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 import torch.multiprocessing as mp
  22
  23 ######################################################################
  24
  25 parser = argparse.ArgumentParser(
  26     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  27 )
  28
  29 parser.add_argument("--log_filename", type=str, default="train.log")
  30
  31 parser.add_argument("--result_dir", type=str, default=None)
  32
  33 parser.add_argument("--seed", type=int, default=0)
  34
  35 parser.add_argument("--resume", action="store_true", default=False)
  36
  37 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
  38
  39 ########################################
  40
  41 parser.add_argument("--nb_epochs", type=int, default=10000)
  42
  43 parser.add_argument("--batch_size", type=int, default=None)
  44
  45 parser.add_argument("--physical_batch_size", type=int, default=None)
  46
  47 parser.add_argument("--nb_train_samples", type=int, default=None)
  48
  49 parser.add_argument("--nb_test_samples", type=int, default=None)
  50
  51 parser.add_argument("--nb_new_c_quizzes_for_train", type=int, default=None)
  52
  53 parser.add_argument("--nb_new_c_quizzes_for_test", 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("--gpus", type=str, default="all")
  82
  83 parser.add_argument("--nb_gpts", type=int, default=5)
  84
  85 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.9)
  86
  87 parser.add_argument("--proba_understands", type=float, default=0.9)
  88
  89 parser.add_argument("--proba_not_understands", type=float, default=0.5)
  90
  91 parser.add_argument("--generation_temperature", type=float, default=1.0)
  92
  93 parser.add_argument("--dirty_debug", action="store_true", default=False)
  94
  95 ######################################################################
  96
  97 grids_tasks = ", ".join(
  98     [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
  99 )
 100
 101 parser.add_argument(
 102     "--grids_tasks",
 103     type=str,
 104     default=None,
 105     help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
 106 )
 107
 108 ######################################################################
 109
 110 parser.add_argument("--sky_height", type=int, default=6)
 111
 112 parser.add_argument("--sky_width", type=int, default=8)
 113
 114 parser.add_argument("--sky_nb_birds", type=int, default=3)
 115
 116 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 117
 118 parser.add_argument("--sky_speed", type=int, default=3)
 119
 120 ######################################################################
 121
 122 args = parser.parse_args()
 123
 124 if args.result_dir is None:
 125     args.result_dir = f"results_culture"
 126
 127 ######################################################################
 128
 129 default_args = {
 130     "model": "37M",
 131     "batch_size": 25,
 132     "nb_train_samples": 100000,
 133     "nb_test_samples": 10000,
 134 }
 135
 136 for k, v in default_args.items():
 137     if getattr(args, k) is None:
 138         setattr(args, k, v)
 139
 140 ######################################################################
 141
 142 default_model_args = {
 143     "17K": {
 144         "dim_model": 32,
 145         "dim_keys": 32,
 146         "dim_hidden": 32,
 147         "nb_heads": 2,
 148         "nb_blocks": 2,
 149     },
 150     "4M": {
 151         "dim_model": 256,
 152         "dim_keys": 32,
 153         "dim_hidden": 1024,
 154         "nb_heads": 4,
 155         "nb_blocks": 6,
 156     },
 157     "37M": {
 158         "dim_model": 512,
 159         "dim_keys": 64,
 160         "dim_hidden": 2048,
 161         "nb_heads": 8,
 162         "nb_blocks": 12,
 163     },
 164     "122M": {
 165         "dim_model": 768,
 166         "dim_keys": 64,
 167         "dim_hidden": 2048,
 168         "nb_heads": 8,
 169         "nb_blocks": 24,
 170     },
 171     "352M": {
 172         "dim_model": 1024,
 173         "dim_keys": 64,
 174         "dim_hidden": 2048,
 175         "nb_heads": 8,
 176         "nb_blocks": 48,
 177     },
 178 }
 179
 180 if args.model in default_model_args:
 181     for k, v in default_model_args[args.model].items():
 182         if getattr(args, k) is None:
 183             setattr(args, k, v)
 184 else:
 185     raise ValueError(f"Unknown model {args.model}")
 186
 187 ######################################################################
 188
 189 if args.resume:
 190     assert os.path.isdir(args.result_dir)
 191
 192 else:
 193     try:
 194         os.mkdir(args.result_dir)
 195     except FileExistsError:
 196         print(f"result directory {args.result_dir} already exists")
 197         exit(1)
 198
 199 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 200
 201 if args.seed >= 0:
 202     # torch.backends.cudnn.deterministic = True
 203     # torch.backends.cudnn.benchmark = False
 204     # torch.use_deterministic_algorithms(True)
 205     torch.manual_seed(args.seed)
 206     if torch.cuda.is_available():
 207         torch.cuda.manual_seed_all(args.seed)
 208
 209 ######################################################################
 210
 211
 212 def log_string(s):
 213     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 214
 215     if log_file is not None:
 216         log_file.write(t + s + "\n")
