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     description="An implementation of GPT with cache.",
  33     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  34 )
  35
  36 parser.add_argument("--log_filename", type=str, default="train.log")
  37
  38 parser.add_argument("--result_dir", type=str, default=None)
  39
  40 parser.add_argument("--seed", type=int, default=0)
  41
  42 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  43
  44 ########################################
  45
  46 parser.add_argument("--nb_epochs", type=int, default=10000)
  47
  48 parser.add_argument("--batch_size", type=int, default=None)
  49
  50 parser.add_argument("--physical_batch_size", type=int, default=None)
  51
  52 parser.add_argument("--nb_train_samples", type=int, default=None)
  53
  54 parser.add_argument("--nb_test_samples", type=int, default=None)
  55
  56 parser.add_argument("--learning_rate", type=float, default=5e-4)
  57
  58 ########################################
  59
  60 parser.add_argument("--model", type=str, default=None)
  61
  62 parser.add_argument("--dim_model", type=int, default=None)
  63
  64 parser.add_argument("--dim_keys", type=int, default=None)
  65
  66 parser.add_argument("--dim_hidden", type=int, default=None)
  67
  68 parser.add_argument("--nb_heads", type=int, default=None)
  69
  70 parser.add_argument("--nb_blocks", type=int, default=None)
  71
  72 parser.add_argument("--dropout", type=float, default=0.1)
  73
  74 ########################################
  75
  76 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  77
  78 parser.add_argument("--problem", type=str, default="grids")
  79
  80 parser.add_argument("--nb_threads", type=int, default=-1)
  81
  82 parser.add_argument("--nb_gpts", type=int, default=5)
  83
  84 parser.add_argument("--min_to_validate", type=int, default=None)
  85
  86 parser.add_argument("--max_to_validate", type=int, default=None)
  87
  88 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  89
  90 parser.add_argument("--generation_temperature", type=float, default=2.0)
  91
  92 parser.add_argument("--deterministic_validation", action="store_true", default=False)
  93
  94 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
  95
  96 parser.add_argument("--dirty_debug", action="store_true", default=False)
  97
  98 ######################################################################
  99
 100 parser.add_argument("--sky_height", type=int, default=6)
 101
 102 parser.add_argument("--sky_width", type=int, default=8)
 103
 104 parser.add_argument("--sky_nb_birds", type=int, default=3)
 105
 106 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 107
 108 parser.add_argument("--sky_speed", type=int, default=3)
 109
 110 ######################################################################
 111
 112 args = parser.parse_args()
 113
 114 if args.min_to_validate is None:
 115     args.min_to_validate = args.nb_gpts - 1
 116
 117 if args.max_to_validate is None:
 118     args.max_to_validate = args.nb_gpts - 1
 119
 120 if args.result_dir is None:
 121     args.result_dir = f"results_culture"
 122
 123 ######################################################################
 124
 125 default_args = {
 126     "model": "37M",
 127     "batch_size": 25,
 128     "nb_train_samples": 100000,
 129     "nb_test_samples": 10000,
 130 }
 131
 132 for k, v in default_args.items():
 133     if getattr(args, k) is None:
 134         setattr(args, k, v)
 135
 136 ######################################################################
 137
 138 default_model_args = {
 139     "17K": {
 140         "dim_model": 32,
 141         "dim_keys": 32,
 142         "dim_hidden": 32,
 143         "nb_heads": 2,
 144         "nb_blocks": 2,
 145     },
 146     "4M": {
 147         "dim_model": 256,
 148         "dim_keys": 32,
 149         "dim_hidden": 1024,
 150         "nb_heads": 4,
 151         "nb_blocks": 6,
 152     },
 153     "37M": {
 154         "dim_model": 512,
 155         "dim_keys": 64,
 156         "dim_hidden": 2048,
 157         "nb_heads": 8,
 158         "nb_blocks": 12,
 159     },
 160     "122M": {
 161         "dim_model": 768,
 162         "dim_keys": 64,
 163         "dim_hidden": 2048,
 164         "nb_heads": 8,
 165         "nb_blocks": 24,
 166     },
 167     "352M": {
 168         "dim_model": 1024,
 169         "dim_keys": 64,
 170         "dim_hidden": 2048,
 171         "nb_heads": 8,
 172         "nb_blocks": 48,
 173     },
 174 }
 175
 176 if args.model in default_model_args:
 177     for k, v in default_model_args[args.model].items():
 178         if getattr(args, k) is None:
 179             setattr(args, k, v)
 180 else:
 181     raise ValueError(f"Unknown model {args.model}")
 182
 183 ######################################################################
 184
 185 try:
 186     os.mkdir(args.result_dir)
 187 except FileExistsError:
 188     print(f"result directory {args.result_dir} already exists")
 189     exit(1)
 190
 191 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 192
 193 if args.seed >= 0:
 194     # torch.backends.cudnn.deterministic = True
 195     # torch.backends.cudnn.benchmark = False
 196     # torch.use_deterministic_algorithms(True)
 197     torch.manual_seed(args.seed)
 198     if torch.cuda.is_available():
 199         torch.cuda.manual_seed_all(args.seed)
 200
