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 log_string(f"argv {' '.join(sys.argv)}")
222 log_string(f"args.{n} {getattr(args, n)}")
225 ######################################################################
227 if args.gpus == "all":
228 gpus_idx = range(torch.cuda.device_count())
230 gpus_idx = [int(k) for k in args.gpus.split(",")]
232 gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
234 if torch.cuda.is_available():
235 main_device = gpus[0]
237 assert len(gpus) == 0
238 main_device = torch.device("cpu")
241 args.nb_train_samples = 2500
242 args.nb_test_samples = 100
244 if args.physical_batch_size is None:
245 args.physical_batch_size = args.batch_size
247 assert args.batch_size % args.physical_batch_size == 0
249 assert args.nb_train_samples % args.batch_size == 0
250 assert args.nb_test_samples % args.batch_size == 0
252 if args.problem == "sky":
254 height=args.sky_height,
255 width=args.sky_width,
256 nb_birds=args.sky_nb_birds,
257 nb_iterations=args.sky_nb_iterations,
258 speed=args.sky_speed,
259 max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
261 nb_threads=args.nb_threads,
263 back_accuracy = False
264 elif args.problem == "grids":
265 problem = grids.Grids(
266 max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
268 nb_threads=args.nb_threads,
269 tasks=args.grids_tasks,
275 problem.save_some_examples(args.result_dir)
277 quiz_machine = quiz_machine.QuizMachine(
279 nb_train_samples=args.nb_train_samples,
280 nb_test_samples=args.nb_test_samples,
281 back_accuracy=back_accuracy,
282 batch_size=args.physical_batch_size,
283 result_dir=args.result_dir,
288 ######################################################################
290 log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
292 vocabulary_size = quiz_machine.vocabulary_size()
294 log_string(f"vocabulary_size {vocabulary_size}")
296 ######################################################################
299 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
300 with torch.autograd.no_grad():
301 model.eval().to(local_device)
303 nb_test_samples, acc_test_loss = 0, 0.0
304 nb_samples_accumulated = 0
306 for input in quiz_machine.batches(model, split="test"):
307 input = input.to(local_device)
309 bs = model(mygpt.BracketedSequence(input))
312 loss = F.cross_entropy(output.transpose(1, 2), input)
314 acc_test_loss += loss.item() * input.size(0)
316 nb_test_samples += input.size(0)
318 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
320 log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
322 model.main_test_accuracy = quiz_machine.produce_results(
325 result_dir=args.result_dir,
326 deterministic_synthesis=deterministic_synthesis,
330 def one_epoch(model, quiz_machine, local_device=main_device):
331 model.to(local_device).train()
333 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
335 nb_train_samples, acc_train_loss = 0, 0.0
337 for input in quiz_machine.batches(model, split="train"):
338 input = input.to(local_device)
340 if nb_train_samples % args.batch_size == 0:
341 optimizer.zero_grad()
343 output = model(mygpt.BracketedSequence(input)).x
344 loss = F.cross_entropy(output.transpose(1, 2), input)
345 acc_train_loss += loss.item() * input.size(0)
347 nb_train_samples += input.size(0)
351 if nb_train_samples % args.batch_size == 0:
354 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
356 log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
358 run_tests(model, quiz_machine, deterministic_synthesis=False)
360 model.to(main_device)
363 ######################################################################
366 def standard_validity(logproba):
367 l = logproba.sort(dim=-1).values
368 return (l[:, 0] < math.log(args.proba_not_understands)) & (
369 l[:, 1] > math.log(args.proba_understands)
373 def valid_c_quizzes(recorded, criteria):
374 result = [q[criteria(lp)] for q, lp in recorded]
375 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
378 ######################################################################
381 def create_c_quizzes(
387 quizzes_and_logproba_records = []
389 nb_to_create = nb_for_train + nb_for_test
391 # ------------------------------------------------------------
393 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
395 with open(file_name, "w") as logp_file:
397 valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
400 # Select a model at random to generate the new quizzes
402 model_for_generation = models[torch.randint(len(models), (1,))]
404 c_quizzes = quiz_machine.generate_quizzes(
406 model_for_generation=model_for_generation,
407 temperature=args.generation_temperature,
410 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
412 if c_quizzes.size(0) > 0:
413 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
415 s = " ".join([str(x.item()) for x in l])
416 logp_file.write(s + "\n")
417 quizzes_and_logproba_records.append((c_quizzes, logproba))
419 nb_validated = valid_c_quizzes(
420 quizzes_and_logproba_records, standard_validity
424 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
427 # store the new c_quizzes which have been validated
429 new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
