main.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math, sys, argparse, time, tqdm, os, datetime, warnings
   9
  10 import torch, torchvision
  11 from torch import nn
  12 from torch.nn import functional as F
  13
  14 import ffutils
  15
  16 import mygpt
  17 import sky, grids, quiz_machine
  18
  19 # world quizzes vs. culture quizzes
  20
  21 ######################################################################
  22
  23 if torch.cuda.is_available():
  24     device = torch.device("cuda")
  25     torch.backends.cuda.matmul.allow_tf32 = True
  26 else:
  27     device = torch.device("cpu")
  28
  29 ######################################################################
  30
  31 parser = argparse.ArgumentParser(
  32     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  33 )
  34
  35 parser.add_argument("--log_filename", type=str, default="train.log")
  36
  37 parser.add_argument("--result_dir", type=str, default=None)
  38
  39 parser.add_argument("--seed", type=int, default=0)
  40
  41 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  42
  43 ########################################
  44
  45 parser.add_argument("--nb_epochs", type=int, default=10000)
  46
  47 parser.add_argument("--batch_size", type=int, default=None)
  48
  49 parser.add_argument("--physical_batch_size", type=int, default=None)
  50
  51 parser.add_argument("--nb_train_samples", type=int, default=None)
  52
  53 parser.add_argument("--nb_test_samples", type=int, default=None)
  54
  55 parser.add_argument("--learning_rate", type=float, default=5e-4)
  56
  57 ########################################
  58
  59 parser.add_argument("--model", type=str, default=None)
  60
  61 parser.add_argument("--dim_model", type=int, default=None)
  62
  63 parser.add_argument("--dim_keys", type=int, default=None)
  64
  65 parser.add_argument("--dim_hidden", type=int, default=None)
  66
  67 parser.add_argument("--nb_heads", type=int, default=None)
  68
  69 parser.add_argument("--nb_blocks", type=int, default=None)
  70
  71 parser.add_argument("--dropout", type=float, default=0.1)
  72
  73 ########################################
  74
  75 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  76
  77 parser.add_argument("--problem", type=str, default="grids")
  78
  79 parser.add_argument("--nb_threads", type=int, default=-1)
  80
  81 parser.add_argument("--nb_gpts", type=int, default=5)
  82
  83 parser.add_argument("--min_to_validate", type=int, default=None)
  84
  85 parser.add_argument("--max_to_validate", type=int, default=None)
  86
  87 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  88
  89 parser.add_argument("--generation_temperature", type=float, default=2.0)
  90
  91 parser.add_argument("--deterministic_validation", action="store_true", default=False)
  92
  93 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
  94
  95 parser.add_argument("--dirty_debug", action="store_true", default=False)
  96
  97 ######################################################################
  98
  99 parser.add_argument("--sky_height", type=int, default=6)
 100
 101 parser.add_argument("--sky_width", type=int, default=8)
 102
 103 parser.add_argument("--sky_nb_birds", type=int, default=3)
 104
 105 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 106
 107 parser.add_argument("--sky_speed", type=int, default=3)
 108
 109 ######################################################################
 110
 111 args = parser.parse_args()
 112
 113 if args.min_to_validate is None:
 114     args.min_to_validate = args.nb_gpts - 1
 115
 116 if args.max_to_validate is None:
 117     args.max_to_validate = args.nb_gpts - 1
 118
 119 if args.result_dir is None:
 120     args.result_dir = f"results_culture"
 121
 122 ######################################################################
 123
 124 default_args = {
 125     "model": "37M",
 126     "batch_size": 25,
 127     "nb_train_samples": 100000,
 128     "nb_test_samples": 10000,
 129 }
 130
 131 for k, v in default_args.items():
 132     if getattr(args, k) is None:
 133         setattr(args, k, v)
 134
 135 ######################################################################
 136
 137 default_model_args = {
 138     "17K": {
 139         "dim_model": 32,
 140         "dim_keys": 32,
 141         "dim_hidden": 32,
 142         "nb_heads": 2,
 143         "nb_blocks": 2,
 144     },
 145     "4M": {
 146         "dim_model": 256,
 147         "dim_keys": 32,
 148         "dim_hidden": 1024,
 149         "nb_heads": 4,
 150         "nb_blocks": 6,
 151     },
 152     "37M": {
 153         "dim_model": 512,
 154         "dim_keys": 64,
 155         "dim_hidden": 2048,
 156         "nb_heads": 8,
 157         "nb_blocks": 12,
 158     },
 159     "122M": {
 160         "dim_model": 768,
 161         "dim_keys": 64,
 162         "dim_hidden": 2048,
 163         "nb_heads": 8,
 164         "nb_blocks": 24,
 165     },
 166     "352M": {
 167         "dim_model": 1024,
 168         "dim_keys": 64,
 169         "dim_hidden": 2048,
 170         "nb_heads": 8,
 171         "nb_blocks": 48,
 172     },
 173 }
 174
 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:
 178             setattr(args, k, v)
 179 else:
 180     raise ValueError(f"Unknown model {args.model}")
 181
 182 ######################################################################
 183
 184 try:
 185     os.mkdir(args.result_dir)
 186 except FileExistsError:
 187     print(f"result directory {args.result_dir} already exists")
 188     exit(1)
