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("--learning_rate", type=float, default=5e-4)
  52
  53 ########################################
  54
  55 parser.add_argument("--model", type=str, default=None)
  56
  57 parser.add_argument("--dim_model", type=int, default=None)
  58
  59 parser.add_argument("--dim_keys", type=int, default=None)
  60
  61 parser.add_argument("--dim_hidden", type=int, default=None)
  62
  63 parser.add_argument("--nb_heads", type=int, default=None)
  64
  65 parser.add_argument("--nb_blocks", type=int, default=None)
  66
  67 parser.add_argument("--dropout", type=float, default=0.1)
  68
  69 ########################################
  70
  71 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  72
  73 parser.add_argument("--problem", type=str, default="grids")
  74
  75 parser.add_argument("--nb_threads", type=int, default=1)
  76
  77 parser.add_argument("--gpus", type=str, default="all")
  78
  79 parser.add_argument("--nb_gpts", type=int, default=5)
  80
  81 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  82
  83 parser.add_argument("--proba_understands", type=float, default=0.99)
  84
  85 parser.add_argument("--proba_not_understands", type=float, default=0.5)
  86
  87 parser.add_argument("--generation_temperature", type=float, default=2.0)
  88
  89 parser.add_argument("--dirty_debug", action="store_true", default=False)
  90
  91 ######################################################################
  92
  93 grids_tasks = ", ".join(
  94     [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
  95 )
  96
  97 parser.add_argument(
  98     "--grids_tasks",
  99     type=str,
 100     default=None,
 101     help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
 102 )
 103
 104 ######################################################################
 105
 106 parser.add_argument("--sky_height", type=int, default=6)
 107
 108 parser.add_argument("--sky_width", type=int, default=8)
 109
 110 parser.add_argument("--sky_nb_birds", type=int, default=3)
 111
 112 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 113
 114 parser.add_argument("--sky_speed", type=int, default=3)
 115
 116 ######################################################################
 117
 118 args = parser.parse_args()
 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 if args.resume:
 186     assert os.path.isdir(args.result_dir)
 187
 188 else:
 189     try:
 190         os.mkdir(args.result_dir)
 191     except FileExistsError:
 192         print(f"result directory {args.result_dir} already exists")
 193         exit(1)
 194
 195 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 196
 197 if args.seed >= 0:
 198     # torch.backends.cudnn.deterministic = True
 199     # torch.backends.cudnn.benchmark = False
 200     # torch.use_deterministic_algorithms(True)
 201     torch.manual_seed(args.seed)
 202     if torch.cuda.is_available():
 203         torch.cuda.manual_seed_all(args.seed)
 204
 205 ######################################################################
 206
 207
 208 def log_string(s):
 209     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 210
 211     if log_file is not None:
 212         log_file.write(t + s + "\n")
 213         log_file.flush()
 214
 215     print(t + s)
 216     sys.stdout.flush()
 217
 218
 219 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 220
 221 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py")
 222
 223 log_string(f"argv {' '.join(sys.argv)}")
 224
 225 for n in vars(args):
 226     log_string(f"args.{n} {getattr(args, n)}")
 227
 228
 229 ######################################################################
 230
 231 if args.gpus == "all":
 232     gpus_idx = range(torch.cuda.device_count())
 233 else:
 234     gpus_idx = [int(k) for k in args.gpus.split(",")]
 235
 236 gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
 237
 238 if torch.cuda.is_available():
 239     main_device = gpus[0]
 240 else:
 241     assert len(gpus) == 0
 242     main_device = torch.device("cpu")
 243
 244 if args.dirty_debug:
 245     args.nb_train_samples = 2500
 246     args.nb_test_samples = 100
 247
 248 if args.physical_batch_size is None:
 249     args.physical_batch_size = args.batch_size
 250 else:
 251     assert args.batch_size % args.physical_batch_size == 0
 252
 253 assert args.nb_train_samples % args.batch_size == 0
 254 assert args.nb_test_samples % args.batch_size == 0
 255
 256 if args.problem == "sky":
 257     problem = sky.Sky(
 258         height=args.sky_height,
 259         width=args.sky_width,
