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=0.9)
  92
  93 parser.add_argument("--generation_temperature", type=float, default=2.0)
  94
  95 parser.add_argument("--dirty_debug", action="store_true", default=False)
  96
  97 ######################################################################
  98
  99 parser.add_argument("--sky_height", type=int, default=6)
 100
 101 parser.add_argument("--sky_width", type=int, default=8)
 102
 103 parser.add_argument("--sky_nb_birds", type=int, default=3)
 104
 105 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 106
 107 parser.add_argument("--sky_speed", type=int, default=3)
 108
 109 ######################################################################
 110
 111 args = parser.parse_args()
 112
 113 if args.min_to_validate is None:
 114     args.min_to_validate = args.nb_gpts - 1
 115
 116 if args.max_to_validate is None:
 117     args.max_to_validate = args.nb_gpts - 1
 118
 119 if args.result_dir is None:
 120     args.result_dir = f"results_culture"
 121
 122 ######################################################################
 123
 124 default_args = {
 125     "model": "37M",
 126     "batch_size": 25,
 127     "nb_train_samples": 100000,
 128     "nb_test_samples": 10000,
 129 }
 130
 131 for k, v in default_args.items():
 132     if getattr(args, k) is None:
 133         setattr(args, k, v)
 134
 135 ######################################################################
 136
 137 default_model_args = {
 138     "17K": {
 139         "dim_model": 32,
 140         "dim_keys": 32,
 141         "dim_hidden": 32,
 142         "nb_heads": 2,
 143         "nb_blocks": 2,
 144     },
 145     "4M": {
 146         "dim_model": 256,
 147         "dim_keys": 32,
 148         "dim_hidden": 1024,
 149         "nb_heads": 4,
 150         "nb_blocks": 6,
 151     },
 152     "37M": {
 153         "dim_model": 512,
 154         "dim_keys": 64,
 155         "dim_hidden": 2048,
 156         "nb_heads": 8,
 157         "nb_blocks": 12,
 158     },
 159     "122M": {
 160         "dim_model": 768,
 161         "dim_keys": 64,
 162         "dim_hidden": 2048,
 163         "nb_heads": 8,
 164         "nb_blocks": 24,
 165     },
 166     "352M": {
 167         "dim_model": 1024,
 168         "dim_keys": 64,
 169         "dim_hidden": 2048,
 170         "nb_heads": 8,
 171         "nb_blocks": 48,
 172     },
 173 }
 174
 175 if args.model in default_model_args:
 176     for k, v in default_model_args[args.model].items():
 177         if getattr(args, k) is None:
 178             setattr(args, k, v)
 179 else:
 180     raise ValueError(f"Unknown model {args.model}")
 181
 182 ######################################################################
 183
 184 try:
 185     os.mkdir(args.result_dir)
 186 except FileExistsError:
 187     print(f"result directory {args.result_dir} already exists")
 188     exit(1)
 189
 190 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 191
 192 if args.seed >= 0:
 193     # torch.backends.cudnn.deterministic = True
 194     # torch.backends.cudnn.benchmark = False
 195     # torch.use_deterministic_algorithms(True)
 196     torch.manual_seed(args.seed)
 197     if torch.cuda.is_available():
 198         torch.cuda.manual_seed_all(args.seed)
 199
 200 ######################################################################
 201
 202
 203 def log_string(s):
 204     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 205
 206     if log_file is not None:
 207         log_file.write(t + s + "\n")
 208         log_file.flush()
 209
 210     print(t + s)
 211     sys.stdout.flush()
 212
 213
 214 log_string(f"argv {' '.join(sys.argv)}")
 215
 216 for n in vars(args):
 217     log_string(f"args.{n} {getattr(args, n)}")
 218
 219
 220 ######################################################################
 221
 222 if args.dirty_debug:
 223     args.nb_train_samples = 2500
 224     args.nb_test_samples = 100
 225
 226 if args.physical_batch_size is None:
 227     args.physical_batch_size = args.batch_size
 228 else:
 229     assert args.batch_size % args.physical_batch_size == 0
 230
 231 assert args.nb_train_samples % args.batch_size == 0
 232 assert args.nb_test_samples % args.batch_size == 0
 233
 234 if args.problem == "sky":
 235     problem = sky.Sky(
 236         height=args.sky_height,
 237         width=args.sky_width,
 238         nb_birds=args.sky_nb_birds,
 239         nb_iterations=args.sky_nb_iterations,
 240         speed=args.sky_speed,
 241         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 242         chunk_size=100,
 243         nb_threads=args.nb_threads,
 244     )
 245     back_accuracy = False
 246 elif args.problem == "grids":
 247     problem = grids.Grids(
 248         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 249         chunk_size=100,
 250         nb_threads=args.nb_threads,
 251     )
 252     back_accuracy = True
 253 else:
 254     raise ValueError
 255
 256 problem.save_some_examples(args.result_dir)
 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
 280 ######################################################################
 281
 282
 283 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
 284     if local_device is None:
 285         local_device = device
 286
