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, reasoning, 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("--problem", type=str, default="sky")
84 parser.add_argument("--nb_gpts", type=int, default=5)
86 parser.add_argument("--min_to_validate", type=int, default=None)
88 parser.add_argument("--max_to_validate", type=int, default=None)
90 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
92 parser.add_argument("--generation_temperature", type=float, default=2.0)
94 parser.add_argument("--deterministic_validation", action="store_true", default=False)
96 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
98 parser.add_argument("--dirty_debug", action="store_true", default=False)
100 ######################################################################
102 parser.add_argument("--sky_height", type=int, default=6)
104 parser.add_argument("--sky_width", type=int, default=8)
106 parser.add_argument("--sky_nb_birds", type=int, default=3)
108 parser.add_argument("--sky_nb_iterations", type=int, default=2)
110 parser.add_argument("--sky_speed", type=int, default=3)
112 ######################################################################
114 args = parser.parse_args()
116 if args.min_to_validate is None:
117 args.min_to_validate = args.nb_gpts - 1
119 if args.max_to_validate is None:
120 args.max_to_validate = args.nb_gpts - 1
122 if args.result_dir is None:
123 args.result_dir = f"results_culture"
125 ######################################################################
128 args.accuracy_to_make_c_quizzes = 0.0
129 nb_new_c_quizzes_for_train = 100
130 nb_new_c_quizzes_for_test = 10
132 ######################################################################
137 "nb_train_samples": 100000,
138 "nb_test_samples": 10000,
141 for k, v in default_args.items():
142 if getattr(args, k) is None:
145 ######################################################################
147 default_model_args = {
185 if args.model in default_model_args:
186 for k, v in default_model_args[args.model].items():
187 if getattr(args, k) is None:
190 raise ValueError(f"Unknown model {args.model}")
192 ######################################################################
195 os.mkdir(args.result_dir)
196 except FileExistsError:
197 print(f"result directory {args.result_dir} already exists")
200 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
203 # torch.backends.cudnn.deterministic = True
204 # torch.backends.cudnn.benchmark = False
205 # torch.use_deterministic_algorithms(True)
206 torch.manual_seed(args.seed)
207 if torch.cuda.is_available():
208 torch.cuda.manual_seed_all(args.seed)
210 ######################################################################
214 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
216 if log_file is not None:
217 log_file.write(t + s + "\n")
224 log_string(f"argv {' '.join(sys.argv)}")
227 log_string(f"args.{n} {getattr(args, n)}")
230 ######################################################################
233 args.nb_train_samples = 2500
234 args.nb_test_samples = 100
236 if args.physical_batch_size is None:
237 args.physical_batch_size = args.batch_size
239 assert args.batch_size % args.physical_batch_size == 0
241 assert args.nb_train_samples % args.batch_size == 0
242 assert args.nb_test_samples % args.batch_size == 0
244 if args.problem == "sky":
246 height=args.sky_height,
247 width=args.sky_width,
248 nb_birds=args.sky_nb_birds,
249 nb_iterations=args.sky_nb_iterations,
250 speed=args.sky_speed,
252 back_accuracy = False
253 elif args.problem == "reasoning":
254 problem = reasoning.Reasoning(device=device)
259 quizz_machine = quizz_machine.QuizzMachine(
261 nb_train_samples=args.nb_train_samples,
262 nb_test_samples=args.nb_test_samples,
263 back_accuracy=back_accuracy,
264 batch_size=args.physical_batch_size,
265 result_dir=args.result_dir,
270 ######################################################################
272 log_string(f"device {device}")
274 vocabulary_size = quizz_machine.vocabulary_size()
276 log_string(f"vocabulary_size {vocabulary_size}")
278 ######################################################################
280 # Compute the entropy of the training tokens
283 for input in quizz_machine.batches(split="train", desc="train-entropy"):
284 token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
287 token_probas = token_count / token_count.sum()
288 entropy = -torch.xlogy(token_probas, token_probas).sum()
289 train_set_perplexity = math.exp(entropy)
291 ######################################################################
292 # A bit of paranoia never hurts
294 if args.max_percents_of_test_in_train >= 0:
296 def subsets_as_tuples(batches, cs):
298 for batch in batches:
300 s.add(tuple([v.item() for v in x]))
306 nb_test, nb_in_train = 0, 0
307 for test_subset in subsets_as_tuples(
308 quizz_machine.batches(split="test", desc="test-check"), 25000
311 for train_subset in subsets_as_tuples(
312 quizz_machine.batches(split="train", desc="train-check"), 25000
314 in_train.update(test_subset.intersection(train_subset))
315 nb_in_train += len(in_train)
316 nb_test += len(test_subset)
319 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
323 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
324 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
326 ##############################
329 def one_epoch(model, quizz_machine):