 260         nb_birds=args.sky_nb_birds,
 261         nb_iterations=args.sky_nb_iterations,
 262         speed=args.sky_speed,
 263         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 264         chunk_size=100,
 265         nb_threads=args.nb_threads,
 266     )
 267     back_accuracy = False
 268 elif args.problem == "grids":
 269     problem = grids.Grids(
 270         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 271         chunk_size=100,
 272         nb_threads=args.nb_threads,
 273         tasks=args.grids_tasks,
 274     )
 275     back_accuracy = True
 276 else:
 277     raise ValueError
 278
 279 problem.save_some_examples(args.result_dir)
 280
 281 quiz_machine = quiz_machine.QuizMachine(
 282     problem=problem,
 283     nb_train_samples=args.nb_train_samples,
 284     nb_test_samples=args.nb_test_samples,
 285     back_accuracy=back_accuracy,
 286     batch_size=args.physical_batch_size,
 287     result_dir=args.result_dir,
 288     logger=log_string,
 289     device=main_device,
 290 )
 291
 292 ######################################################################
 293
 294 log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
 295
 296 vocabulary_size = quiz_machine.vocabulary_size()
 297
 298 log_string(f"vocabulary_size {vocabulary_size}")
 299
 300 ######################################################################
 301
 302
 303 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
 304     with torch.autograd.no_grad():
 305         model.eval().to(local_device)
 306
 307         nb_test_samples, acc_test_loss = 0, 0.0
 308         nb_samples_accumulated = 0
 309
 310         for input in quiz_machine.batches(model, split="test"):
 311             input = input.to(local_device)
 312
 313             bs = model(mygpt.BracketedSequence(input))
 314             output = bs.x
 315
 316             loss = F.cross_entropy(output.transpose(1, 2), input)
 317
 318             acc_test_loss += loss.item() * input.size(0)
 319
 320             nb_test_samples += input.size(0)
 321
 322         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 323
 324         log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
 325
 326         model.main_test_accuracy = quiz_machine.produce_results(
 327             n_epoch=n_epoch,
 328             model=model,
 329             result_dir=args.result_dir,
 330             deterministic_synthesis=deterministic_synthesis,
 331         )
 332
 333
 334 def one_epoch(model, quiz_machine, local_device=main_device):
 335     model.to(local_device).train()
 336
 337     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 338
 339     nb_train_samples, acc_train_loss = 0, 0.0
 340
 341     for input in quiz_machine.batches(model, split="train"):
 342         input = input.to(local_device)
 343
 344         if nb_train_samples % args.batch_size == 0:
 345             optimizer.zero_grad()
 346
 347         output = model(mygpt.BracketedSequence(input)).x
 348         loss = F.cross_entropy(output.transpose(1, 2), input)
 349         acc_train_loss += loss.item() * input.size(0)
 350
 351         nb_train_samples += input.size(0)
 352
 353         loss.backward()
 354
 355         if nb_train_samples % args.batch_size == 0:
 356             optimizer.step()
 357
 358     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 359
 360     log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
 361
 362     run_tests(model, quiz_machine, deterministic_synthesis=False)
 363
 364     model.to(main_device)
 365
 366
 367 ######################################################################
 368
 369
 370 def standard_validity(logproba):
 371     l = logproba.sort(dim=-1).values
 372     return (l[:, 0] < math.log(args.proba_not_understands)) & (
 373         l[:, 1] > math.log(args.proba_understands)
 374     )
 375
 376
 377 def valid_quizzes_and_logprobas(recorded, criteria):
 378     validated_quizzes, validated_logprobas = [], []
 379     for q, lp in recorded:
 380         validated_indices = criteria(lp)
 381         validated_quizzes.append(q[validated_indices])
 382         validated_logprobas.append(lp[validated_indices])
 383
 384     if len(validated_quizzes) > 0:
 385         return torch.cat(validated_quizzes, dim=0), torch.cat(
 386             validated_logprobas, dim=0
 387         )
 388     else:
 389         return None, None
 390
 391
 392 ######################################################################
 393
 394
 395 def create_c_quizzes(models, quiz_machine, nb_for_train=1000, nb_for_test=100):