 201 ######################################################################
 202
 203
 204 def log_string(s):
 205     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 206
 207     if log_file is not None:
 208         log_file.write(t + s + "\n")
 209         log_file.flush()
 210
 211     print(t + s)
 212     sys.stdout.flush()
 213
 214
 215 log_string(f"argv {' '.join(sys.argv)}")
 216
 217 for n in vars(args):
 218     log_string(f"args.{n} {getattr(args, n)}")
 219
 220
 221 ######################################################################
 222
 223 if args.dirty_debug:
 224     args.nb_train_samples = 2500
 225     args.nb_test_samples = 100
 226
 227 if args.physical_batch_size is None:
 228     args.physical_batch_size = args.batch_size
 229 else:
 230     assert args.batch_size % args.physical_batch_size == 0
 231
 232 assert args.nb_train_samples % args.batch_size == 0
 233 assert args.nb_test_samples % args.batch_size == 0
 234
 235 if args.problem == "sky":
 236     problem = sky.Sky(
 237         height=args.sky_height,
 238         width=args.sky_width,
 239         nb_birds=args.sky_nb_birds,
 240         nb_iterations=args.sky_nb_iterations,
 241         speed=args.sky_speed,
 242         max_nb_cached_chunks=args.nb_train_samples // 100,
 243         chunk_size=100,
 244         nb_threads=args.nb_threads,
 245     )
 246     back_accuracy = False
 247 elif args.problem == "grids":
 248     problem = grids.Grids(
 249         device=device,
 250         max_nb_cached_chunks=args.nb_train_samples // 100,
 251         chunk_size=100,
 252         nb_threads=args.nb_threads,
 253     )
 254     back_accuracy = True
 255 else:
 256     raise ValueError
 257
 258 quiz_machine = quiz_machine.QuizMachine(
 259     problem=problem,
 260     nb_train_samples=args.nb_train_samples,
 261     nb_test_samples=args.nb_test_samples,
 262     back_accuracy=back_accuracy,
 263     batch_size=args.physical_batch_size,
 264     result_dir=args.result_dir,
 265     logger=log_string,
 266     device=device,
 267 )
 268
 269 ######################################################################
 270
 271 log_string(f"device {device}")
 272
 273 vocabulary_size = quiz_machine.vocabulary_size()
 274
 275 log_string(f"vocabulary_size {vocabulary_size}")
 276
 277 ######################################################################
 278
 279 # Compute the entropy of the training tokens
 280
 281 token_count = 0
 282 for input in quiz_machine.batches(split="train", desc="train-entropy"):
 283     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 284         (0, 1)
 285     )
 286 token_probas = token_count / token_count.sum()
 287 entropy = -torch.xlogy(token_probas, token_probas).sum()
 288 train_set_perplexity = math.exp(entropy)
 289
 290 ######################################################################
 291 # A bit of paranoia never hurts
 292
 293 if args.max_percents_of_test_in_train >= 0:
 294
 295     def subsets_as_tuples(batches, cs):
 296         s = set()
 297         for batch in batches:
 298             for x in batch:
 299                 s.add(tuple([v.item() for v in x]))
 300                 if len(s) == cs:
 301                     yield s
 302                     s = set()
 303         yield s
 304
 305     nb_test, nb_in_train = 0, 0
 306     for test_subset in subsets_as_tuples(
 307         quiz_machine.batches(split="test", desc="test-check"), 25000
 308     ):
 309         in_train = set()
 310         for train_subset in subsets_as_tuples(
 311             quiz_machine.batches(split="train", desc="train-check"), 25000
 312         ):
 313             in_train.update(test_subset.intersection(train_subset))
 314         nb_in_train += len(in_train)
 315         nb_test += len(test_subset)
 316
 317     log_string(
 318         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 319     )
 320
 321     assert (
 322         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 323     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 324
 325 ##############################
 326
 327
 328 def one_epoch(model, quiz_machine):
 329     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 330
 331     model.train()
 332
 333     nb_train_samples, acc_train_loss = 0, 0.0
 334
 335     for input in quiz_machine.batches(split="train"):
 336         input = input.to(device)
 337
 338         if nb_train_samples % args.batch_size == 0:
 339             optimizer.zero_grad()
 340
 341         output = model(mygpt.BracketedSequence(input)).x
 342         loss = F.cross_entropy(output.transpose(1, 2), input)
 343         acc_train_loss += loss.item() * input.size(0)
 344
 345         nb_train_samples += input.size(0)
 346
 347         loss.backward()
 348
 349         if nb_train_samples % args.batch_size == 0:
 350             optimizer.step()
 351
 352     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 353
 354     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 355
 356
 357 ######################################################################
 358
 359
 360 def run_tests(model, quiz_machine, deterministic_synthesis):
 361     with torch.autograd.no_grad():
 362         model.eval()
 363
 364         nb_test_samples, acc_test_loss = 0, 0.0
 365         nb_samples_accumulated = 0
 366
 367         for input in quiz_machine.batches(split="test"):
 368             input = input.to(device)
