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=False)
84 parser.add_argument("--problem", type=str, default="sky")
86 parser.add_argument("--nb_gpts", type=int, default=5)
88 parser.add_argument("--nb_models_for_generation", type=int, default=1)
90 parser.add_argument("--generation_mode", type=str, default="groupthink")
92 parser.add_argument("--min_to_validate", type=int, default=4)
94 parser.add_argument("--max_to_validate", type=int, default=4)
96 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
98 parser.add_argument("--dirty_debug", action="store_true", default=False)
100 parser.add_argument("--sky_height", type=int, default=6)
102 parser.add_argument("--sky_width", type=int, default=8)
104 parser.add_argument("--sky_nb_birds", type=int, default=3)
106 parser.add_argument("--sky_nb_iterations", type=int, default=2)
108 parser.add_argument("--sky_speed", type=int, default=3)
110 ######################################################################
112 args = parser.parse_args()
114 if args.result_dir is None:
115 args.result_dir = f"results_culture"
117 ######################################################################
120 args.accuracy_to_make_c_quizzes = 0.0
121 nb_new_c_quizzes_for_train = 100
122 nb_new_c_quizzes_for_test = 10
124 ######################################################################
129 "nb_train_samples": 100000,
130 "nb_test_samples": 10000,
133 for k, v in default_args.items():
134 if getattr(args, k) is None:
137 ######################################################################
139 default_model_args = {
177 if args.model in default_model_args:
178 for k, v in default_model_args[args.model].items():
179 if getattr(args, k) is None:
182 raise ValueError(f"Unknown model {args.model}")
184 ######################################################################
187 os.mkdir(args.result_dir)
188 except FileExistsError:
189 print(f"result directory {args.result_dir} already exists")
192 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
195 # torch.backends.cudnn.deterministic = True
196 # torch.backends.cudnn.benchmark = False
197 # torch.use_deterministic_algorithms(True)
198 torch.manual_seed(args.seed)
199 if torch.cuda.is_available():
200 torch.cuda.manual_seed_all(args.seed)
202 ######################################################################
206 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
208 if log_file is not None:
209 log_file.write(t + s + "\n")
216 log_string(f"argv {' '.join(sys.argv)}")
219 log_string(f"args.{n} {getattr(args, n)}")
222 ######################################################################
225 args.nb_train_samples = 2500
226 args.nb_test_samples = 100
228 if args.physical_batch_size is None:
229 args.physical_batch_size = args.batch_size
231 assert args.batch_size % args.physical_batch_size == 0
233 assert args.nb_train_samples % args.batch_size == 0
234 assert args.nb_test_samples % args.batch_size == 0
236 if args.problem == "sky":
238 height=args.sky_height,
239 width=args.sky_width,
240 nb_birds=args.sky_nb_birds,
241 nb_iterations=args.sky_nb_iterations,
242 speed=args.sky_speed,
244 elif args.problem == "wireworld":
245 problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
249 quizz_machine = quizz_machine.QuizzMachine(
251 nb_train_samples=args.nb_train_samples,
252 nb_test_samples=args.nb_test_samples,
253 batch_size=args.physical_batch_size,
254 result_dir=args.result_dir,
259 ######################################################################
261 log_string(f"device {device}")
263 vocabulary_size = quizz_machine.vocabulary_size()
265 log_string(f"vocabulary_size {vocabulary_size}")
267 ######################################################################
269 # Compute the entropy of the training tokens
272 for input in quizz_machine.batches(split="train", desc="train-entropy"):
273 token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
276 token_probas = token_count / token_count.sum()
277 entropy = -torch.xlogy(token_probas, token_probas).sum()
278 train_set_perplexity = math.exp(entropy)
280 ######################################################################
281 # A bit of paranoia never hurts
283 if args.max_percents_of_test_in_train >= 0:
285 def subsets_as_tuples(batches, cs):
287 for batch in batches:
289 s.add(tuple([v.item() for v in x]))
295 nb_test, nb_in_train = 0, 0
296 for test_subset in subsets_as_tuples(
297 quizz_machine.batches(split="test", desc="test-check"), 25000
300 for train_subset in subsets_as_tuples(
301 quizz_machine.batches(split="train", desc="train-check"), 25000
303 in_train.update(test_subset.intersection(train_subset))
304 nb_in_train += len(in_train)
305 nb_test += len(test_subset)
308 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
312 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
313 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
315 ##############################
318 def one_epoch(model, quizz_machine):
319 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
323 nb_train_samples, acc_train_loss = 0, 0.0
325 for input in quizz_machine.batches(split="train"):
326 input = input.to(device)
