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 from problem import MultiThreadProblem
  19
  20 # world quizzes vs. culture quizzes
  21
  22 ######################################################################
  23
  24 if torch.cuda.is_available():
  25     device = torch.device("cuda")
  26     torch.backends.cuda.matmul.allow_tf32 = True
  27 else:
  28     device = torch.device("cpu")
  29
  30 ######################################################################
  31
  32 parser = argparse.ArgumentParser(
  33     description="An implementation of GPT with cache.",
  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("--multi_thread_problem", action="store_true", default=False)
  82
  83 parser.add_argument("--nb_gpts", type=int, default=5)
  84
  85 parser.add_argument("--min_to_validate", type=int, default=None)
  86
  87 parser.add_argument("--max_to_validate", type=int, default=None)
  88
  89 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  90
  91 parser.add_argument("--generation_temperature", type=float, default=2.0)
  92
  93 parser.add_argument("--deterministic_validation", action="store_true", default=False)
  94
  95 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
  96
  97 parser.add_argument("--dirty_debug", action="store_true", default=False)
  98
  99 ######################################################################
 100
 101 parser.add_argument("--sky_height", type=int, default=6)
 102
 103 parser.add_argument("--sky_width", type=int, default=8)
 104
 105 parser.add_argument("--sky_nb_birds", type=int, default=3)
 106
 107 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 108
 109 parser.add_argument("--sky_speed", type=int, default=3)
 110
 111 ######################################################################
 112
 113 args = parser.parse_args()
 114
 115 if args.min_to_validate is None:
 116     args.min_to_validate = args.nb_gpts - 1
 117
 118 if args.max_to_validate is None:
 119     args.max_to_validate = args.nb_gpts - 1
 120
 121 if args.result_dir is None:
 122     args.result_dir = f"results_culture"
 123
 124 ######################################################################
 125
 126 default_args = {
 127     "model": "37M",
 128     "batch_size": 100,
 129     "nb_train_samples": 100000,
 130     "nb_test_samples": 10000,
 131 }
 132
 133 for k, v in default_args.items():
 134     if getattr(args, k) is None:
 135         setattr(args, k, v)
 136
 137 ######################################################################
 138
 139 default_model_args = {
 140     "17K": {
 141         "dim_model": 32,
 142         "dim_keys": 32,
 143         "dim_hidden": 32,
 144         "nb_heads": 2,
 145         "nb_blocks": 2,
 146     },
 147     "4M": {
 148         "dim_model": 256,
 149         "dim_keys": 32,
 150         "dim_hidden": 1024,
 151         "nb_heads": 4,
 152         "nb_blocks": 6,
 153     },
 154     "37M": {
 155         "dim_model": 512,
 156         "dim_keys": 64,
 157         "dim_hidden": 2048,
 158         "nb_heads": 8,
 159         "nb_blocks": 12,
 160     },
 161     "122M": {
 162         "dim_model": 768,
 163         "dim_keys": 64,
 164         "dim_hidden": 2048,
 165         "nb_heads": 8,
 166         "nb_blocks": 24,
 167     },
 168     "352M": {
 169         "dim_model": 1024,
 170         "dim_keys": 64,
 171         "dim_hidden": 2048,
 172         "nb_heads": 8,
 173         "nb_blocks": 48,
 174     },
 175 }
 176
 177 if args.model in default_model_args:
 178     for k, v in default_model_args[args.model].items():
 179         if getattr(args, k) is None:
 180             setattr(args, k, v)
 181 else:
 182     raise ValueError(f"Unknown model {args.model}")
 183
 184 ######################################################################
 185
 186 try:
 187     os.mkdir(args.result_dir)
 188 except FileExistsError:
 189     print(f"result directory {args.result_dir} already exists")
 190     exit(1)
 191
 192 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 193
 194 if args.seed >= 0:
 195     # torch.backends.cudnn.deterministic = True
 196     # torch.backends.cudnn.benchmark = False
 197     # torch.use_deterministic_algorithms(True)
 198     torch.manual_seed(args.seed)
 199     if torch.cuda.is_available():
 200         torch.cuda.manual_seed_all(args.seed)
 201
 202 ######################################################################
 203
 204
 205 def log_string(s):
 206     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 207
 208     if log_file is not None:
 209         log_file.write(t + s + "\n")
 210         log_file.flush()
 211
 212     print(t + s)
 213     sys.stdout.flush()
 214
 215
 216 log_string(f"argv {' '.join(sys.argv)}")
 217
 218 for n in vars(args):
 219     log_string(f"args.{n} {getattr(args, n)}")
 220
 221
 222 ######################################################################
 223
 224 if args.dirty_debug:
 225     args.nb_train_samples = 2500
 226     args.nb_test_samples = 100
 227
 228 if args.physical_batch_size is None:
 229     args.physical_batch_size = args.batch_size
 230 else:
 231     assert args.batch_size % args.physical_batch_size == 0
 232
 233 assert args.nb_train_samples % args.batch_size == 0
 234 assert args.nb_test_samples % args.batch_size == 0
 235
 236 if args.problem == "sky":
 237     problem = sky.Sky(
 238         height=args.sky_height,
 239         width=args.sky_width,
 240         nb_birds=args.sky_nb_birds,
 241         nb_iterations=args.sky_nb_iterations,
 242         speed=args.sky_speed,
 243     )
 244     back_accuracy = False
 245 elif args.problem == "grids":
 246     problem = grids.Grids(device=device)
 247     back_accuracy = True
 248 else:
 249     raise ValueError