 369
 370             bs = model(mygpt.BracketedSequence(input))
 371             output = bs.x
 372
 373             loss = F.cross_entropy(output.transpose(1, 2), input)
 374
 375             acc_test_loss += loss.item() * input.size(0)
 376
 377             nb_test_samples += input.size(0)
 378
 379         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 380
 381         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 382
 383         model.main_test_accuracy = quiz_machine.produce_results(
 384             n_epoch=n_epoch,
 385             model=model,
 386             result_dir=args.result_dir,
 387             deterministic_synthesis=deterministic_synthesis,
 388         )
 389
 390
 391 ######################################################################
 392
 393
 394 def standard_validity(logproba):
 395     l = logproba.sort(dim=-1).values
 396     return logical_and(l[0] < math.log(0.5), l[1] > math.log(0.95))
 397
 398
 399 def valid_c_quizzes(recorded, criteria):
 400     result = [q[criteria(lp)] for q, lp in recorded]
 401     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 402
 403
 404 ######################################################################
 405
 406
 407 def create_c_quizzes(
 408     models,
 409     quiz_machine,
 410     nb_for_train=1000,
 411     nb_for_test=100,
 412 ):
 413     quizzes_and_logproba_records = []
 414
 415     nb_to_create = nb_for_train + nb_for_test
 416
 417     # ------------------------------------------------------------
 418
 419     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 420
 421     with open(file_name, "w") as logp_file:
 422         while (
 423             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 424             < nb_to_create
 425         ):
 426             # Select a model at random to generate the new quizzes
 427
 428             model_for_generation = models[torch.randint(len(models), (1,))]
 429
 430             c_quizzes = quiz_machine.generate_quizzes(
 431                 nb_to_create,
 432                 model_for_generation=model_for_generation,
 433                 temperature=args.generation_temperature,
 434             )
 435
 436             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 437
 438             if c_quizzes.size(0) > 0:
 439                 logproba = c_quizzes.new(c_quizzes.size(0), len(models))
 440                 for q, l in zip(
 441                     c_quizzes.split(args.batch_size), logits.split(args.batch_size)
 442                 ):
 443                     for model in models:
 444                         l[model.id] = F.cross_entropy(model(q))
 445
 446                 for l in logproba:
 447                     s = " ".join([str(x.item()) for x in l])
 448                     logp_file.write(s + "\n")
 449
 450                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 451
 452             nb_validated = valid_c_quizzes(
 453                 quizzes_and_logproba_records, standard_validity
 454             ).size(0)
 455
 456             log_string(
 457                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 458             )
 459
 460     # store the new c_quizzes which have been validated
 461
 462     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 463
 464     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 465
 466     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 467     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 468
 469     # save a bunch of images to investigate what quizzes with a
 470     # certain nb of correct predictions look like
 471
 472     q = new_c_quizzes[:72]
 473
 474     if q.size(0) > 0:
 475         quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
 476
 477
 478 ######################################################################
 479
 480
 481 def create_c_quizzes_(
 482     models,
 483     quiz_machine,
 484     nb_for_train=1000,
 485     nb_for_test=100,
 486 ):
 487     quizzes_and_nb_correct_records = []
 488
 489     nb_to_create = nb_for_train + nb_for_test
 490
 491     # ------------------------------------------------------------
 492
 493     standard_validity = lambda nb_correct: torch.logical_and(
 494         nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
 495     )
 496
 497     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 498
 499     with open(file_name, "w") as logp_file:
 500         while (
 501             valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
 502             < nb_to_create
 503         ):
 504             # Select a model at random to generate the new quizzes
 505
 506             model_for_generation = models[torch.randint(len(models), (1,))]
 507
 508             c_quizzes = quiz_machine.generate_quizzes(
 509                 nb_to_create,
 510                 model_for_generation=model_for_generation,
 511                 temperature=args.generation_temperature,
 512             )
 513
 514             # if args.prediction_correctness:
 515
 516             # else:
 517             # logproba = quiz_machine.new(quiz_machine.size(0), len(models))
 518             # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)):
 519             # for model in models:
 520             # l[...] = F.cross_entropy(model(q))
 521
 522             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 523