328 if nb_train_samples % args.batch_size == 0:
329 optimizer.zero_grad()
331 output = model(mygpt.BracketedSequence(input)).x
332 loss = F.cross_entropy(output.transpose(1, 2), input)
333 acc_train_loss += loss.item() * input.size(0)
335 nb_train_samples += input.size(0)
339 if nb_train_samples % args.batch_size == 0:
342 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
344 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
347 ######################################################################
350 def run_tests(model, quizz_machine, deterministic_synthesis):
351 with torch.autograd.no_grad():
354 nb_test_samples, acc_test_loss = 0, 0.0
355 nb_samples_accumulated = 0
357 for input in quizz_machine.batches(split="test"):
358 input = input.to(device)
360 bs = model(mygpt.BracketedSequence(input))
363 loss = F.cross_entropy(output.transpose(1, 2), input)
365 acc_test_loss += loss.item() * input.size(0)
367 nb_test_samples += input.size(0)
369 main_test_accuracy = quizz_machine.produce_results(
372 result_dir=args.result_dir,
373 deterministic_synthesis=deterministic_synthesis,
376 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
378 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
380 model.main_test_accuracy = main_test_accuracy
383 ######################################################################
386 def create_c_quizzes(
391 min_ave_seq_logproba=None,
393 # We will store the generated quizzes for each number of
395 recorded = dict([(n, []) for n in range(len(models) + 1)])
398 sum_logits, sum_nb_c_quizzes = 0, 0
401 return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
406 sum([x.size(0) for x in recorded[n]])
407 for n in range(args.min_to_validate, args.max_to_validate + 1)
411 nb_to_create = nb_for_train + nb_for_test
413 while nb_validated() < nb_to_create:
418 ) = quizz_machine.gang_create_c_quizzes(
420 nb_models_for_generation=args.nb_models_for_generation,
422 mode=args.generation_mode,
423 reverse_cleanup=args.reverse_cleanup,
424 min_ave_seq_logproba=min_ave_seq_logproba,
426 result_dir=args.result_dir,
429 sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
430 sum_nb_c_quizzes += new_c_quizzes.size(0)
433 nb_correct = torch.randint(
434 len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
437 for n in range(nb_correct.max() + 1):
438 recorded[n].append(new_c_quizzes[nb_correct == n].clone())
440 nv = [recorded[n][-1].size(0) for n in recorded.keys()]
442 log_string(f"keep c_quizzes kept {nv} total {nb_validated()} / {nb_to_create}")
444 # concatenate and shuffle
445 for n in recorded.keys():
446 if len(recorded[n]) > 0:
447 q = torch.cat(recorded[n], dim=0)
448 q = q[torch.randperm(q.size(0), device=q.device)]
453 new_c_quizzes = torch.cat(
454 [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
458 new_c_quizzes = new_c_quizzes[
459 torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
460 : nb_for_train + nb_for_test
464 quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
465 quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
467 for n in recorded.keys():
470 if n >= args.min_to_validate and n <= args.max_to_validate
473 quizz_machine.problem.save_quizzes(
476 f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
479 return sum_logits / sum_nb_c_quizzes
482 ######################################################################
486 for k in range(args.nb_gpts):
488 vocabulary_size=vocabulary_size,
489 dim_model=args.dim_model,
490 dim_keys=args.dim_keys,
491 dim_hidden=args.dim_hidden,
492 nb_heads=args.nb_heads,
493 nb_blocks=args.nb_blocks,
495 dropout=args.dropout,
498 model.main_test_accuracy = 0.0
504 nb_parameters = sum(p.numel() for p in models[0].parameters())
505 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
507 ######################################################################
509 min_ave_seq_logproba = None
511 for n_epoch in range(args.nb_epochs):
512 log_string(f"--- epoch {n_epoch} ----------------------------------------")
514 a = [(model.id, float(model.main_test_accuracy)) for model in models]
515 a.sort(key=lambda p: p[0])
516 s = " ".join([f"{p[1]*100:.02f}%" for p in a])
517 log_string(f"current accuracies {s}")
519 # select the model with lowest accuracy
520 models.sort(key=lambda model: model.main_test_accuracy)
524 f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
528 one_epoch(model, quizz_machine)
530 quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
533 f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
537 run_tests(model, quizz_machine, deterministic_synthesis=False)
540 f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
543 if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
544 ave_seq_logproba = create_c_quizzes(
547 nb_for_train=nb_new_c_quizzes_for_train,
548 nb_for_test=nb_new_c_quizzes_for_test,
549 min_ave_seq_logproba=min_ave_seq_logproba,
552 # We keep the first average logits as a reference
553 # if min_ave_seq_logproba is None:
554 # min_ave_seq_logproba = ave_seq_logproba
557 # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
562 run_tests(model, quizz_machine, deterministic_synthesis=False)
565 ######################################################################