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("--resume", action="store_true", default=False)
37 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
39 ########################################
41 parser.add_argument("--nb_epochs", type=int, default=10000)
43 parser.add_argument("--batch_size", type=int, default=None)
45 parser.add_argument("--physical_batch_size", type=int, default=None)
47 parser.add_argument("--nb_train_samples", type=int, default=None)
49 parser.add_argument("--nb_test_samples", type=int, default=None)
51 parser.add_argument("--learning_rate", type=float, default=5e-4)
53 ########################################
55 parser.add_argument("--model", type=str, default=None)
57 parser.add_argument("--dim_model", type=int, default=None)
59 parser.add_argument("--dim_keys", type=int, default=None)
61 parser.add_argument("--dim_hidden", type=int, default=None)
63 parser.add_argument("--nb_heads", type=int, default=None)
65 parser.add_argument("--nb_blocks", type=int, default=None)
67 parser.add_argument("--dropout", type=float, default=0.1)
69 ########################################
71 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
73 parser.add_argument("--problem", type=str, default="grids")
75 parser.add_argument("--nb_threads", type=int, default=1)
77 parser.add_argument("--gpus", type=str, default="all")
79 parser.add_argument("--nb_gpts", type=int, default=5)
81 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
83 parser.add_argument("--proba_understands", type=float, default=0.99)
85 parser.add_argument("--proba_not_understands", type=float, default=0.5)
87 parser.add_argument("--generation_temperature", type=float, default=2.0)
89 parser.add_argument("--dirty_debug", action="store_true", default=False)
91 ######################################################################
93 grids_tasks = ", ".join(
94 [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
101 help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
104 ######################################################################
106 parser.add_argument("--sky_height", type=int, default=6)
108 parser.add_argument("--sky_width", type=int, default=8)
110 parser.add_argument("--sky_nb_birds", type=int, default=3)
112 parser.add_argument("--sky_nb_iterations", type=int, default=2)
114 parser.add_argument("--sky_speed", type=int, default=3)
116 ######################################################################
118 args = parser.parse_args()
120 if args.result_dir is None:
121 args.result_dir = f"results_culture"
123 ######################################################################
128 "nb_train_samples": 100000,
129 "nb_test_samples": 10000,
132 for k, v in default_args.items():
133 if getattr(args, k) is None:
136 ######################################################################
138 default_model_args = {
176 if args.model in default_model_args:
177 for k, v in default_model_args[args.model].items():
178 if getattr(args, k) is None:
181 raise ValueError(f"Unknown model {args.model}")
183 ######################################################################
186 assert os.path.isdir(args.result_dir)
190 os.mkdir(args.result_dir)
191 except FileExistsError:
192 print(f"result directory {args.result_dir} already exists")
195 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
198 # torch.backends.cudnn.deterministic = True
199 # torch.backends.cudnn.benchmark = False
200 # torch.use_deterministic_algorithms(True)
201 torch.manual_seed(args.seed)
202 if torch.cuda.is_available():
203 torch.cuda.manual_seed_all(args.seed)
205 ######################################################################
209 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
211 if log_file is not None:
212 log_file.write(t + s + "\n")
219 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
221 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py")
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 model.to(local_device).train()
337 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
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_quizzes_and_logprobas(recorded, criteria):
378 validated_quizzes, validated_logprobas = [], []
379 for q, lp in recorded:
380 validated_indices = criteria(lp)
381 validated_quizzes.append(q[validated_indices])
382 validated_logprobas.append(lp[validated_indices])
384 if len(validated_quizzes) > 0:
385 return torch.cat(validated_quizzes, dim=0), torch.cat(
386 validated_logprobas, dim=0
392 ######################################################################
395 def create_c_quizzes(models, quiz_machine, nb_for_train=1000, nb_for_test=100):
396 nb_to_create = nb_for_train + nb_for_test
398 recorded_quizzes_logprobas = []
402 while nb_validated < nb_to_create:
403 model_for_generation = models[torch.randint(len(models), (1,))]
405 c_quizzes = quiz_machine.generate_quizzes(
407 model_for_generation=model_for_generation,
408 temperature=args.generation_temperature,
411 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
413 if c_quizzes.size(0) > 0:
414 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
415 recorded_quizzes_logprobas.append((c_quizzes, logproba))
417 validated_quizzes, validated_logprobas = valid_quizzes_and_logprobas(
418 recorded_quizzes_logprobas, standard_validity
421 if validated_quizzes is not None:
422 nb_validated = validated_quizzes.size(0)
425 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
428 # store the new c_quizzes which have been validated
430 quiz_machine.reverse_random_half_in_place(validated_quizzes)
431 quiz_machine.store_c_quizzes(validated_quizzes[:nb_for_train], for_train=True)
432 quiz_machine.store_c_quizzes(