 396     nb_to_create = nb_for_train + nb_for_test
 397
 398     recorded_quizzes_logprobas = []
 399
 400     nb_validated = 0
 401
 402     while nb_validated < nb_to_create:
 403         model_for_generation = models[torch.randint(len(models), (1,))]
 404
 405         c_quizzes = quiz_machine.generate_quizzes(
 406             nb_to_create,
 407             model_for_generation=model_for_generation,
 408             temperature=args.generation_temperature,
 409         )
 410
 411         c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 412
 413         if c_quizzes.size(0) > 0:
 414             logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 415             recorded_quizzes_logprobas.append((c_quizzes, logproba))
 416
 417             validated_quizzes, validated_logprobas = valid_quizzes_and_logprobas(
 418                 recorded_quizzes_logprobas, standard_validity
 419             )
 420
 421             if validated_quizzes is not None:
 422                 nb_validated = validated_quizzes.size(0)
 423
 424         log_string(
 425             f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 426         )
 427
 428     # store the new c_quizzes which have been validated
 429
 430     quiz_machine.reverse_random_half_in_place(validated_quizzes)
 431     quiz_machine.store_c_quizzes(validated_quizzes[:nb_for_train], for_train=True)
 432     quiz_machine.store_c_quizzes(
 433         validated_quizzes[nb_for_train:nb_to_create], for_train=False
 434     )
 435
 436     ######################################################################
 437     # save the log probas
 438
 439     file_name = os.path.join(
 440         args.result_dir, f"culture_c_quiz_all_{n_epoch:04d}_logp.dat"
 441     )
 442
 443     with open(file_name, "w") as logp_file:
 444         for _, ll in recorded_quizzes_logprobas:
 445             for l in ll:
 446                 s = " ".join([str(x.item()) for x in l])
 447                 logp_file.write(s + "\n")
 448
 449     ######################################################################
 450     # save images with their logprobas
 451
 452     vq = validated_quizzes[:72]
 453     vl = validated_logprobas[:72]
 454
 455     if vq.size(0) > 0:
 456         prefix = f"culture_c_quiz_{n_epoch:04d}"
 457
 458         file_name = os.path.join(args.result_dir, prefix + "_logp.dat")
 459         with open(file_name, "w") as logp_file:
 460             for l in vl:
 461                 s = " ".join([str(x.item()) for x in l])
 462                 logp_file.write(s + "\n")
 463
 464         quiz_machine.save_quiz_illustrations(args.result_dir, prefix, vq)
 465
 466
 467 ######################################################################
 468
 469 models = []
 470
 471 for k in range(args.nb_gpts):
 472     log_string(f"creating model {k} and its w_quizzes")
 473     model = mygpt.MyGPT(
 474         vocabulary_size=vocabulary_size,
 475         dim_model=args.dim_model,
 476         dim_keys=args.dim_keys,
 477         dim_hidden=args.dim_hidden,
 478         nb_heads=args.nb_heads,
 479         nb_blocks=args.nb_blocks,
 480         causal=True,
 481         dropout=args.dropout,
 482     ).to(main_device)
 483
 484     model.main_test_accuracy = 0.0
 485     model.id = k
 486
 487     model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
 488     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 489     model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
 490     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 491
 492     models.append(model)
 493
 494 ######################################################################
 495
 496 if args.resume:
 497     try:
 498         for model in models:
 499             filename = f"gpt_{model.id:03d}.pth"
 500
 501             try:
 502                 d = torch.load(os.path.join(args.result_dir, filename))
 503                 model.load_state_dict(d[0])
 504                 model.main_test_accuracy = d[1]
 505                 log_string(f"successfully loaded {filename}")
 506             except FileNotFoundError:
 507                 log_string(f"cannot find {filename}")
 508                 pass
 509
 510         try:
 511             filename = "c_quizzes.pth"
 512             quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
 513             log_string(f"successfully loaded {filename}")
 514         except FileNotFoundError:
 515             log_string(f"cannot find {filename}")
 516             pass
 517
 518     except:
 519         log_string(f"error when loading {filename}.")