 524             if c_quizzes.size(0) > 0:
 525                 nb_correct, seq_logproba = quiz_machine.compute_correctness(
 526                     c_quizzes,
 527                     models,
 528                     bidirectional_validation=args.bidirectional_validation,
 529                     deterministic_validation=args.deterministic_validation,
 530                 )
 531
 532                 for n, l in zip(nb_correct, seq_logproba):
 533                     s = " ".join([str(x.item()) for x in l])
 534                     logp_file.write(f"{n} {s}\n")
 535
 536                 if args.dirty_debug:
 537                     nb_correct = torch.randint(
 538                         len(models) + 1, nb_correct.size(), device=c_quizzes.device
 539                     )
 540
 541                 quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 542
 543             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 544             nv = " ".join([str(x.item()) for x in nv])
 545
 546             nb_validated = valid_c_quizzes(
 547                 quizzes_and_nb_correct_records, standard_validity
 548             ).size(0)
 549
 550             log_string(
 551                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 552             )
 553
 554     # store the new c_quizzes which have been validated
 555
 556     new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
 557
 558     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 559
 560     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 561     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 562
 563     # save a bunch of images to investigate what quizzes with a
 564     # certain nb of correct predictions look like
 565
 566     for n in range(len(models) + 1):
 567         s = (
 568             "_validated"
 569             if n >= args.min_to_validate and n <= args.max_to_validate
 570             else ""
 571         )
 572
 573         q = valid_c_quizzes(
 574             quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
 575         )[:72]
 576
 577         quiz_machine.reverse_random_half_in_place(q)
 578
 579         if q.size(0) > 0:
 580             quiz_machine.save_quizzes(
 581                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 582             )
 583
 584
 585 ######################################################################
 586
 587 models = []
 588
 589 for k in range(args.nb_gpts):
 590     model = mygpt.MyGPT(
 591         vocabulary_size=vocabulary_size,
 592         dim_model=args.dim_model,
 593         dim_keys=args.dim_keys,
 594         dim_hidden=args.dim_hidden,
 595         nb_heads=args.nb_heads,
 596         nb_blocks=args.nb_blocks,
 597         causal=True,
 598         dropout=args.dropout,
 599     ).to(device)
 600
 601     model.main_test_accuracy = 0.0
 602     model.id = k
 603
 604     models.append(model)
 605
 606
 607 nb_parameters = sum(p.numel() for p in models[0].parameters())
 608 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 609
 610 ######################################################################
 611
 612 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 613 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 614
 615 log_string(
 616     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}"
 617 )
 618
 619 ######################################################################
 620
 621 if args.dirty_debug:
 622     args.accuracy_to_make_c_quizzes = 0.0
 623     args.nb_gpts = 2
 624     nb_new_c_quizzes_for_train = 100
 625     nb_new_c_quizzes_for_test = 10
 626
 627 ######################################################################
 628
 629 for n_epoch in range(args.nb_epochs):
 630     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 631
 632     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 633     log_string(f"current_test_accuracies {cta}")
 634
 635     ##################################################
 636     # Select, improve, and eval the worst model
 637
 638     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 639
 640     log_string(
 641         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 642     )
 643
 644     one_epoch(weakest_model, quiz_machine)
 645
 646     log_string(
 647         f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 648     )
 649
 650     run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
 651
 652     log_string(
 653         f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 654     )
 655
 656     ##################################################
 657     # Replace a fraction of the w_quizzes with fresh ones
 658
 659     quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 660
 661     ##################################################
 662     # If all the models are good enough, generate new quizzes and
 663     # re-compute the test errors
 664
 665     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 666         create_c_quizzes(
 667             models,
 668             quiz_machine,
 669             nb_for_train=nb_new_c_quizzes_for_train,
 670             nb_for_test=nb_new_c_quizzes_for_test,
 671         )
 672
 673         for model in models:
 674             run_tests(model, quiz_machine, deterministic_synthesis=False)
 675
 676
 677 ######################################################################