 217         log_file.flush()
 218
 219     print(t + s)
 220     sys.stdout.flush()
 221
 222
 223 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 224
 225 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py")
 226
 227 log_string(f"argv {' '.join(sys.argv)}")
 228
 229 for n in vars(args):
 230     log_string(f"args.{n} {getattr(args, n)}")
 231
 232
 233 ######################################################################
 234
 235 if args.gpus == "all":
 236     gpus_idx = range(torch.cuda.device_count())
 237 else:
 238     gpus_idx = [int(k) for k in args.gpus.split(",")]
 239
 240 gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
 241
 242 if torch.cuda.is_available():
 243     main_device = gpus[0]
 244 else:
 245     assert len(gpus) == 0
 246     main_device = torch.device("cpu")
 247
 248 if args.dirty_debug:
 249     args.nb_train_samples = 2500
 250     args.nb_test_samples = 100
 251
 252 if args.physical_batch_size is None:
 253     args.physical_batch_size = args.batch_size
 254 else:
 255     assert args.batch_size % args.physical_batch_size == 0
 256
 257 assert args.nb_train_samples % args.batch_size == 0
 258 assert args.nb_test_samples % args.batch_size == 0
 259
 260 if args.problem == "sky":
 261     problem = sky.Sky(
 262         height=args.sky_height,
 263         width=args.sky_width,
 264         nb_birds=args.sky_nb_birds,
 265         nb_iterations=args.sky_nb_iterations,
 266         speed=args.sky_speed,
 267         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 268         chunk_size=100,
 269         nb_threads=args.nb_threads,
 270     )
 271     back_accuracy = False
 272 elif args.problem == "grids":
 273     problem = grids.Grids(
 274         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 275         chunk_size=100,
 276         nb_threads=args.nb_threads,
 277         tasks=args.grids_tasks,
 278     )
 279     back_accuracy = True
 280 else:
 281     raise ValueError
 282
 283 problem.save_some_examples(args.result_dir)
 284
 285 quiz_machine = quiz_machine.QuizMachine(
 286     problem=problem,
 287     nb_train_samples=args.nb_train_samples,
 288     nb_test_samples=args.nb_test_samples,
 289     back_accuracy=back_accuracy,
 290     batch_size=args.physical_batch_size,
 291     result_dir=args.result_dir,
 292     logger=log_string,
 293     device=main_device,
 294 )
 295
 296 ######################################################################
 297
 298 log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
 299
 300 vocabulary_size = quiz_machine.vocabulary_size()
 301
 302 log_string(f"vocabulary_size {vocabulary_size}")
 303
 304 ######################################################################
 305
 306
 307 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
 308     with torch.autograd.no_grad():
 309         model.eval().to(local_device)
 310
 311         nb_test_samples, acc_test_loss = 0, 0.0
 312         nb_samples_accumulated = 0
 313
 314         for input in quiz_machine.batches(model, split="test"):
 315             input = input.to(local_device)
 316
 317             bs = model(mygpt.BracketedSequence(input))
 318             output = bs.x
 319
 320             loss = F.cross_entropy(output.transpose(1, 2), input)
 321
 322             acc_test_loss += loss.item() * input.size(0)
 323
 324             nb_test_samples += input.size(0)
 325
 326         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 327
 328         log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
 329
 330         model.main_test_accuracy = quiz_machine.produce_results(
 331             n_epoch=n_epoch,
 332             model=model,
 333             result_dir=args.result_dir,
 334             deterministic_synthesis=deterministic_synthesis,
 335         )
 336
 337
 338 def one_epoch(model, quiz_machine, local_device=main_device):
 339     model.to(local_device).train()
 340
 341     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 342
 343     nb_train_samples, acc_train_loss = 0, 0.0
 344
 345     for input in quiz_machine.batches(model, split="train"):
 346         input = input.to(local_device)
 347
 348         if nb_train_samples % args.batch_size == 0:
 349             optimizer.zero_grad()
 350
 351         output = model(mygpt.BracketedSequence(input)).x
 352         loss = F.cross_entropy(output.transpose(1, 2), input)
 353         acc_train_loss += loss.item() * input.size(0)
 354
 355         nb_train_samples += input.size(0)
 356
 357         loss.backward()
 358
 359         if nb_train_samples % args.batch_size == 0:
 360             optimizer.step()
 361
 362     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 363
 364     log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
 365
 366     run_tests(model, quiz_machine, deterministic_synthesis=False)
 367
 368     model.to(main_device)
 369
 370
 371 ######################################################################
 372
 373 # This is the key routine that decides what generated quizzes to keep
 374
 375
 376 # token_logprobas are NxMxT where M is the number of models
 377
 378
 379 def compute_valid_quizzes_(token_logprobas):
 380     warnings.warn("validation with uniform constraints", RuntimeWarning)
