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 import threading
  20
  21 # world quizzes vs. culture quizzes
  22
  23 ######################################################################
  24
  25 if torch.cuda.is_available():
  26     device = torch.device("cuda")
  27     torch.backends.cuda.matmul.allow_tf32 = True
  28 else:
  29     device = torch.device("cpu")
  30
  31 ######################################################################
  32
  33 parser = argparse.ArgumentParser(
  34     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  35 )
  36
  37 parser.add_argument("--log_filename", type=str, default="train.log")
  38
  39 parser.add_argument("--result_dir", type=str, default=None)
  40
  41 parser.add_argument("--seed", type=int, default=0)
  42
  43 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
  44
  45 ########################################
  46
  47 parser.add_argument("--nb_epochs", type=int, default=10000)
  48
  49 parser.add_argument("--batch_size", type=int, default=None)
  50
  51 parser.add_argument("--physical_batch_size", type=int, default=None)
  52
  53 parser.add_argument("--nb_train_samples", type=int, default=None)
  54
  55 parser.add_argument("--nb_test_samples", type=int, default=None)
  56
  57 parser.add_argument("--learning_rate", type=float, default=5e-4)
  58
  59 ########################################
  60
  61 parser.add_argument("--model", type=str, default=None)
  62
  63 parser.add_argument("--dim_model", type=int, default=None)
  64
  65 parser.add_argument("--dim_keys", type=int, default=None)
  66
  67 parser.add_argument("--dim_hidden", type=int, default=None)
  68
  69 parser.add_argument("--nb_heads", type=int, default=None)
  70
  71 parser.add_argument("--nb_blocks", type=int, default=None)
  72
  73 parser.add_argument("--dropout", type=float, default=0.1)
  74
  75 ########################################
  76
  77 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  78
  79 parser.add_argument("--problem", type=str, default="grids")
  80
  81 parser.add_argument("--nb_threads", type=int, default=1)
  82
  83 parser.add_argument("--nb_gpus", type=int, default=1)
  84
  85 parser.add_argument("--nb_gpts", type=int, default=5)
  86
  87 parser.add_argument("--min_to_validate", type=int, default=None)
  88
  89 parser.add_argument("--max_to_validate", type=int, default=None)
  90
  91 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  92
  93 parser.add_argument("--generation_temperature", type=float, default=2.0)
  94
  95 parser.add_argument("--deterministic_validation", action="store_true", default=False)
  96
  97 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
  98
  99 parser.add_argument("--dirty_debug", action="store_true", default=False)
 100
 101 ######################################################################
 102
 103 parser.add_argument("--sky_height", type=int, default=6)
 104
 105 parser.add_argument("--sky_width", type=int, default=8)
 106
 107 parser.add_argument("--sky_nb_birds", type=int, default=3)
 108
 109 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 110
 111 parser.add_argument("--sky_speed", type=int, default=3)
 112
 113 ######################################################################
 114
 115 args = parser.parse_args()
 116
 117 if args.min_to_validate is None:
 118     args.min_to_validate = args.nb_gpts - 1
 119
 120 if args.max_to_validate is None:
 121     args.max_to_validate = args.nb_gpts - 1
 122
 123 if args.result_dir is None:
 124     args.result_dir = f"results_culture"
 125
 126 ######################################################################
 127
 128 default_args = {
 129     "model": "37M",
 130     "batch_size": 25,
 131     "nb_train_samples": 100000,
 132     "nb_test_samples": 10000,
 133 }
 134
 135 for k, v in default_args.items():
 136     if getattr(args, k) is None:
 137         setattr(args, k, v)
 138
 139 ######################################################################
 140
 141 default_model_args = {
 142     "17K": {
 143         "dim_model": 32,
 144         "dim_keys": 32,
 145         "dim_hidden": 32,
 146         "nb_heads": 2,
 147         "nb_blocks": 2,
 148     },
 149     "4M": {
 150         "dim_model": 256,
 151         "dim_keys": 32,
 152         "dim_hidden": 1024,
 153         "nb_heads": 4,
 154         "nb_blocks": 6,
 155     },
 156     "37M": {
 157         "dim_model": 512,
 158         "dim_keys": 64,
 159         "dim_hidden": 2048,
 160         "nb_heads": 8,
 161         "nb_blocks": 12,
 162     },
 163     "122M": {
 164         "dim_model": 768,
 165         "dim_keys": 64,
 166         "dim_hidden": 2048,
 167         "nb_heads": 8,
 168         "nb_blocks": 24,
 169     },
 170     "352M": {
 171         "dim_model": 1024,
 172         "dim_keys": 64,
 173         "dim_hidden": 2048,
 174         "nb_heads": 8,
 175         "nb_blocks": 48,
 176     },
 177 }
 178
 179 if args.model in default_model_args:
 180     for k, v in default_model_args[args.model].items():
 181         if getattr(args, k) is None:
 182             setattr(args, k, v)
 183 else:
 184     raise ValueError(f"Unknown model {args.model}")
 185
 186 ######################################################################
 187
 188 try:
 189     os.mkdir(args.result_dir)
 190 except FileExistsError:
 191     print(f"result directory {args.result_dir} already exists")
 192     exit(1)
 193
 194 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 195
 196 if args.seed >= 0:
 197     # torch.backends.cudnn.deterministic = True
 198     # torch.backends.cudnn.benchmark = False
 199     # torch.use_deterministic_algorithms(True)
 200     torch.manual_seed(args.seed)
