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, wireworld, quizz_machine
18 # world quizzes vs. culture quizzes
20 ######################################################################
22 nb_new_c_quizzes_for_train = 1000
23 nb_new_c_quizzes_for_test = 100
25 ######################################################################
27 if torch.cuda.is_available():
28 device = torch.device("cuda")
29 torch.backends.cuda.matmul.allow_tf32 = True
31 device = torch.device("cpu")
33 ######################################################################
35 parser = argparse.ArgumentParser(
36 description="An implementation of GPT with cache.",
37 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
40 parser.add_argument("--log_filename", type=str, default="train.log")
42 parser.add_argument("--result_dir", type=str, default=None)
44 parser.add_argument("--seed", type=int, default=0)
46 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
48 ########################################
50 parser.add_argument("--nb_epochs", type=int, default=10000)
52 parser.add_argument("--batch_size", type=int, default=None)
54 parser.add_argument("--physical_batch_size", type=int, default=None)
56 parser.add_argument("--nb_train_samples", type=int, default=None)
58 parser.add_argument("--nb_test_samples", type=int, default=None)
60 parser.add_argument("--learning_rate", type=float, default=1e-3)
62 ########################################
64 parser.add_argument("--model", type=str, default=None)
66 parser.add_argument("--dim_model", type=int, default=None)
68 parser.add_argument("--dim_keys", type=int, default=None)
70 parser.add_argument("--dim_hidden", type=int, default=None)
72 parser.add_argument("--nb_heads", type=int, default=None)
74 parser.add_argument("--nb_blocks", type=int, default=None)
76 parser.add_argument("--dropout", type=float, default=0.1)
78 ########################################
80 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
82 parser.add_argument("--reverse_cleanup", action="store_true", default=True)
84 parser.add_argument("--validation_forward_only", action="store_true", default=False)
86 parser.add_argument("--problem", type=str, default="sky")
88 parser.add_argument("--nb_gpts", type=int, default=5)
90 parser.add_argument("--min_to_validate", type=int, default=4)
92 parser.add_argument("--max_to_validate", type=int, default=4)
94 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
96 parser.add_argument("--dirty_debug", action="store_true", default=False)
98 parser.add_argument("--sky_height", type=int, default=6)
100 parser.add_argument("--sky_width", type=int, default=8)
102 parser.add_argument("--sky_nb_birds", type=int, default=3)
104 parser.add_argument("--sky_nb_iterations", type=int, default=2)
106 parser.add_argument("--sky_speed", type=int, default=3)
108 ######################################################################
110 args = parser.parse_args()
112 if args.result_dir is None:
113 args.result_dir = f"results_culture"
115 ######################################################################
118 args.accuracy_to_make_c_quizzes = 0.0
119 nb_new_c_quizzes_for_train = 100
120 nb_new_c_quizzes_for_test = 10
122 ######################################################################
127 "nb_train_samples": 100000,
128 "nb_test_samples": 10000,
131 for k, v in default_args.items():
132 if getattr(args, k) is None:
135 ######################################################################
137 default_model_args = {
175 if args.model in default_model_args:
176 for k, v in default_model_args[args.model].items():
177 if getattr(args, k) is None:
180 raise ValueError(f"Unknown model {args.model}")
182 ######################################################################
185 os.mkdir(args.result_dir)
186 except FileExistsError:
187 print(f"result directory {args.result_dir} already exists")
190 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
193 # torch.backends.cudnn.deterministic = True
194 # torch.backends.cudnn.benchmark = False
195 # torch.use_deterministic_algorithms(True)
196 torch.manual_seed(args.seed)
197 if torch.cuda.is_available():
198 torch.cuda.manual_seed_all(args.seed)
200 ######################################################################
204 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
206 if log_file is not None:
207 log_file.write(t + s + "\n")
214 log_string(f"argv {' '.join(sys.argv)}")
217 log_string(f"args.{n} {getattr(args, n)}")
220 ######################################################################
223 args.nb_train_samples = 2500
224 args.nb_test_samples = 100
226 if args.physical_batch_size is None:
227 args.physical_batch_size = args.batch_size
229 assert args.batch_size % args.physical_batch_size == 0
231 assert args.nb_train_samples % args.batch_size == 0
232 assert args.nb_test_samples % args.batch_size == 0
234 if args.problem == "sky":
236 height=args.sky_height,
237 width=args.sky_width,
238 nb_birds=args.sky_nb_birds,
239 nb_iterations=args.sky_nb_iterations,
240 speed=args.sky_speed,
242 elif args.problem == "wireworld":
243 problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
247 quizz_machine = quizz_machine.QuizzMachine(
249 nb_train_samples=args.nb_train_samples,
250 nb_test_samples=args.nb_test_samples,
251 batch_size=args.physical_batch_size,
252 result_dir=args.result_dir,
257 ######################################################################
259 log_string(f"device {device}")
261 vocabulary_size = quizz_machine.vocabulary_size()
263 log_string(f"vocabulary_size {vocabulary_size}")
265 ######################################################################
267 # Compute the entropy of the training tokens
270 for input in quizz_machine.batches(split="train", desc="train-entropy"):
271 token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
274 token_probas = token_count / token_count.sum()