 189
 190 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 191
 192 if args.seed >= 0:
 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)
 199
 200 ######################################################################
 201
 202
 203 def log_string(s):
 204     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 205
 206     if log_file is not None:
 207         log_file.write(t + s + "\n")
 208         log_file.flush()
 209
 210     print(t + s)
 211     sys.stdout.flush()
 212
 213
 214 log_string(f"argv {' '.join(sys.argv)}")
 215
 216 for n in vars(args):
 217     log_string(f"args.{n} {getattr(args, n)}")
 218
 219
 220 ######################################################################
 221
 222 if args.dirty_debug:
 223     args.nb_train_samples = 2500
 224     args.nb_test_samples = 100
 225
 226 if args.physical_batch_size is None:
 227     args.physical_batch_size = args.batch_size
 228 else:
 229     assert args.batch_size % args.physical_batch_size == 0
 230
 231 assert args.nb_train_samples % args.batch_size == 0
 232 assert args.nb_test_samples % args.batch_size == 0
 233
 234 if args.problem == "sky":
 235     problem = sky.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,
 241         max_nb_cached_chunks=args.nb_train_samples // 100,
 242         chunk_size=100,
 243         nb_threads=args.nb_threads,
 244     )
 245     back_accuracy = False
 246 elif args.problem == "grids":
 247     problem = grids.Grids(
 248         max_nb_cached_chunks=args.nb_train_samples // 100,
 249         chunk_size=100,
 250         nb_threads=args.nb_threads,
 251     )
 252     back_accuracy = True
 253 else:
 254     raise ValueError
 255
 256 quiz_machine = quiz_machine.QuizMachine(
 257     problem=problem,
 258     nb_train_samples=args.nb_train_samples,
 259     nb_test_samples=args.nb_test_samples,
 260     back_accuracy=back_accuracy,
 261     batch_size=args.physical_batch_size,
 262     result_dir=args.result_dir,
 263     logger=log_string,
 264     device=device,
 265 )
 266
 267 ######################################################################
 268
 269 log_string(f"device {device}")
 270
 271 vocabulary_size = quiz_machine.vocabulary_size()
 272
 273 log_string(f"vocabulary_size {vocabulary_size}")
 274
 275 ######################################################################
 276 ##############################
 277
 278
 279 def one_epoch(model, quiz_machine):
 280     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 281
 282     model.train()
 283
 284     nb_train_samples, acc_train_loss = 0, 0.0
 285
 286     for input in quiz_machine.batches(model, split="train"):
 287         input = input.to(device)
 288
 289         if nb_train_samples % args.batch_size == 0:
 290             optimizer.zero_grad()
 291
 292         output = model(mygpt.BracketedSequence(input)).x
 293         loss = F.cross_entropy(output.transpose(1, 2), input)
 294         acc_train_loss += loss.item() * input.size(0)
 295
 296         nb_train_samples += input.size(0)
 297
 298         loss.backward()
 299
 300         if nb_train_samples % args.batch_size == 0:
 301             optimizer.step()
 302
 303     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 304
 305     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 306
 307
 308 ######################################################################
 309
 310
 311 def run_tests(model, quiz_machine, deterministic_synthesis):
 312     with torch.autograd.no_grad():
 313         model.eval()
 314
 315         nb_test_samples, acc_test_loss = 0, 0.0
 316         nb_samples_accumulated = 0
 317
 318         for input in quiz_machine.batches(model, split="test"):
 319             input = input.to(device)
 320
 321             bs = model(mygpt.BracketedSequence(input))
 322             output = bs.x
 323
 324             loss = F.cross_entropy(output.transpose(1, 2), input)
 325
 326             acc_test_loss += loss.item() * input.size(0)
 327
 328             nb_test_samples += input.size(0)
 329
 330         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 331
 332         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 333
 334         model.main_test_accuracy = quiz_machine.produce_results(
 335             n_epoch=n_epoch,
 336             model=model,
 337             result_dir=args.result_dir,
 338             deterministic_synthesis=deterministic_synthesis,
 339         )
 340
 341
 342 ######################################################################
 343
 344
 345 def standard_validity(logproba):
 346     l = logproba.sort(dim=-1).values
 347     return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
 348     # warnings.warn("TEST!!!", RuntimeWarning)
 349     # print(l.exp())
 350     # return (l[:, 0] < math.log(0.99))
 351
 352
 353 def valid_c_quizzes(recorded, criteria):
 354     result = [q[criteria(lp)] for q, lp in recorded]
 355     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 356
 357
 358 ######################################################################
 359
 360
 361 def create_c_quizzes(
 362     models,
 363     quiz_machine,
 364     nb_for_train=1000,
 365     nb_for_test=100,
 366 ):
 367     quizzes_and_logproba_records = []
 368
 369     nb_to_create = nb_for_train + nb_for_test
 370
 371     # ------------------------------------------------------------
 372
 373     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 374
 375     with open(file_name, "w") as logp_file:
 376         while (