 201     if torch.cuda.is_available():
 202         torch.cuda.manual_seed_all(args.seed)
 203
 204 ######################################################################
 205
 206
 207 def log_string(s):
 208     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 209
 210     if log_file is not None:
 211         log_file.write(t + s + "\n")
 212         log_file.flush()
 213
 214     print(t + s)
 215     sys.stdout.flush()
 216
 217
 218 log_string(f"argv {' '.join(sys.argv)}")
 219
 220 for n in vars(args):
 221     log_string(f"args.{n} {getattr(args, n)}")
 222
 223
 224 ######################################################################
 225
 226 if args.dirty_debug:
 227     args.nb_train_samples = 2500
 228     args.nb_test_samples = 100
 229
 230 if args.physical_batch_size is None:
 231     args.physical_batch_size = args.batch_size
 232 else:
 233     assert args.batch_size % args.physical_batch_size == 0
 234
 235 assert args.nb_train_samples % args.batch_size == 0
 236 assert args.nb_test_samples % args.batch_size == 0
 237
 238 if args.problem == "sky":
 239     problem = sky.Sky(
 240         height=args.sky_height,
 241         width=args.sky_width,
 242         nb_birds=args.sky_nb_birds,
 243         nb_iterations=args.sky_nb_iterations,
 244         speed=args.sky_speed,
 245         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 246         chunk_size=100,
 247         nb_threads=args.nb_threads,
 248     )
 249     back_accuracy = False
 250 elif args.problem == "grids":
 251     problem = grids.Grids(
 252         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 253         chunk_size=100,
 254         nb_threads=args.nb_threads,
 255     )
 256     back_accuracy = True
 257 else:
 258     raise ValueError
 259
 260 quiz_machine = quiz_machine.QuizMachine(
 261     problem=problem,
 262     nb_train_samples=args.nb_train_samples,
 263     nb_test_samples=args.nb_test_samples,
 264     back_accuracy=back_accuracy,
 265     batch_size=args.physical_batch_size,
 266     result_dir=args.result_dir,
 267     logger=log_string,
 268     device=device,
 269 )
 270
 271 ######################################################################
 272
 273 log_string(f"device {device}")
 274
 275 vocabulary_size = quiz_machine.vocabulary_size()
 276
 277 log_string(f"vocabulary_size {vocabulary_size}")
 278
 279 ######################################################################
 280
 281
 282 ######################################################################
 283
 284
 285 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
 286     if local_device is None:
 287         local_device = device
 288
 289     with torch.autograd.no_grad():
 290         model.eval().to(local_device)
 291
 292         nb_test_samples, acc_test_loss = 0, 0.0
 293         nb_samples_accumulated = 0
 294
 295         for input in quiz_machine.batches(model, split="test"):
 296             input = input.to(local_device)
 297
 298             bs = model(mygpt.BracketedSequence(input))
 299             output = bs.x
 300
 301             loss = F.cross_entropy(output.transpose(1, 2), input)
 302
 303             acc_test_loss += loss.item() * input.size(0)
 304
 305             nb_test_samples += input.size(0)
 306
 307         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 308
 309         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 310
 311         model.main_test_accuracy = quiz_machine.produce_results(
 312             n_epoch=n_epoch,
 313             model=model,
 314             result_dir=args.result_dir,
 315             deterministic_synthesis=deterministic_synthesis,
 316         )
 317
 318
 319 def one_epoch(model, quiz_machine, local_device=None):
 320     if local_device is None:
 321         local_device = device
 322
 323     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 324
 325     model.to(local_device).train()
 326
 327     nb_train_samples, acc_train_loss = 0, 0.0
 328
 329     for input in quiz_machine.batches(model, split="train"):
 330         input = input.to(local_device)
 331
 332         if nb_train_samples % args.batch_size == 0:
 333             optimizer.zero_grad()
 334
 335         output = model(mygpt.BracketedSequence(input)).x
 336         loss = F.cross_entropy(output.transpose(1, 2), input)
 337         acc_train_loss += loss.item() * input.size(0)
 338
 339         nb_train_samples += input.size(0)
 340
 341         loss.backward()
 342
 343         if nb_train_samples % args.batch_size == 0:
 344             optimizer.step()
 345
 346     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 347
 348     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 349
 350     run_tests(model, quiz_machine, deterministic_synthesis=False)
 351
 352     model.TRAINING_LOCK.release()
 353
 354
 355 ######################################################################
 356
 357
 358 def standard_validity(logproba):
 359     l = logproba.sort(dim=-1).values
 360     return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
 361     # warnings.warn("TEST!!!", RuntimeWarning)
 362     # print(l.exp())
 363     # return (l[:, 0] < math.log(0.99))
 364
 365
 366 def valid_c_quizzes(recorded, criteria):
 367     result = [q[criteria(lp)] for q, lp in recorded]
 368     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 369
 370
 371 ######################################################################
 372
 373
 374 def create_c_quizzes(
 375     models,
 376     quiz_machine,
 377     nb_for_train=1000,
 378     nb_for_test=100,
 379 ):
 380     quizzes_and_logproba_records = []
 381
 382     nb_to_create = nb_for_train + nb_for_test
 383