275 entropy = -torch.xlogy(token_probas, token_probas).sum()
276 train_set_perplexity = math.exp(entropy)
278 ######################################################################
279 # A bit of paranoia never hurts
281 if args.max_percents_of_test_in_train >= 0:
283 def subsets_as_tuples(batches, cs):
285 for batch in batches:
287 s.add(tuple([v.item() for v in x]))
293 nb_test, nb_in_train = 0, 0
294 for test_subset in subsets_as_tuples(
295 quizz_machine.batches(split="test", desc="test-check"), 25000
298 for train_subset in subsets_as_tuples(
299 quizz_machine.batches(split="train", desc="train-check"), 25000
301 in_train.update(test_subset.intersection(train_subset))
302 nb_in_train += len(in_train)
303 nb_test += len(test_subset)
306 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
310 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
311 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
313 ##############################
316 def one_epoch(model, quizz_machine):
317 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
321 nb_train_samples, acc_train_loss = 0, 0.0
323 for input in quizz_machine.batches(split="train"):
324 input = input.to(device)
326 if nb_train_samples % args.batch_size == 0:
327 optimizer.zero_grad()
329 output = model(mygpt.BracketedSequence(input)).x
330 loss = F.cross_entropy(output.transpose(1, 2), input)
331 acc_train_loss += loss.item() * input.size(0)
333 nb_train_samples += input.size(0)
337 if nb_train_samples % args.batch_size == 0:
340 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
342 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
345 ######################################################################
348 def run_tests(model, quizz_machine, deterministic_synthesis):
349 with torch.autograd.no_grad():
352 nb_test_samples, acc_test_loss = 0, 0.0
353 nb_samples_accumulated = 0
355 for input in quizz_machine.batches(split="test"):
356 input = input.to(device)
358 bs = model(mygpt.BracketedSequence(input))
361 loss = F.cross_entropy(output.transpose(1, 2), input)
363 acc_test_loss += loss.item() * input.size(0)
365 nb_test_samples += input.size(0)
367 model.main_test_accuracy = quizz_machine.produce_results(
370 result_dir=args.result_dir,
371 deterministic_synthesis=deterministic_synthesis,
374 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
376 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
379 ######################################################################
382 def valid_c_quizzes(recorded, criteria):
383 result = [q[criteria(c)] for q, c in recorded]
384 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
387 ######################################################################
390 def create_c_quizzes(
398 nb_to_create = nb_for_train + nb_for_test
400 # ------------------------------------------------------------
402 standard_validity = lambda nb_correct: torch.logical_and(
403 nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
406 while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create:
407 model_for_generation = models[torch.randint(len(models), (1,))]
409 c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes(
411 model_for_generation=model_for_generation,
412 reverse_cleanup=args.reverse_cleanup,
415 nb_correct = quizz_machine.compute_correctness(
416 c_quizzes, models, both_directions=not args.validation_forward_only
420 nb_correct = torch.randint(
421 len(models) + 1, nb_correct.size(), device=c_quizzes.device
424 recorded.append((c_quizzes, nb_correct))
426 nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
427 nv = " ".join([str(x.item()) for x in nv])
429 nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
432 f"keep c_quizzes kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
435 # store the new c_quizzes which have been validated
437 new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
439 quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
440 quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
442 # save a bunch of images to investigate what quizzes with a
443 # certain nb of correct predictions look like
445 for n in range(len(models) + 1):
448 if n >= args.min_to_validate and n <= args.max_to_validate
452 q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
455 quizz_machine.problem.save_quizzes(
458 f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
462 ######################################################################
466 for k in range(args.nb_gpts):
468 vocabulary_size=vocabulary_size,
469 dim_model=args.dim_model,
470 dim_keys=args.dim_keys,
471 dim_hidden=args.dim_hidden,
472 nb_heads=args.nb_heads,
473 nb_blocks=args.nb_blocks,
475 dropout=args.dropout,
478 model.main_test_accuracy = 0.0
484 nb_parameters = sum(p.numel() for p in models[0].parameters())
485 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
487 ######################################################################
489 for n_epoch in range(args.nb_epochs):
490 log_string(f"--- epoch {n_epoch} ----------------------------------------")
492 weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
495 f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
499 one_epoch(weakest_model, quizz_machine)
502 f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
506 run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
509 f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
512 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
513 log_string(f"current_test_accuracies {cta}")
515 # replace a fraction of the w_quizzes with fresh ones
516 quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
518 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
522 nb_for_train=nb_new_c_quizzes_for_train,
523 nb_for_test=nb_new_c_quizzes_for_test,
528 run_tests(model, quizz_machine, deterministic_synthesis=False)
531 ######################################################################