 520         exit(1)
 521
 522 ######################################################################
 523
 524 nb_parameters = sum(p.numel() for p in models[0].parameters())
 525 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 526
 527 ######################################################################
 528
 529 # Compute the entropy of the training tokens
 530
 531 token_count = 0
 532 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 533     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 534         (0, 1)
 535     )
 536 token_probas = token_count / token_count.sum()
 537 entropy = -torch.xlogy(token_probas, token_probas).sum()
 538 train_set_perplexity = math.exp(entropy)
 539
 540 ######################################################################
 541 # A bit of paranoia never hurts
 542
 543 if args.max_percents_of_test_in_train >= 0:
 544
 545     def subsets_as_tuples(batches, cs):
 546         s = set()
 547         for batch in batches:
 548             for x in batch:
 549                 s.add(tuple([v.item() for v in x]))
 550                 if len(s) == cs:
 551                     yield s
 552                     s = set()
 553         yield s
 554
 555     nb_test, nb_in_train = 0, 0
 556     for test_subset in subsets_as_tuples(
 557         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 558     ):
 559         in_train = set()
 560         for train_subset in subsets_as_tuples(
 561             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 562         ):
 563             in_train.update(test_subset.intersection(train_subset))
 564         nb_in_train += len(in_train)
 565         nb_test += len(test_subset)
 566
 567     log_string(
 568         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 569     )
 570
 571     assert (
 572         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 573     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 574
 575 ######################################################################
 576
 577 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 578 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 579
 580 log_string(
 581     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}"
 582 )
 583
 584 ######################################################################
 585
 586 if args.dirty_debug:
 587     args.accuracy_to_make_c_quizzes = 0.0
 588     args.nb_gpts = 2
 589     nb_new_c_quizzes_for_train = 100
 590     nb_new_c_quizzes_for_test = 10
 591
 592     def standard_validity(logproba):
 593         l = logproba.sort(dim=-1).values
 594         return l[:, 0] < math.log(0.5)
 595
 596
 597 ######################################################################
 598
 599 for n_epoch in range(args.nb_epochs):
 600     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 601
 602     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 603     log_string(f"current_test_accuracies {cta}")
 604
 605     ##################################################
 606     # Select, improve, and eval the worst model
 607
 608     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 609
 610     weakest_models = ranked_models[: len(gpus)]
 611
 612     threads = []
 613
 614     for gpu, model in zip(gpus, weakest_models):
 615         log_string(f"training model {model.id}")
 616
 617         t = threading.Thread(
 618             target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
 619         )
 620
 621         threads.append(t)
 622
 623         t.start()
 624
 625     for t in threads:
 626         t.join()
 627
 628     # Save the models to disk
 629
 630     for model in weakest_models:
 631         filename = f"gpt_{model.id:03d}.pth"
 632         torch.save(
 633             (model.state_dict(), model.main_test_accuracy),
 634             os.path.join(args.result_dir, filename),
 635         )
 636         log_string(f"wrote {filename}")
 637
 638     # Renew the training samples
 639
 640     for model in weakest_models:
 641         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 642
 643     ##################################################
 644     # If all the models are good enough, generate new quizzes and
 645     # re-compute the test errors
 646
 647     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 648         create_c_quizzes(
 649             models,
 650             quiz_machine,
 651             nb_for_train=nb_new_c_quizzes_for_train,
 652             nb_for_test=nb_new_c_quizzes_for_test,
 653         )
 654
 655         filename = "c_quizzes.pth"
 656         quiz_machine.save_c_quizzes(os.path.join(args.result_dir, filename))
 657         log_string(f"wrote {filename}")
 658
 659 ######################################################################