 384     # ------------------------------------------------------------
 385
 386     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 387
 388     with open(file_name, "w") as logp_file:
 389         while (
 390             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 391             < nb_to_create
 392         ):
 393             # Select a model at random to generate the new quizzes
 394
 395             model_for_generation = models[torch.randint(len(models), (1,))]
 396
 397             c_quizzes = quiz_machine.generate_quizzes(
 398                 nb_to_create,
 399                 model_for_generation=model_for_generation,
 400                 temperature=args.generation_temperature,
 401             )
 402
 403             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 404
 405             if c_quizzes.size(0) > 0:
 406                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 407                 for l in logproba:
 408                     s = " ".join([str(x.item()) for x in l])
 409                     logp_file.write(s + "\n")
 410                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 411
 412             nb_validated = valid_c_quizzes(
 413                 quizzes_and_logproba_records, standard_validity
 414             ).size(0)
 415
 416             log_string(
 417                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 418             )
 419
 420     # store the new c_quizzes which have been validated
 421
 422     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 423
 424     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 425
 426     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 427     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 428
 429     # save a bunch of images to investigate what quizzes with a
 430     # certain nb of correct predictions look like
 431
 432     q = new_c_quizzes[:72]
 433
 434     if q.size(0) > 0:
 435         quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
 436
 437
 438 ######################################################################
 439
 440
 441 def create_c_quizzes_(
 442     models,
 443     quiz_machine,
 444     nb_for_train=1000,
 445     nb_for_test=100,
 446 ):
 447     quizzes_and_nb_correct_records = []
 448
 449     nb_to_create = nb_for_train + nb_for_test
 450
 451     # ------------------------------------------------------------
 452
 453     standard_validity = lambda nb_correct: (nb_correct >= args.min_to_validate) & (
 454         nb_correct <= args.max_to_validate
 455     )
 456
 457     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 458
 459     with open(file_name, "w") as logp_file:
 460         while (
 461             valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
 462             < nb_to_create
 463         ):
 464             # Select a model at random to generate the new quizzes
 465
 466             model_for_generation = models[torch.randint(len(models), (1,))]
 467
 468             c_quizzes = quiz_machine.generate_quizzes(
 469                 nb_to_create,
 470                 model_for_generation=model_for_generation,
 471                 temperature=args.generation_temperature,
 472             )
 473
 474             # if args.prediction_correctness:
 475
 476             # else:
 477             # logproba = quiz_machine.new(quiz_machine.size(0), len(models))
 478             # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)):
 479             # for model in models:
 480             # l[...] = F.cross_entropy(model(q))
 481
 482             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 483
 484             if c_quizzes.size(0) > 0:
 485                 nb_correct, seq_logproba = quiz_machine.compute_correctness(
 486                     c_quizzes,
 487                     models,
 488                     bidirectional_validation=args.bidirectional_validation,
 489                     deterministic_validation=args.deterministic_validation,
 490                 )
 491
 492                 for n, l in zip(nb_correct, seq_logproba):
 493                     s = " ".join([str(x.item()) for x in l])
 494                     logp_file.write(f"{n} {s}\n")
 495
 496                 if args.dirty_debug:
 497                     nb_correct = torch.randint(
 498                         len(models) + 1, nb_correct.size(), device=c_quizzes.device
 499                     )
 500
 501                 quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 502
 503             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 504             nv = " ".join([str(x.item()) for x in nv])
 505
 506             nb_validated = valid_c_quizzes(
 507                 quizzes_and_nb_correct_records, standard_validity
 508             ).size(0)
 509
 510             log_string(
 511                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 512             )
 513
 514     # store the new c_quizzes which have been validated
 515
 516     new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
 517
 518     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 519
 520     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 521     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 522
 523     # save a bunch of images to investigate what quizzes with a
 524     # certain nb of correct predictions look like
 525
 526     for n in range(len(models) + 1):
 527         s = (
 528             "_validated"
 529             if n >= args.min_to_validate and n <= args.max_to_validate
 530             else ""
 531         )
 532
 533         q = valid_c_quizzes(
 534             quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