 287     with torch.autograd.no_grad():
 288         model.eval().to(local_device)
 289
 290         nb_test_samples, acc_test_loss = 0, 0.0
 291         nb_samples_accumulated = 0
 292
 293         for input in quiz_machine.batches(model, split="test"):
 294             input = input.to(local_device)
 295
 296             bs = model(mygpt.BracketedSequence(input))
 297             output = bs.x
 298
 299             loss = F.cross_entropy(output.transpose(1, 2), input)
 300
 301             acc_test_loss += loss.item() * input.size(0)
 302
 303             nb_test_samples += input.size(0)
 304
 305         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 306
 307         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 308
 309         model.main_test_accuracy = quiz_machine.produce_results(
 310             n_epoch=n_epoch,
 311             model=model,
 312             result_dir=args.result_dir,
 313             deterministic_synthesis=deterministic_synthesis,
 314         )
 315
 316
 317 def one_epoch(model, quiz_machine, local_device=None):
 318     if local_device is None:
 319         local_device = device
 320
 321     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 322
 323     model.to(local_device).train()
 324
 325     nb_train_samples, acc_train_loss = 0, 0.0
 326
 327     for input in quiz_machine.batches(model, split="train"):
 328         input = input.to(local_device)
 329
 330         if nb_train_samples % args.batch_size == 0:
 331             optimizer.zero_grad()
 332
 333         output = model(mygpt.BracketedSequence(input)).x
 334         loss = F.cross_entropy(output.transpose(1, 2), input)
 335         acc_train_loss += loss.item() * input.size(0)
 336
 337         nb_train_samples += input.size(0)
 338
 339         loss.backward()
 340
 341         if nb_train_samples % args.batch_size == 0:
 342             optimizer.step()
 343
 344     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 345
 346     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 347
 348     run_tests(model, quiz_machine, deterministic_synthesis=False)
 349
 350     model.TRAINING_LOCK.release()
 351
 352
 353 ######################################################################
 354
 355
 356 def standard_validity(logproba):
 357     l = logproba.sort(dim=-1).values
 358     return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
 359     # warnings.warn("TEST!!!", RuntimeWarning)
 360     # print(l.exp())
 361     # return (l[:, 0] < math.log(0.99))
 362
 363
 364 def valid_c_quizzes(recorded, criteria):
 365     result = [q[criteria(lp)] for q, lp in recorded]
 366     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 367
 368
 369 ######################################################################
 370
 371
 372 def create_c_quizzes(
 373     models,
 374     quiz_machine,
 375     nb_for_train=1000,
 376     nb_for_test=100,
 377 ):
 378     quizzes_and_logproba_records = []
 379
 380     nb_to_create = nb_for_train + nb_for_test
 381
 382     # ------------------------------------------------------------
 383
 384     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 385
 386     with open(file_name, "w") as logp_file:
 387         while (
 388             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 389             < nb_to_create
 390         ):
 391             # Select a model at random to generate the new quizzes
 392
 393             model_for_generation = models[torch.randint(len(models), (1,))]
 394
 395             c_quizzes = quiz_machine.generate_quizzes(
 396                 nb_to_create,
 397                 model_for_generation=model_for_generation,
 398                 temperature=args.generation_temperature,
 399             )
 400
 401             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 402
 403             if c_quizzes.size(0) > 0:
 404                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 405                 for l in logproba:
 406                     s = " ".join([str(x.item()) for x in l])
 407                     logp_file.write(s + "\n")
 408                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 409
 410             nb_validated = valid_c_quizzes(
 411                 quizzes_and_logproba_records, standard_validity
 412             ).size(0)
 413
 414             log_string(
 415                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 416             )
 417
 418     # store the new c_quizzes which have been validated
 419
 420     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 421
 422     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 423
 424     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 425     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 426
 427     # save a bunch of images to investigate what quizzes with a
 428     # certain nb of correct predictions look like
 429
 430     q = new_c_quizzes[:72]
 431
 432     if q.size(0) > 0:
 433         quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
 434
 435
 436 ######################################################################
 437
 438 models = []
 439
 440 for k in range(args.nb_gpts):
 441     log_string(f"creating model {k} and its w_quizzes")
 442     model = mygpt.MyGPT(