 250
 251 if args.multi_thread_problem:
 252     problem = MultiThreadProblem(problem, args.nb_train_samples, chunk_size=1000)
 253
 254 quiz_machine = quiz_machine.QuizMachine(
 255     problem=problem,
 256     nb_train_samples=args.nb_train_samples,
 257     nb_test_samples=args.nb_test_samples,
 258     back_accuracy=back_accuracy,
 259     batch_size=args.physical_batch_size,
 260     result_dir=args.result_dir,
 261     logger=log_string,
 262     device=device,
 263 )
 264
 265 ######################################################################
 266
 267 log_string(f"device {device}")
 268
 269 vocabulary_size = quiz_machine.vocabulary_size()
 270
 271 log_string(f"vocabulary_size {vocabulary_size}")
 272
 273 ######################################################################
 274
 275 # Compute the entropy of the training tokens
 276
 277 token_count = 0
 278 for input in quiz_machine.batches(split="train", desc="train-entropy"):
 279     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 280         (0, 1)
 281     )
 282 token_probas = token_count / token_count.sum()
 283 entropy = -torch.xlogy(token_probas, token_probas).sum()
 284 train_set_perplexity = math.exp(entropy)
 285
 286 ######################################################################
 287 # A bit of paranoia never hurts
 288
 289 if args.max_percents_of_test_in_train >= 0:
 290
 291     def subsets_as_tuples(batches, cs):
 292         s = set()
 293         for batch in batches:
 294             for x in batch:
 295                 s.add(tuple([v.item() for v in x]))
 296                 if len(s) == cs:
 297                     yield s
 298                     s = set()
 299         yield s
 300
 301     nb_test, nb_in_train = 0, 0
 302     for test_subset in subsets_as_tuples(
 303         quiz_machine.batches(split="test", desc="test-check"), 25000
 304     ):
 305         in_train = set()
 306         for train_subset in subsets_as_tuples(
 307             quiz_machine.batches(split="train", desc="train-check"), 25000
 308         ):
 309             in_train.update(test_subset.intersection(train_subset))
 310         nb_in_train += len(in_train)
 311         nb_test += len(test_subset)
 312
 313     log_string(
 314         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 315     )
 316
 317     assert (
 318         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 319     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 320
 321 ##############################
 322
 323
 324 def one_epoch(model, quiz_machine):
 325     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 326
 327     model.train()
 328
 329     nb_train_samples, acc_train_loss = 0, 0.0
 330
 331     for input in quiz_machine.batches(split="train"):
 332         input = input.to(device)
 333
 334         if nb_train_samples % args.batch_size == 0:
 335             optimizer.zero_grad()
 336
 337         output = model(mygpt.BracketedSequence(input)).x
 338         loss = F.cross_entropy(output.transpose(1, 2), input)
 339         acc_train_loss += loss.item() * input.size(0)
 340
 341         nb_train_samples += input.size(0)
 342
 343         loss.backward()
 344
 345         if nb_train_samples % args.batch_size == 0:
 346             optimizer.step()
 347
 348     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 349
 350     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 351
 352
 353 ######################################################################
 354
 355
 356 def run_tests(model, quiz_machine, deterministic_synthesis):
 357     with torch.autograd.no_grad():
 358         model.eval()
 359
 360         nb_test_samples, acc_test_loss = 0, 0.0
 361         nb_samples_accumulated = 0
 362
 363         for input in quiz_machine.batches(split="test"):
 364             input = input.to(device)
 365
 366             bs = model(mygpt.BracketedSequence(input))
 367             output = bs.x
 368
 369             loss = F.cross_entropy(output.transpose(1, 2), input)
 370
 371             acc_test_loss += loss.item() * input.size(0)
 372
 373             nb_test_samples += input.size(0)
 374
 375         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 376
 377         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 378
 379         model.main_test_accuracy = quiz_machine.produce_results(
 380             n_epoch=n_epoch,
 381             model=model,
 382             result_dir=args.result_dir,
 383             deterministic_synthesis=deterministic_synthesis,
 384         )
 385
 386
 387 ######################################################################
 388
 389
 390 def valid_c_quizzes(recorded, criteria):
 391     result = [q[criteria(c)] for q, c in recorded]
 392     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 393
 394
 395 ######################################################################
 396
 397
 398 def create_c_quizzes(
 399     models,
 400     quiz_machine,
 401     nb_for_train=1000,
 402     nb_for_test=100,
 403 ):
 404     quizzes_and_nb_correct_records = []
 405
 406     nb_to_create = nb_for_train + nb_for_test
 407
 408     # ------------------------------------------------------------
 409
 410     standard_validity = lambda nb_correct: torch.logical_and(
 411         nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
 412     )
 413
 414     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 415
 416     with open(file_name, "w") as logp_file:
 417         while (
 418             valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
 419             < nb_to_create
 420         ):
 421             # Select a model at random to generate the new quizzes
 422
 423             model_for_generation = models[torch.randint(len(models), (1,))]