 535         )[:72]
 536
 537         quiz_machine.reverse_random_half_in_place(q)
 538
 539         if q.size(0) > 0:
 540             quiz_machine.save_quizzes(
 541                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 542             )
 543
 544
 545 ######################################################################
 546
 547 models = []
 548
 549 for k in range(args.nb_gpts):
 550     log_string(f"creating model {k} and its w_quizzes")
 551     model = mygpt.MyGPT(
 552         vocabulary_size=vocabulary_size,
 553         dim_model=args.dim_model,
 554         dim_keys=args.dim_keys,
 555         dim_hidden=args.dim_hidden,
 556         nb_heads=args.nb_heads,
 557         nb_blocks=args.nb_blocks,
 558         causal=True,
 559         dropout=args.dropout,
 560     ).to(device)
 561
 562     model.main_test_accuracy = 0.0
 563     model.id = k
 564     model.TRAINING_LOCK = threading.Lock()
 565
 566     model.train_w_quizzes = quiz_machine.generate_token_sequences(
 567         args.nb_train_samples
 568     ).to(device)
 569     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 570     model.test_w_quizzes = quiz_machine.generate_token_sequences(
 571         args.nb_test_samples
 572     ).to(device)
 573     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 574
 575     models.append(model)
 576
 577
 578 nb_parameters = sum(p.numel() for p in models[0].parameters())
 579 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 580
 581 ######################################################################
 582
 583 # Compute the entropy of the training tokens
 584
 585 token_count = 0
 586 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 587     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 588         (0, 1)
 589     )
 590 token_probas = token_count / token_count.sum()
 591 entropy = -torch.xlogy(token_probas, token_probas).sum()
 592 train_set_perplexity = math.exp(entropy)
 593
 594 ######################################################################
 595 # A bit of paranoia never hurts
 596
 597 if args.max_percents_of_test_in_train >= 0:
 598
 599     def subsets_as_tuples(batches, cs):
 600         s = set()
 601         for batch in batches:
 602             for x in batch:
 603                 s.add(tuple([v.item() for v in x]))
 604                 if len(s) == cs:
 605                     yield s
 606                     s = set()
 607         yield s
 608
 609     nb_test, nb_in_train = 0, 0
 610     for test_subset in subsets_as_tuples(
 611         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 612     ):
 613         in_train = set()
 614         for train_subset in subsets_as_tuples(
 615             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 616         ):
 617             in_train.update(test_subset.intersection(train_subset))
 618         nb_in_train += len(in_train)
 619         nb_test += len(test_subset)
 620
 621     log_string(
 622         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 623     )
 624
 625     assert (
 626         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 627     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 628
 629 ######################################################################
 630
 631 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 632 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 633
 634 log_string(
 635     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}"
 636 )
 637
 638 ######################################################################
 639
 640 if args.dirty_debug:
 641     args.accuracy_to_make_c_quizzes = 0.0
 642     args.nb_gpts = 2
 643     nb_new_c_quizzes_for_train = 100
 644     nb_new_c_quizzes_for_test = 10
 645
 646 ######################################################################
 647
 648 for n_epoch in range(args.nb_epochs):
 649     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 650
 651     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 652     log_string(f"current_test_accuracies {cta}")
 653
 654     ##################################################
 655     # Select, improve, and eval the worst model
 656
 657     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 658
 659     weakest_models = ranked_models[: args.nb_gpus]
 660
 661     for gpu_id, model in enumerate(weakest_models):
 662         model.TRAINING_LOCK.acquire()
 663
 664         log_string(
 665             f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 666         )
 667
 668         threading.Thread(
 669             target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
 670         ).start()
 671
 672     for model in weakest_models:
 673         model.TRAINING_LOCK.acquire()
 674         model.TRAINING_LOCK.release()
 675
 676     ##################################################
 677     # Replace a fraction of the w_quizzes with fresh ones
 678
 679     log_string(
 680         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 681     )
 682
 683     # Renew entirely the train set
 684
 685     for model in weakest_models:
 686         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 687
 688     ##################################################
 689     # If all the models are good enough, generate new quizzes and
 690     # re-compute the test errors
 691
 692     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 693         create_c_quizzes(
 694             models,
 695             quiz_machine,
 696             nb_for_train=nb_new_c_quizzes_for_train,
 697             nb_for_test=nb_new_c_quizzes_for_test,
 698         )
 699
 700 ######################################################################