthesis,
334 def one_epoch(model, quiz_machine, local_device=main_device):
335 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
337 model.to(local_device).train()
339 nb_train_samples, acc_train_loss = 0, 0.0
341 for input in quiz_machine.batches(model, split="train"):
342 input = input.to(local_device)
344 if nb_train_samples % args.batch_size == 0:
345 optimizer.zero_grad()
347 output = model(mygpt.BracketedSequence(input)).x
348 loss = F.cross_entropy(output.transpose(1, 2), input)
349 acc_train_loss += loss.item() * input.size(0)
351 nb_train_samples += input.size(0)
355 if nb_train_samples % args.batch_size == 0:
358 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
360 log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
362 run_tests(model, quiz_machine, deterministic_synthesis=False)
364 model.to(main_device)
367 ######################################################################
370 def standard_validity(logproba):
371 l = logproba.sort(dim=-1).values
372 return (l[:, 0] < math.log(args.proba_not_understands)) & (
373 l[:, 1] > math.log(args.proba_understands)
377 def valid_c_quizzes(recorded, criteria):
378 result = [q[criteria(lp)] for q, lp in recorded]
379 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
382 ######################################################################
385 def create_c_quizzes(
391 quizzes_and_logproba_records = []
393 nb_to_create = nb_for_train + nb_for_test
395 # ------------------------------------------------------------
397 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
399 with open(file_name, "w") as logp_file:
401 valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
404 # Select a model at random to generate the new quizzes
406 model_for_generation = models[torch.randint(len(models), (1,))]
408 c_quizzes = quiz_machine.generate_quizzes(
410 model_for_generation=model_for_generation,
411 temperature=args.generation_temperature,
414 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
416 if c_quizzes.size(0) > 0:
417 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
419 s = " ".join([str(x.item()) for x in l])
420 logp_file.write(s + "\n")
421 quizzes_and_logproba_records.append((c_quizzes, logproba))
423 nb_validated = valid_c_quizzes(
424 quizzes_and_logproba_records, standard_validity
428 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
431 # store the new c_quizzes which have been validated
433 new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
435 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
437 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
438 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
440 # save a bunch of images to investigate what quizzes with a
441 # certain nb of correct predictions look like
443 q = new_c_quizzes[:72]
446 quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
449 ######################################################################
453 for k in range(args.nb_gpts):
454 log_string(f"creating model {k} and its w_quizzes")
456 vocabulary_size=vocabulary_size,
457 dim_model=args.dim_model,
458 dim_keys=args.dim_keys,
459 dim_hidden=args.dim_hidden,
460 nb_heads=args.nb_heads,
461 nb_blocks=args.nb_blocks,
463 dropout=args.dropout,
466 model.main_test_accuracy = 0.0
469 model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
470 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
471 model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
472 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
477 nb_parameters = sum(p.numel() for p in models[0].parameters())
478 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
480 ######################################################################
482 # Compute the entropy of the training tokens
485 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
486 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
489 token_probas = token_count / token_count.sum()
490 entropy = -torch.xlogy(token_probas, token_probas).sum()
491 train_set_perplexity = math.exp(entropy)
493 ######################################################################
494 # A bit of paranoia never hurts
496 if args.max_percents_of_test_in_train >= 0:
498 def subsets_as_tuples(batches, cs):
500 for batch in batches:
502 s.add(tuple([v.item() for v in x]))
508 nb_test, nb_in_train = 0, 0
509 for test_subset in subsets_as_tuples(
510 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
513 for train_subset in subsets_as_tuples(
514 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
516 in_train.update(test_subset.intersection(train_subset))
517 nb_in_train += len(in_train)
518 nb_test += len(test_subset)
521 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
525 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
526 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
528 ######################################################################
530 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
531 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
534 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}"
537 ######################################################################
540 args.accuracy_to_make_c_quizzes = 0.0
542 nb_new_c_quizzes_for_train = 100
543 nb_new_c_quizzes_for_test = 10
545 def standard_validity(logproba):
546 l = logproba.sort(dim=-1).values
547 return l[:, 0] < math.log(0.5)
550 ######################################################################
555 filename = f"gpt_{model.id:03d}.pth"
558 model.load_state_dict(torch.load(os.path.join(args.result_dir, filename)))
559 log_string(f"model {model.id} successfully loaded from checkpoint.")
560 nb_loaded_models += 1
562 except FileNotFoundError:
563 log_string(f"starting model {model.id} from scratch.")
566 log_string(f"error when loading {filename}.")
569 assert nb_loaded_models == 0 or nb_loaded_models == len(models)
571 ######################################################################
573 for n_epoch in range(args.nb_epochs):
574 log_string(f"--- epoch {n_epoch} ----------------------------------------")
576 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
577 log_string(f"current_test_accuracies {cta}")
579 ##################################################
580 # Select, improve, and eval the worst model
582 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
584 weakest_models = ranked_models[: len(gpus)]
588 for gpu, model in zip(gpus, weakest_models):
589 log_string(f"training model {model.id}")
591 t = threading.Thread(
592 target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
602 for model in weakest_models:
603 filename = f"gpt_{model.id:03d}.pth"
604 torch.save(model.state_dict(), os.path.join(args.result_dir, filename))
606 ##################################################
607 # Replace a fraction of the w_quizzes with fresh ones
610 f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
613 # Renew entirely the train set
615 for model in weakest_models:
616 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
618 ##################################################
619 # If all the models are good enough, generate new quizzes and
620 # re-compute the test errors
622 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
626 nb_for_train=nb_new_c_quizzes_for_train,
627 nb_for_test=nb_new_c_quizzes_for_test,
630 ######################################################################