431 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
433 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
434 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
438 q = new_c_quizzes[:72]
441 quiz_machine.save_quiz_illustrations(
442 args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q
446 ######################################################################
450 for k in range(args.nb_gpts):
451 log_string(f"creating model {k} and its w_quizzes")
453 vocabulary_size=vocabulary_size,
454 dim_model=args.dim_model,
455 dim_keys=args.dim_keys,
456 dim_hidden=args.dim_hidden,
457 nb_heads=args.nb_heads,
458 nb_blocks=args.nb_blocks,
460 dropout=args.dropout,
463 model.main_test_accuracy = 0.0
466 model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
467 quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
468 model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
469 quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
473 ######################################################################
478 filename = f"gpt_{model.id:03d}.pth"
481 model.load_state_dict(
482 torch.load(os.path.join(args.result_dir, filename))
484 log_string(f"successfully loaded {filename}")
485 except FileNotFoundError:
486 log_string(f"cannot find {filename}")
490 filename = "c_quizzes.pth"
491 quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
492 log_string(f"successfully loaded {filename}")
493 except FileNotFoundError:
494 log_string(f"cannot find {filename}")
498 log_string(f"error when loading {filename}.")
501 ######################################################################
503 nb_parameters = sum(p.numel() for p in models[0].parameters())
504 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
506 ######################################################################
508 # Compute the entropy of the training tokens
511 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
512 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
515 token_probas = token_count / token_count.sum()
516 entropy = -torch.xlogy(token_probas, token_probas).sum()
517 train_set_perplexity = math.exp(entropy)
519 ######################################################################
520 # A bit of paranoia never hurts
522 if args.max_percents_of_test_in_train >= 0:
524 def subsets_as_tuples(batches, cs):
526 for batch in batches:
528 s.add(tuple([v.item() for v in x]))
534 nb_test, nb_in_train = 0, 0
535 for test_subset in subsets_as_tuples(
536 quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
539 for train_subset in subsets_as_tuples(
540 quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
542 in_train.update(test_subset.intersection(train_subset))
543 nb_in_train += len(in_train)
544 nb_test += len(test_subset)
547 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
551 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
552 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
554 ######################################################################
556 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
557 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
560 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}"
563 ######################################################################
566 args.accuracy_to_make_c_quizzes = 0.0
568 nb_new_c_quizzes_for_train = 100
569 nb_new_c_quizzes_for_test = 10
571 def standard_validity(logproba):
572 l = logproba.sort(dim=-1).values
573 return l[:, 0] < math.log(0.5)
576 ######################################################################
578 for n_epoch in range(args.nb_epochs):
579 log_string(f"--- epoch {n_epoch} ----------------------------------------")
581 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
582 log_string(f"current_test_accuracies {cta}")
584 ##################################################
585 # Select, improve, and eval the worst model
587 ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
589 weakest_models = ranked_models[: len(gpus)]
593 for gpu, model in zip(gpus, weakest_models):
594 log_string(f"training model {model.id}")
596 t = threading.Thread(
597 target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
607 for model in weakest_models:
608 filename = f"gpt_{model.id:03d}.pth"
609 torch.save(model.state_dict(), os.path.join(args.result_dir, filename))
610 log_string(f"wrote {filename}")
612 ##################################################
613 # Replace a fraction of the w_quizzes with fresh ones
616 f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
619 # Renew entirely the train set
621 for model in weakest_models:
622 quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
624 ##################################################
625 # If all the models are good enough, generate new quizzes and
626 # re-compute the test errors
628 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
632 nb_for_train=nb_new_c_quizzes_for_train,
633 nb_for_test=nb_new_c_quizzes_for_test,
636 quiz_machine.save_c_quizzes(os.path.join(args.result_dir, "c_quizzes.pth"))
638 ######################################################################