 381     l = token_logprobas.min(dim=-1).values.sort(dim=-1).values
 382     return (l[:, 0] < math.log(0.1)) & (l[:, 1] > math.log(0.5))
 383
 384
 385 def compute_valid_quizzes(token_logprobas):
 386     l = token_logprobas.sum(dim=-1).sort(dim=-1).values
 387     return (l[:, 0] < math.log(args.proba_not_understands)) & (
 388         l[:, 1] > math.log(args.proba_understands)
 389     )
 390
 391
 392 def extract_valid_quizzes_and_logprobas(recorded):
 393     validated_quizzes, validated_logprobas = [], []
 394     for quizzes, token_logprobas in recorded:
 395         validated_indices = compute_valid_quizzes(token_logprobas)
 396         validated_quizzes.append(quizzes[validated_indices])
 397         validated_logprobas.append(token_logprobas[validated_indices])
 398
 399     if len(validated_quizzes) > 0:
 400         return torch.cat(validated_quizzes, dim=0), torch.cat(
 401             validated_logprobas, dim=0
 402         )
 403     else:
 404         return None, None
 405
 406
 407 ######################################################################
 408
 409
 410 def create_c_quizzes(models, quiz_machine, nb_for_train=1000, nb_for_test=100):
 411     nb_to_create = nb_for_train + nb_for_test
 412
 413     recorded_quizzes_logprobas = []
 414
 415     nb_validated = 0
 416
 417     while nb_validated < nb_to_create:
 418         model_for_generation = models[torch.randint(len(models), (1,))]
 419
 420         c_quizzes = quiz_machine.generate_quizzes(
 421             nb_to_create,
 422             model_for_generation=model_for_generation,
 423             temperature=args.generation_temperature,
 424         )
 425
 426         c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 427
 428         if c_quizzes.size(0) > 0:
 429             token_logproba = quiz_machine.solution_token_logprobas(models, c_quizzes)
 430             recorded_quizzes_logprobas.append((c_quizzes, token_logproba))
 431
 432             (
 433                 validated_quizzes,
 434                 validated_logprobas,
 435             ) = extract_valid_quizzes_and_logprobas(recorded_quizzes_logprobas)
 436
 437             if validated_quizzes is not None:
 438                 nb_validated = validated_quizzes.size(0)
 439
 440         log_string(
 441             f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 442         )
 443
 444     # store the new c_quizzes which have been validated
 445
 446     quiz_machine.reverse_random_half_in_place(validated_quizzes)
 447     quiz_machine.store_c_quizzes(validated_quizzes[:nb_for_train], for_train=True)
 448     quiz_machine.store_c_quizzes(
 449         validated_quizzes[nb_for_train:nb_to_create], for_train=False
 450     )
 451
 452     ######################################################################
 453     # save images with their logprobas
 454
 455     vq = validated_quizzes[:72]
 456     vl = validated_logprobas[:72]
 457
 458     if vq.size(0) > 0:
 459         prefix = f"culture_c_quiz_{n_epoch:04d}"
 460         filename = os.path.join(args.result_dir, prefix + "_logp.pth")
 461         torch.save(vl, filename)
 462         # with open(file_name, "w") as logp_file:
 463         # for l in vl:
 464         # s = " ".join([str(x.item()) for x in l])
 465         # logp_file.write(s + "\n")
 466
 467         quiz_machine.save_quiz_illustrations(args.result_dir, prefix, vq)
 468
 469
 470 ######################################################################
 471
 472 models = []
 473
 474 for k in range(args.nb_gpts):
 475     log_string(f"creating model {k} and its w_quizzes")
 476     model = mygpt.MyGPT(
 477         vocabulary_size=vocabulary_size,
 478         dim_model=args.dim_model,
 479         dim_keys=args.dim_keys,
 480         dim_hidden=args.dim_hidden,
 481         nb_heads=args.nb_heads,
 482         nb_blocks=args.nb_blocks,
 483         causal=True,
 484         dropout=args.dropout,
 485     ).to(main_device)
 486
 487     model.main_test_accuracy = 0.0
 488     model.id = k
 489
 490     model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
 491     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 492     model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
 493     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 494
 495     models.append(model)
 496
 497 ######################################################################
 498
 499 if args.resume:
 500     try:
 501         for model in models:
 502             filename = f"gpt_{model.id:03d}.pth"
 503
 504             try:
 505                 d = torch.load(os.path.join(args.result_dir, filename))
 506                 model.load_state_dict(d[0])
 507                 model.main_test_accuracy = d[1]
 508                 log_string(f"successfully loaded {filename}")
 509             except FileNotFoundError:
 510                 log_string(f"cannot find {filename}")
 511                 pass
 512
 513         try:
 514             filename = "c_quizzes.pth"
 515             quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
 516             log_string(f"successfully loaded {filename}")
 517         except FileNotFoundError:
 518             log_string(f"cannot find {filename}")
 519             pass
 520
 521     except:
 522         log_string(f"error when loading {filename}.")
