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
16 import sky, grids, quiz_machine
18 # world quizzes vs. culture quizzes
20 ######################################################################
22 if torch.cuda.is_available():
23 device = torch.device("cuda")
24 torch.backends.cuda.matmul.allow_tf32 = True
26 device = torch.device("cpu")
28 ######################################################################
30 parser = argparse.ArgumentParser(
31 description="An implementation of GPT with cache.",
32 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
35 parser.add_argument("--log_filename", type=str, default="train.log")
37 parser.add_argument("--result_dir", type=str, default=None)
39 parser.add_argument("--seed", type=int, default=0)
41 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
43 ########################################
45 parser.add_argument("--nb_epochs", type=int, default=10000)
47 parser.add_argument("--batch_size", type=int, default=None)
49 parser.add_argument("--physical_batch_size", type=int, default=None)
51 parser.add_argument("--nb_train_samples", type=int, default=None)
53 parser.add_argument("--nb_test_samples", type=int, default=None)
55 parser.add_argument("--learning_rate", type=float, default=5e-4)
57 ########################################
59 parser.add_argument("--model", type=str, default=None)
61 parser.add_argument("--dim_model", type=int, default=None)
63 parser.add_argument("--dim_keys", type=int, default=None)
65 parser.add_argument("--dim_hidden", type=int, default=None)
67 parser.add_argument("--nb_heads", type=int, default=None)
69 parser.add_argument("--nb_blocks", type=int, default=None)
71 parser.add_argument("--dropout", type=float, default=0.1)
73 ########################################
75 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
77 parser.add_argument("--problem", type=str, default="grids")
79 parser.add_argument("--nb_gpts", type=int, default=5)
81 parser.add_argument("--min_to_validate", type=int, default=None)
83 parser.add_argument("--max_to_validate", type=int, default=None)
85 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
87 parser.add_argument("--generation_temperature", type=float, default=2.0)
89 parser.add_argument("--deterministic_validation", action="store_true", default=False)
91 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
93 parser.add_argument("--dirty_debug", action="store_true", default=False)
95 ######################################################################
97 parser.add_argument("--sky_height", type=int, default=6)
99 parser.add_argument("--sky_width", type=int, default=8)
101 parser.add_argument("--sky_nb_birds", type=int, default=3)
103 parser.add_argument("--sky_nb_iterations", type=int, default=2)
105 parser.add_argument("--sky_speed", type=int, default=3)
107 ######################################################################
109 args = parser.parse_args()
111 if args.min_to_validate is None:
112 args.min_to_validate = args.nb_gpts - 1
114 if args.max_to_validate is None:
115 args.max_to_validate = args.nb_gpts - 1
117 if args.result_dir is None:
118 args.result_dir = f"results_culture"
120 ######################################################################
125 "nb_train_samples": 100000,
126 "nb_test_samples": 10000,
129 for k, v in default_args.items():
130 if getattr(args, k) is None:
133 ######################################################################
135 default_model_args = {
173 if args.model in default_model_args:
174 for k, v in default_model_args[args.model].items():
175 if getattr(args, k) is None:
178 raise ValueError(f"Unknown model {args.model}")
180 ######################################################################
183 os.mkdir(args.result_dir)
184 except FileExistsError:
185 print(f"result directory {args.result_dir} already exists")
188 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
191 # torch.backends.cudnn.deterministic = True
192 # torch.backends.cudnn.benchmark = False
193 # torch.use_deterministic_algorithms(True)
194 torch.manual_seed(args.seed)
195 if torch.cuda.is_available():
196 torch.cuda.manual_seed_all(args.seed)
198 ######################################################################
202 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
204 if log_file is not None:
205 log_file.write(t + s + "\n")
212 log_string(f"argv {' '.join(sys.argv)}")
215 log_string(f"args.{n} {getattr(args, n)}")
218 ######################################################################
221 args.nb_train_samples = 2500
222 args.nb_test_samples = 100
224 if args.physical_batch_size is None:
225 args.physical_batch_size = args.batch_size
227 assert args.batch_size % args.physical_batch_size == 0
229 assert args.nb_train_samples % args.batch_size == 0
230 assert args.nb_test_samples % args.batch_size == 0
232 if args.problem == "sky":
234 height=args.sky_height,
235 width=args.sky_width,
236 nb_birds=args.sky_nb_birds,
237 nb_iterations=args.sky_nb_iterations,
238 speed=args.sky_speed,
240 back_accuracy = False
241 elif args.problem == "grids":
242 problem = grids.Grids(device=device)
247 quiz_machine = quiz_machine.QuizMachine(
249 nb_train_samples=args.nb_train_samples,
250 nb_test_samples=args.nb_test_samples,
251 back_accuracy=back_accuracy,
252 batch_size=args.physical_batch_size,
253 result_dir=args.result_dir,
258 ######################################################################
260 log_string(f"device {device}")
262 vocabulary_size = quiz_machine.vocabulary_size()
264 log_string(f"vocabulary_size {vocabulary_size}")
266 ######################################################################
268 # Compute the entropy of the training tokens
271 for input in quiz_machine.batches(split="train", desc="train-entropy"):
272 token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
275 token_probas = token_count / token_count.sum()
276 entropy = -torch.xlogy(token_probas, token_probas).sum()
277 train_set_perplexity = math.exp(entropy)
279 ######################################################################
280 # A bit of paranoia never hurts
282 if args.max_percents_of_test_in_train >= 0:
284 def subsets_as_tuples(batches, cs):
286 for batch in batches:
288 s.add(tuple([v.item() for v in x]))
294 nb_test, nb_in_train = 0, 0
295 for test_subset in subsets_as_tuples(
296 quiz_machine.batches(split="test", desc="test-check"), 25000
299 for train_subset in subsets_as_tuples(
300 quiz_machine.batches(split="train", desc="train-check"), 25000
302 in_train.update(test_subset.intersection(train_subset))