 377             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 378             < nb_to_create
 379         ):
 380             # Select a model at random to generate the new quizzes
 381
 382             model_for_generation = models[torch.randint(len(models), (1,))]
 383
 384             c_quizzes = quiz_machine.generate_quizzes(
 385                 nb_to_create,
 386                 model_for_generation=model_for_generation,
 387                 temperature=args.generation_temperature,
 388             )
 389
 390             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 391
 392             if c_quizzes.size(0) > 0:
 393                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 394                 for l in logproba:
 395                     s = " ".join([str(x.item()) for x in l])
 396                     logp_file.write(s + "\n")
 397                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 398
 399             nb_validated = valid_c_quizzes(
 400                 quizzes_and_logproba_records, standard_validity
 401             ).size(0)
 402
 403             log_string(
 404                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 405             )
 406
 407     # store the new c_quizzes which have been validated
 408
 409     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 410
 411     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 412
 413     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 414     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 415
 416     # save a bunch of images to investigate what quizzes with a
 417     # certain nb of correct predictions look like
 418
 419     q = new_c_quizzes[:72]
 420
 421     if q.size(0) > 0:
 422         quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
 423
 424
 425 ######################################################################
 426
 427
 428 def create_c_quizzes_(
 429     models,
 430     quiz_machine,
 431     nb_for_train=1000,
 432     nb_for_test=100,
 433 ):
 434     quizzes_and_nb_correct_records = []
 435
 436     nb_to_create = nb_for_train + nb_for_test
 437
 438     # ------------------------------------------------------------
 439
 440     standard_validity = lambda nb_correct: (nb_correct >= args.min_to_validate) & (
 441         nb_correct <= args.max_to_validate
 442     )
 443
 444     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 445
 446     with open(file_name, "w") as logp_file:
 447         while (
 448             valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
 449             < nb_to_create
 450         ):
 451             # Select a model at random to generate the new quizzes
 452
 453             model_for_generation = models[torch.randint(len(models), (1,))]
 454
 455             c_quizzes = quiz_machine.generate_quizzes(
 456                 nb_to_create,
 457                 model_for_generation=model_for_generation,
 458                 temperature=args.generation_temperature,
 459             )
 460
 461             # if args.prediction_correctness:
 462
 463             # else:
 464             # logproba = quiz_machine.new(quiz_machine.size(0), len(models))
 465             # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)):
 466             # for model in models:
 467             # l[...] = F.cross_entropy(model(q))
 468
 469             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 470
 471             if c_quizzes.size(0) > 0:
 472                 nb_correct, seq_logproba = quiz_machine.compute_correctness(
 473                     c_quizzes,
 474                     models,
 475                     bidirectional_validation=args.bidirectional_validation,
 476                     deterministic_validation=args.deterministic_validation,
 477                 )
 478
 479                 for n, l in zip(nb_correct, seq_logproba):
 480                     s = " ".join([str(x.item()) for x in l])
 481                     logp_file.write(f"{n} {s}\n")
 482
 483                 if args.dirty_debug:
 484                     nb_correct = torch.randint(
 485                         len(models) + 1, nb_correct.size(), device=c_quizzes.device
 486                     )
 487
 488                 quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 489
 490             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 491             nv = " ".join([str(x.item()) for x in nv])
 492
 493             nb_validated = valid_c_quizzes(
 494                 quizzes_and_nb_correct_records, standard_validity
 495             ).size(0)
 496
 497             log_string(
 498                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 499             )
 500
 501     # store the new c_quizzes which have been validated
 502
 503     new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
 504
 505     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 506
 507     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 508     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 509
 510     # save a bunch of images to investigate what quizzes with a
 511     # certain nb of correct predictions look like
 512
 513     for n in range(len(models) + 1):
 514         s = (
 515             "_validated"
 516             if n >= args.min_to_validate and n <= args.max_to_validate
 517             else ""
 518         )
 519
 520         q = valid_c_quizzes(
 521             quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
 522         )[:72]