 523         exit(1)
 524
 525 ######################################################################
 526
 527 nb_parameters = sum(p.numel() for p in models[0].parameters())
 528 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 529
 530 ######################################################################
 531
 532 # Compute the entropy of the training tokens
 533
 534 token_count = 0
 535 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 536     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 537         (0, 1)
 538     )
 539 token_probas = token_count / token_count.sum()
 540 entropy = -torch.xlogy(token_probas, token_probas).sum()
 541 train_set_perplexity = math.exp(entropy)
 542
 543 ######################################################################
 544 # A bit of paranoia never hurts
 545
 546 if args.max_percents_of_test_in_train >= 0:
 547
 548     def subsets_as_tuples(batches, cs):
 549         s = set()
 550         for batch in batches:
 551             for x in batch:
 552                 s.add(tuple([v.item() for v in x]))
 553                 if len(s) == cs:
 554                     yield s
 555                     s = set()
 556         yield s
 557
 558     nb_test, nb_in_train = 0, 0
 559     for test_subset in subsets_as_tuples(
 560         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 561     ):
 562         in_train = set()
 563         for train_subset in subsets_as_tuples(
 564             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 565         ):
 566             in_train.update(test_subset.intersection(train_subset))
 567         nb_in_train += len(in_train)
 568         nb_test += len(test_subset)
 569
 570     log_string(
 571         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 572     )
 573
 574     assert (
 575         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 576     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 577
 578 ######################################################################
 579
 580 if args.nb_new_c_quizzes_for_train is None:
 581     args.nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 582
 583 if args.nb_new_c_quizzes_for_test is None:
 584     args.nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 585
 586 log_string(
 587     f"nb_new_c_quizzes_for_train {args.nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {args.nb_new_c_quizzes_for_test}"
 588 )
 589
 590 ######################################################################
 591
 592 if args.dirty_debug:
 593     args.accuracy_to_make_c_quizzes = 0.0
 594     args.nb_gpts = 2
 595     args.nb_new_c_quizzes_for_train = 100
 596     args.nb_new_c_quizzes_for_test = 10
 597
 598
 599 ######################################################################
 600
 601 for n_epoch in range(args.nb_epochs):
 602     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 603
 604     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 605     log_string(f"current_test_accuracies {cta}")
 606
 607     ##################################################
 608     # If all the models are good enough, generate new quizzes and
 609     # re-compute the test errors
 610
 611     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 612         create_c_quizzes(
 613             models,
 614             quiz_machine,
 615             nb_for_train=args.nb_new_c_quizzes_for_train,
 616             nb_for_test=args.nb_new_c_quizzes_for_test,
 617         )
 618
 619         filename = "c_quizzes.pth"
 620         quiz_machine.save_c_quizzes(os.path.join(args.result_dir, filename))
 621         log_string(f"wrote {filename}")
 622
 623         # Force one epoch of training
 624         for model in models:
 625             model.main_test_accuracy = 0.0
 626
 627     ##################################################
 628     # Select, improve, and eval the worst model
 629
 630     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 631
 632     weakest_models = ranked_models[: len(gpus)]
 633
 634     threads = []
 635
 636     for gpu, model in zip(gpus, weakest_models):
 637         log_string(f"training model {model.id}")
 638
 639         t = threading.Thread(
 640             target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
 641         )
 642
 643         threads.append(t)
 644
 645         t.start()
 646
 647     for t in threads:
 648         t.join()
 649
 650     # Save the models to disk
 651
 652     for model in weakest_models:
 653         filename = f"gpt_{model.id:03d}.pth"
 654         torch.save(
 655             (model.state_dict(), model.main_test_accuracy),
 656             os.path.join(args.result_dir, filename),
 657         )
 658         log_string(f"wrote {filename}")
 659
 660     # Renew the training samples
 661
 662     for model in weakest_models:
 663         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 664
 665
 666 ######################################################################