 424
 425             c_quizzes = quiz_machine.generate_quizzes(
 426                 nb_to_create,
 427                 model_for_generation=model_for_generation,
 428                 temperature=args.generation_temperature,
 429             )
 430
 431             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 432
 433             if c_quizzes.size(0) > 0:
 434                 nb_correct, seq_logproba = quiz_machine.compute_correctness(
 435                     c_quizzes,
 436                     models,
 437                     bidirectional_validation=args.bidirectional_validation,
 438                     deterministic_validation=args.deterministic_validation,
 439                 )
 440
 441                 for n, l in zip(nb_correct, seq_logproba):
 442                     s = " ".join([str(x.item()) for x in l])
 443                     logp_file.write(f"{n} {s}\n")
 444
 445                 if args.dirty_debug:
 446                     nb_correct = torch.randint(
 447                         len(models) + 1, nb_correct.size(), device=c_quizzes.device
 448                     )
 449
 450                 quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 451
 452             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 453             nv = " ".join([str(x.item()) for x in nv])
 454
 455             nb_validated = valid_c_quizzes(
 456                 quizzes_and_nb_correct_records, standard_validity
 457             ).size(0)
 458
 459             log_string(
 460                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 461             )
 462
 463     # store the new c_quizzes which have been validated
 464
 465     new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
 466
 467     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 468
 469     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 470     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 471
 472     # save a bunch of images to investigate what quizzes with a
 473     # certain nb of correct predictions look like
 474
 475     for n in range(len(models) + 1):
 476         s = (
 477             "_validated"
 478             if n >= args.min_to_validate and n <= args.max_to_validate
 479             else ""
 480         )
 481
 482         q = valid_c_quizzes(
 483             quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
 484         )[:72]
 485
 486         quiz_machine.reverse_random_half_in_place(q)
 487
 488         if q.size(0) > 0:
 489             quiz_machine.save_quizzes(
 490                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 491             )
 492
 493
 494 ######################################################################
 495
 496 models = []
 497
 498 for k in range(args.nb_gpts):
 499     model = mygpt.MyGPT(
 500         vocabulary_size=vocabulary_size,
 501         dim_model=args.dim_model,
 502         dim_keys=args.dim_keys,
 503         dim_hidden=args.dim_hidden,
 504         nb_heads=args.nb_heads,
 505         nb_blocks=args.nb_blocks,
 506         causal=True,
 507         dropout=args.dropout,
 508     ).to(device)
 509
 510     model.main_test_accuracy = 0.0
 511     model.id = k
 512
 513     models.append(model)
 514
 515
 516 nb_parameters = sum(p.numel() for p in models[0].parameters())
 517 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 518
 519 ######################################################################
 520
 521 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 522 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 523
 524 log_string(
 525     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}"
 526 )
 527
 528 ######################################################################
 529
 530 if args.dirty_debug:
 531     args.accuracy_to_make_c_quizzes = 0.0
 532     args.nb_gpts = 2
 533     nb_new_c_quizzes_for_train = 100
 534     nb_new_c_quizzes_for_test = 10
 535
 536 ######################################################################
 537
 538 for n_epoch in range(args.nb_epochs):
 539     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 540
 541     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 542     log_string(f"current_test_accuracies {cta}")
 543
 544     ##################################################
 545     # Select, improve, and eval the worst model
 546
 547     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 548
 549     log_string(
 550         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 551     )
 552
 553     one_epoch(weakest_model, quiz_machine)
 554
 555     log_string(
 556         f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 557     )
 558
 559     run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
 560
 561     log_string(
 562         f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 563     )
 564
 565     ##################################################
 566     # Replace a fraction of the w_quizzes with fresh ones
 567
 568     quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 569
 570     ##################################################
 571     # If all the models are good enough, generate new quizzes and
 572     # re-compute the test errors
 573
 574     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 575         create_c_quizzes(
 576             models,
 577             quiz_machine,
 578             nb_for_train=nb_new_c_quizzes_for_train,
 579             nb_for_test=nb_new_c_quizzes_for_test,
 580         )
 581
 582         for model in models:
 583             run_tests(model, quiz_machine, deterministic_synthesis=False)
 584
 585
 586 ######################################################################