330 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
334 nb_train_samples, acc_train_loss = 0, 0.0
336 for input in quizz_machine.batches(split="train"):
337 input = input.to(device)
339 if nb_train_samples % args.batch_size == 0:
340 optimizer.zero_grad()
342 output = model(mygpt.BracketedSequence(input)).x
343 loss = F.cross_entropy(output.transpose(1, 2), input)
344 acc_train_loss += loss.item() * input.size(0)
346 nb_train_samples += input.size(0)
350 if nb_train_samples % args.batch_size == 0:
353 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
355 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
358 ######################################################################
361 def run_tests(model, quizz_machine, deterministic_synthesis):
362 with torch.autograd.no_grad():
365 nb_test_samples, acc_test_loss = 0, 0.0
366 nb_samples_accumulated = 0
368 for input in quizz_machine.batches(split="test"):
369 input = input.to(device)
371 bs = model(mygpt.BracketedSequence(input))
374 loss = F.cross_entropy(output.transpose(1, 2), input)
376 acc_test_loss += loss.item() * input.size(0)
378 nb_test_samples += input.size(0)
380 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
382 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
384 model.main_test_accuracy = quizz_machine.produce_results(
387 result_dir=args.result_dir,
388 deterministic_synthesis=deterministic_synthesis,
392 ######################################################################
395 def valid_c_quizzes(recorded, criteria):
396 result = [q[criteria(c)] for q, c in recorded]
397 return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
400 ######################################################################
403 def create_c_quizzes(
411 nb_to_create = nb_for_train + nb_for_test
413 # ------------------------------------------------------------
415 standard_validity = lambda nb_correct: torch.logical_and(
416 nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
419 file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
420 with open(file_name, "w") as logp_file:
421 while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create:
422 # Select a model at random to generate the new quizzes
424 model_for_generation = models[torch.randint(len(models), (1,))]
426 c_quizzes = quizz_machine.generate_quizzes(
428 model_for_generation=model_for_generation,
429 temperature=args.generation_temperature,
432 nb_correct, seq_logproba = quizz_machine.compute_correctness(
435 bidirectional_validation=args.bidirectional_validation,
436 deterministic_validation=args.deterministic_validation,
439 for n, l in zip(nb_correct, seq_logproba):
440 s = " ".join([str(x.item()) for x in l])
441 logp_file.write(f"{n} {s}\n")
444 nb_correct = torch.randint(
445 len(models) + 1, nb_correct.size(), device=c_quizzes.device
448 recorded.append((c_quizzes, nb_correct))
450 nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
451 nv = " ".join([str(x.item()) for x in nv])
453 nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
456 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
459 # store the new c_quizzes which have been validated
461 new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
463 quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
464 quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
466 # save a bunch of images to investigate what quizzes with a
467 # certain nb of correct predictions look like
469 for n in range(len(models) + 1):
472 if n >= args.min_to_validate and n <= args.max_to_validate
476 q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
479 quizz_machine.save_quizzes(
480 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
484 ######################################################################
488 for k in range(args.nb_gpts):
490 vocabulary_size=vocabulary_size,
491 dim_model=args.dim_model,
492 dim_keys=args.dim_keys,
493 dim_hidden=args.dim_hidden,
494 nb_heads=args.nb_heads,
495 nb_blocks=args.nb_blocks,
497 dropout=args.dropout,
500 model.main_test_accuracy = 0.0
506 nb_parameters = sum(p.numel() for p in models[0].parameters())
507 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
509 ######################################################################
511 for n_epoch in range(args.nb_epochs):
512 log_string(f"--- epoch {n_epoch} ----------------------------------------")
514 cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
515 log_string(f"current_test_accuracies {cta}")
517 # Select, improve, and eval the worst model
519 weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
522 f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
525 one_epoch(weakest_model, quizz_machine)
528 f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
531 run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
534 f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
537 # Replace a fraction of the w_quizzes with fresh ones
539 quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
541 # If all the models are good enough, generate new quizzes and
542 # re-compute the test errors
544 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
548 nb_for_train=nb_new_c_quizzes_for_train,
549 nb_for_test=nb_new_c_quizzes_for_test,
553 run_tests(model, quizz_machine, deterministic_synthesis=False)
556 ######################################################################