 523
 524         quiz_machine.reverse_random_half_in_place(q)
 525
 526         if q.size(0) > 0:
 527             quiz_machine.save_quizzes(
 528                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 529             )
 530
 531
 532 ######################################################################
 533
 534 models = []
 535
 536 for k in range(args.nb_gpts):
 537     model = mygpt.MyGPT(
 538         vocabulary_size=vocabulary_size,
 539         dim_model=args.dim_model,
 540         dim_keys=args.dim_keys,
 541         dim_hidden=args.dim_hidden,
 542         nb_heads=args.nb_heads,
 543         nb_blocks=args.nb_blocks,
 544         causal=True,
 545         dropout=args.dropout,
 546     ).to(device)
 547
 548     model.main_test_accuracy = 0.0
 549     model.id = k
 550
 551     model.train_w_quizzes = quiz_machine.generate_token_sequences(
 552         args.nb_train_samples
 553     ).to(device)
 554     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 555     model.test_w_quizzes = quiz_machine.generate_token_sequences(
 556         args.nb_test_samples
 557     ).to(device)
 558     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 559
 560     models.append(model)
 561
 562
 563 nb_parameters = sum(p.numel() for p in models[0].parameters())
 564 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 565
 566 ######################################################################
 567
 568 # Compute the entropy of the training tokens
 569
 570 token_count = 0
 571 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 572     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 573         (0, 1)
 574     )
 575 token_probas = token_count / token_count.sum()
 576 entropy = -torch.xlogy(token_probas, token_probas).sum()
 577 train_set_perplexity = math.exp(entropy)
 578
 579 ######################################################################
 580 # A bit of paranoia never hurts
 581
 582 if args.max_percents_of_test_in_train >= 0:
 583
 584     def subsets_as_tuples(batches, cs):
 585         s = set()
 586         for batch in batches:
 587             for x in batch:
 588                 s.add(tuple([v.item() for v in x]))
 589                 if len(s) == cs:
 590                     yield s
 591                     s = set()
 592         yield s
 593
 594     nb_test, nb_in_train = 0, 0
 595     for test_subset in subsets_as_tuples(
 596         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 597     ):
 598         in_train = set()
 599         for train_subset in subsets_as_tuples(
 600             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 601         ):
 602             in_train.update(test_subset.intersection(train_subset))
 603         nb_in_train += len(in_train)
 604         nb_test += len(test_subset)
 605
 606     log_string(
 607         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 608     )
 609
 610     assert (
 611         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 612     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 613
 614 ######################################################################
 615
 616 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 617 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 618
 619 log_string(
 620     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}"
 621 )
 622
 623 ######################################################################
 624
 625 if args.dirty_debug:
 626     args.accuracy_to_make_c_quizzes = 0.0
 627     args.nb_gpts = 2
 628     nb_new_c_quizzes_for_train = 100
 629     nb_new_c_quizzes_for_test = 10
 630
 631 ######################################################################
 632
 633 for n_epoch in range(args.nb_epochs):
 634     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 635
 636     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 637     log_string(f"current_test_accuracies {cta}")
 638
 639     ##################################################
 640     # Select, improve, and eval the worst model
 641
 642     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 643
 644     log_string(
 645         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 646     )
 647
 648     one_epoch(weakest_model, quiz_machine)
 649
 650     log_string(
 651         f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 652     )
 653
 654     run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
 655
 656     log_string(
 657         f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 658     )
 659
 660     ##################################################
 661     # Replace a fraction of the w_quizzes with fresh ones
 662
 663     log_string(
 664         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 665     )
 666     quiz_machine.renew_w_quizzes(model, args.nb_train_samples // args.nb_gpts)
 667
 668     ##################################################
 669     # If all the models are good enough, generate new quizzes and
 670     # re-compute the test errors
 671
 672     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 673         create_c_quizzes(
 674             models,
 675             quiz_machine,
 676             nb_for_train=nb_new_c_quizzes_for_train,
 677             nb_for_test=nb_new_c_quizzes_for_test,
 678         )
 679
 680         for model in models:
 681             run_tests(model, quiz_machine, deterministic_synthesis=False)
 682
 683
 684 ######################################################################