303 nb_in_train += len(in_train)
304 nb_test += len(test_subset)
307 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
311 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
312 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
314 ##############################
317 def one_epoch(model, quiz_machine):
318 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
322 nb_train_samples, acc_train_loss = 0, 0.0
324 for input in quiz_machine.batches(split="train"):
325 input = input.to(device)
327 if nb_train_samples % args.batch_size == 0:
328 optimizer.zero_grad()
330 output = model(mygpt.BracketedSequence(input)).x
331 loss = F.cross_entropy(output.transpose(1, 2), input)
332 acc_train_loss += loss.item() * input.size(0)
334 nb_train_samples += input.size(0)
338 if nb_train_samples % args.batch_size == 0:
341 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
343 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
346 ######################################################################
349 def run_tests(model, quiz_machine, deterministic_synthesis):
350 with torch.autograd.no_grad():
353 nb_test_samples, acc_test_loss = 0, 0.0
354 nb_samples_accumulated = 0
356 for input in quiz_machine.batches(split="test"):
357 input = input.to(device)
359 bs = model(mygpt.BracketedSequence(input))
362 loss = F.cross_entropy(output.transpose(1, 2), input)
364 acc_test_loss += loss.item() * input.size(0)
366 nb_test_samples += input.size(0)
368 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
370 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
372 model.main_test_accuracy = quiz_machine.produce_results(
375 result_dir=args.result_dir,
376 deterministic_synthesis=deterministic_synthesis,
380 ######################################################################
383 def valid_c_quizzes(recorded, criteria):
384 result = [q[criteria(c)] for q, c in recorded]
385 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
388 ######################################################################
391 def create_c_quizzes(
397 quizzes_and_nb_correct_records = []
399 nb_to_create = nb_for_train + nb_for_test
401 # ------------------------------------------------------------
403 standard_validity = lambda nb_correct: torch.logical_and(
404 nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
407 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
409 with open(file_name, "w") as logp_file:
411 valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
414 # Select a model at random to generate the new quizzes
416 model_for_generation = models[torch.randint(len(models), (1,))]
418 c_quizzes = quiz_machine.generate_quizzes(
420 model_for_generation=model_for_generation,
421 temperature=args.generation_temperature,
424 c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
426 if c_quizzes.size(0) > 0:
427 nb_correct, seq_logproba = quiz_machine.compute_correctness(
430 bidirectional_validation=args.bidirectional_validation,
431 deterministic_validation=args.deterministic_validation,
434 for n, l in zip(nb_correct, seq_logproba):
435 s = " ".join([str(x.item()) for x in l])
436 logp_file.write(f"{n} {s}\n")
439 nb_correct = torch.randint(
440 len(models) + 1, nb_correct.size(), device=c_quizzes.device
443 quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
445 nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
446 nv = " ".join([str(x.item()) for x in nv])
448 nb_validated = valid_c_quizzes(
449 quizzes_and_nb_correct_records, standard_validity
453 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
456 # store the new c_quizzes which have been validated
458 new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
460 quiz_machine.reverse_random_half_in_place(new_c_quizzes)
462 quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
463 quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
465 # save a bunch of images to investigate what quizzes with a
466 # certain nb of correct predictions look like
468 for n in range(len(models) + 1):
471 if n >= args.min_to_validate and n <= args.max_to_validate
476 quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
479 quiz_machine.reverse_random_half_in_place(q)
482 quiz_machine.save_quizzes(
483 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
487 ######################################################################
491 for k in range(args.nb_gpts):
493 vocabulary_size=vocabulary_size,
494 dim_model=args.dim_model,
495 dim_keys=args.dim_keys,
496 dim_hidden=args.dim_hidden,
497 nb_heads=args.nb_heads,
498 nb_blocks=args.nb_blocks,
500 dropout=args.dropout,
503 model.main_test_accuracy = 0.0
509 nb_parameters = sum(p.numel() for p in models[0].parameters())
510 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
512 ######################################################################
514 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
515 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
518 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}"
521 ######################################################################
524 args.accuracy_to_make_c_quizzes = 0.0
526 nb_new_c_quizzes_for_train = 100
527 nb_new_c_quizzes_for_test = 10
529 ######################################################################
531 for n_epoch in range(args.nb_epochs):
532 log_string(f"--- epoch {n_epoch} ----------------------------------------")
534 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
535 log_string(f"current_test_accuracies {cta}")
537 ##################################################
538 # Select, improve, and eval the worst model
540 weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
543 f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
546 one_epoch(weakest_model, quiz_machine)
549 f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
552 run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
555 f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
558 ##################################################
559 # Replace a fraction of the w_quizzes with fresh ones
561 quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
563 ##################################################
564 # If all the models are good enough, generate new quizzes and
565 # re-compute the test errors
567 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
571 nb_for_train=nb_new_c_quizzes_for_train,
572 nb_for_test=nb_new_c_quizzes_for_test,
576 run_tests(model, quiz_machine, deterministic_synthesis=False)
579 ######################################################################