 443         vocabulary_size=vocabulary_size,
 444         dim_model=args.dim_model,
 445         dim_keys=args.dim_keys,
 446         dim_hidden=args.dim_hidden,
 447         nb_heads=args.nb_heads,
 448         nb_blocks=args.nb_blocks,
 449         causal=True,
 450         dropout=args.dropout,
 451     ).to(device)
 452
 453     model.main_test_accuracy = 0.0
 454     model.id = k
 455     model.TRAINING_LOCK = threading.Lock()
 456
 457     model.train_w_quizzes = quiz_machine.generate_token_sequences(
 458         args.nb_train_samples
 459     ).to(device)
 460     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 461     model.test_w_quizzes = quiz_machine.generate_token_sequences(
 462         args.nb_test_samples
 463     ).to(device)
 464     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 465
 466     models.append(model)
 467
 468
 469 nb_parameters = sum(p.numel() for p in models[0].parameters())
 470 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 471
 472 ######################################################################
 473
 474 # Compute the entropy of the training tokens
 475
 476 token_count = 0
 477 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 478     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 479         (0, 1)
 480     )
 481 token_probas = token_count / token_count.sum()
 482 entropy = -torch.xlogy(token_probas, token_probas).sum()
 483 train_set_perplexity = math.exp(entropy)
 484
 485 ######################################################################
 486 # A bit of paranoia never hurts
 487
 488 if args.max_percents_of_test_in_train >= 0:
 489
 490     def subsets_as_tuples(batches, cs):
 491         s = set()
 492         for batch in batches:
 493             for x in batch:
 494                 s.add(tuple([v.item() for v in x]))
 495                 if len(s) == cs:
 496                     yield s
 497                     s = set()
 498         yield s
 499
 500     nb_test, nb_in_train = 0, 0
 501     for test_subset in subsets_as_tuples(
 502         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 503     ):
 504         in_train = set()
 505         for train_subset in subsets_as_tuples(
 506             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 507         ):
 508             in_train.update(test_subset.intersection(train_subset))
 509         nb_in_train += len(in_train)
 510         nb_test += len(test_subset)
 511
 512     log_string(
 513         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 514     )
 515
 516     assert (
 517         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 518     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 519
 520 ######################################################################
 521
 522 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 523 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 524
 525 log_string(
 526     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}"
 527 )
 528
 529 ######################################################################
 530
 531 if args.dirty_debug:
 532     args.accuracy_to_make_c_quizzes = 0.0
 533     args.nb_gpts = 2
 534     nb_new_c_quizzes_for_train = 100
 535     nb_new_c_quizzes_for_test = 10
 536
 537     def standard_validity(logproba):
 538         l = logproba.sort(dim=-1).values
 539         return l[:, 0] < math.log(0.99)
 540
 541
 542 ######################################################################
 543
 544 for n_epoch in range(args.nb_epochs):
 545     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 546
 547     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 548     log_string(f"current_test_accuracies {cta}")
 549
 550     ##################################################
 551     # Select, improve, and eval the worst models
 552
 553     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 554
 555     weakest_models = ranked_models[: args.nb_gpus]
 556
 557     for gpu_id, model in enumerate(weakest_models):
 558         model.TRAINING_LOCK.acquire()
 559
 560         log_string(
 561             f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 562         )
 563
 564         threading.Thread(
 565             target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
 566         ).start()
 567
 568     for model in weakest_models:
 569         model.TRAINING_LOCK.acquire()
 570         model.TRAINING_LOCK.release()
 571
 572     ##################################################
 573     # Renew the train sets
 574
 575     log_string(
 576         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 577     )
 578
 579     for model in weakest_models:
 580         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 581
 582     ##################################################
 583     # If all the models are good enough, generate new quizzes and
 584     # re-compute the test errors
 585
 586     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 587         create_c_quizzes(
 588             models,
 589             quiz_machine,
 590             nb_for_train=nb_new_c_quizzes_for_train,
 591             nb_for_test=nb_new_c_quizzes_for_test,
 592         )
 593
 594 ######################################################################