433 validated_quizzes[nb_for_train:nb_to_create], for_train=False
436 ######################################################################
437 # save the log probas
439 file_name = os.path.join(
440 args.result_dir, f"culture_c_quiz_all_{n_epoch:04d}_logp.dat"
443 with open(file_name, "w") as logp_file:
444 for _, ll in recorded_quizzes_logprobas:
446 s = " ".join([str(x.item()) for x in l])
447 logp_file.write(s + "\n")
449 ######################################################################
450 # save images with their logprobas
452 vq = validated_quizzes[:72]
453 vl = validated_logprobas[:72]
456 prefix = f"culture_c_quiz_{n_epoch:04d}"
458 file_name = os.path.join(args.result_dir, prefix + "_logp.dat")
459 with open(file_name, "w") as logp_file:
461 s = " ".join([str(x.item()) for x in l])
462 logp_file.write(s + "\n")
464 quiz_machine.save_quiz_illustrations(args.result_dir, prefix, vq)
467 ######################################################################
471 for k in range(args.nb_gpts):
472 log_string(f"creating model {k} and its w_quizzes")
474 vocabulary_size=vocabulary_size,
475 dim_model=args.dim_model,
476 dim_keys=args.dim_keys,
477 dim_hidden=args.dim_hidden,
478 nb_heads=args.nb_heads,
479 nb_blocks=args.nb_blocks,
481 dropout=args.dropout,
484 model.main_test_accuracy = 0.0
487 model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
488 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
489 model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
490 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
494 ######################################################################
499 filename = f"gpt_{model.id:03d}.pth"
502 d = torch.load(os.path.join(args.result_dir, filename))
503 model.load_state_dict(d[0])
504 model.main_test_accuracy = d[1]
505 log_string(f"successfully loaded {filename}")
506 except FileNotFoundError:
507 log_string(f"cannot find {filename}")
511 filename = "c_quizzes.pth"
512 quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
513 log_string(f"successfully loaded {filename}")
514 except FileNotFoundError:
515 log_string(f"cannot find {filename}")
519 log_string(f"error when loading {filename}.")
522 ######################################################################
524 nb_parameters = sum(p.numel() for p in models[0].parameters())
525 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
527 ######################################################################
529 # Compute the entropy of the training tokens
532 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
533 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
536 token_probas = token_count / token_count.sum()
537 entropy = -torch.xlogy(token_probas, token_probas).sum()
538 train_set_perplexity = math.exp(entropy)
540 ######################################################################
541 # A bit of paranoia never hurts
543 if args.max_percents_of_test_in_train >= 0:
545 def subsets_as_tuples(batches, cs):
547 for batch in batches:
549 s.add(tuple([v.item() for v in x]))
555 nb_test, nb_in_train = 0, 0
556 for test_subset in subsets_as_tuples(
557 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
560 for train_subset in subsets_as_tuples(
561 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
563 in_train.update(test_subset.intersection(train_subset))
564 nb_in_train += len(in_train)
565 nb_test += len(test_subset)
568 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
572 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
573 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
575 ######################################################################
577 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
578 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
581 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}"
584 ######################################################################
587 args.accuracy_to_make_c_quizzes = 0.0
589 nb_new_c_quizzes_for_train = 100
590 nb_new_c_quizzes_for_test = 10
592 def standard_validity(logproba):
593 l = logproba.sort(dim=-1).values
594 return l[:, 0] < math.log(0.5)
597 ######################################################################
599 for n_epoch in range(args.nb_epochs):
600 log_string(f"--- epoch {n_epoch} ----------------------------------------")
602 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
603 log_string(f"current_test_accuracies {cta}")
605 ##################################################
606 # Select, improve, and eval the worst model
608 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
610 weakest_models = ranked_models[: len(gpus)]
614 for gpu, model in zip(gpus, weakest_models):
615 log_string(f"training model {model.id}")
617 t = threading.Thread(
618 target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
628 # Save the models to disk
630 for model in weakest_models:
631 filename = f"gpt_{model.id:03d}.pth"
633 (model.state_dict(), model.main_test_accuracy),
634 os.path.join(args.result_dir, filename),
636 log_string(f"wrote {filename}")
638 # Renew the training samples
640 for model in weakest_models:
641 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
643 ##################################################
644 # If all the models are good enough, generate new quizzes and
645 # re-compute the test errors
647 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
651 nb_for_train=nb_new_c_quizzes_for_train,
652 nb_for_test=nb_new_c_quizzes_for_test,
655 filename = "c_quizzes.pth"
656 quiz_machine.save_c_quizzes(os.path.join(args.result_dir, filename))
657 log_string(f"wrote {filename}")
659 ######################################################################