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 import mygpt
  16 import sky, grids, quiz_machine
  17
  18 # world quizzes vs. culture quizzes
  19
  20 ######################################################################
  21
  22 if torch.cuda.is_available():
  23     device = torch.device("cuda")
  24     torch.backends.cuda.matmul.allow_tf32 = True
  25 else:
  26     device = torch.device("cpu")
  27
  28 ######################################################################
  29
  30 parser = argparse.ArgumentParser(
  31     description="An implementation of GPT with cache.",
  32     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  33 )
  34
  35 parser.add_argument("--log_filename", type=str, default="train.log")
  36
  37 parser.add_argument("--result_dir", type=str, default=None)
  38
  39 parser.add_argument("--seed", type=int, default=0)
  40
  41 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  42
  43 ########################################
  44
  45 parser.add_argument("--nb_epochs", type=int, default=10000)
  46
  47 parser.add_argument("--batch_size", type=int, default=None)
  48
  49 parser.add_argument("--physical_batch_size", type=int, default=None)
  50
  51 parser.add_argument("--nb_train_samples", type=int, default=None)
  52
  53 parser.add_argument("--nb_test_samples", 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_gpts", type=int, default=5)
  80
  81 parser.add_argument("--min_to_validate", type=int, default=None)
  82
  83 parser.add_argument("--max_to_validate", type=int, default=None)
  84
  85 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  86
  87 parser.add_argument("--generation_temperature", type=float, default=2.0)
  88
  89 parser.add_argument("--deterministic_validation", action="store_true", default=False)
  90
  91 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
  92
  93 parser.add_argument("--dirty_debug", action="store_true", default=False)
  94
  95 ######################################################################
  96
  97 parser.add_argument("--sky_height", type=int, default=6)
  98
  99 parser.add_argument("--sky_width", type=int, default=8)
 100
 101 parser.add_argument("--sky_nb_birds", type=int, default=3)
 102
 103 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 104
 105 parser.add_argument("--sky_speed", type=int, default=3)
 106
 107 ######################################################################
 108
 109 args = parser.parse_args()
 110
 111 if args.min_to_validate is None:
 112     args.min_to_validate = args.nb_gpts - 1
 113
 114 if args.max_to_validate is None:
 115     args.max_to_validate = args.nb_gpts - 1
 116
 117 if args.result_dir is None:
 118     args.result_dir = f"results_culture"
 119
 120 ######################################################################
 121
 122 default_args = {
 123     "model": "37M",
 124     "batch_size": 100,
 125     "nb_train_samples": 100000,
 126     "nb_test_samples": 10000,
 127 }
 128
 129 for k, v in default_args.items():
 130     if getattr(args, k) is None:
 131         setattr(args, k, v)
 132
 133 ######################################################################
 134
 135 default_model_args = {
 136     "17K": {
 137         "dim_model": 32,
 138         "dim_keys": 32,
 139         "dim_hidden": 32,
 140         "nb_heads": 2,
 141         "nb_blocks": 2,
 142     },
 143     "4M": {
 144         "dim_model": 256,
 145         "dim_keys": 32,
 146         "dim_hidden": 1024,
 147         "nb_heads": 4,
 148         "nb_blocks": 6,
 149     },
 150     "37M": {
 151         "dim_model": 512,
 152         "dim_keys": 64,
 153         "dim_hidden": 2048,
 154         "nb_heads": 8,
 155         "nb_blocks": 12,
 156     },
 157     "122M": {
 158         "dim_model": 768,
 159         "dim_keys": 64,
 160         "dim_hidden": 2048,
 161         "nb_heads": 8,
 162         "nb_blocks": 24,
 163     },
 164     "352M": {
 165         "dim_model": 1024,
 166         "dim_keys": 64,
 167         "dim_hidden": 2048,
 168         "nb_heads": 8,
 169         "nb_blocks": 48,
 170     },
 171 }
 172
 173 if args.model in default_model_args:
 174     for k, v in default_model_args[args.model].items():
 175         if getattr(args, k) is None:
 176             setattr(args, k, v)
 177 else:
 178     raise ValueError(f"Unknown model {args.model}")
 179
 180 ######################################################################
 181
 182 try:
 183     os.mkdir(args.result_dir)
 184 except FileExistsError:
 185     print(f"result directory {args.result_dir} already exists")
 186     exit(1)
 187
 188 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 189
 190 if args.seed >= 0:
 191     # torch.backends.cudnn.deterministic = True
 192     # torch.backends.cudnn.benchmark = False
 193     # torch.use_deterministic_algorithms(True)
 194     torch.manual_seed(args.seed)
 195     if torch.cuda.is_available():
 196         torch.cuda.manual_seed_all(args.seed)
 197
 198 ######################################################################
 199
 200
 201 def log_string(s):
 202     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 203
 204     if log_file is not None:
 205         log_file.write(t + s + "\n")
 206         log_file.flush()
 207
 208     print(t + s)
 209     sys.stdout.flush()
 210
 211
 212 log_string(f"argv {' '.join(sys.argv)}")
 213
 214 for n in vars(args):
 215     log_string(f"args.{n} {getattr(args, n)}")
 216
 217
 218 ######################################################################
 219
 220 if args.dirty_debug:
 221     args.nb_train_samples = 2500
 222     args.nb_test_samples = 100
 223
 224 if args.physical_batch_size is None:
 225     args.physical_batch_size = args.batch_size
 226 else:
 227     assert args.batch_size % args.physical_batch_size == 0
 228
 229 assert args.nb_train_samples % args.batch_size == 0
 230 assert args.nb_test_samples % args.batch_size == 0
 231
 232 if args.problem == "sky":
 233     problem = sky.Sky(
 234         height=args.sky_height,
 235         width=args.sky_width,
 236         nb_birds=args.sky_nb_birds,
 237         nb_iterations=args.sky_nb_iterations,
 238         speed=args.sky_speed,
 239     )
 240     back_accuracy = False
 241 elif args.problem == "grids":
 242     problem = grids.Grids(device=device)
 243     back_accuracy = True
 244 else:
 245     raise ValueError
 246
 247 quiz_machine = quiz_machine.QuizMachine(
 248     problem=problem,
 249     nb_train_samples=args.nb_train_samples,
 250     nb_test_samples=args.nb_test_samples,
 251     back_accuracy=back_accuracy,
 252     batch_size=args.physical_batch_size,
 253     result_dir=args.result_dir,
 254     logger=log_string,
 255     device=device,
 256 )
 257
 258 ######################################################################
 259
 260 log_string(f"device {device}")
 261
 262 vocabulary_size = quiz_machine.vocabulary_size()
 263
 264 log_string(f"vocabulary_size {vocabulary_size}")
 265
 266 ######################################################################
 267
 268 # Compute the entropy of the training tokens
 269
 270 token_count = 0
 271 for input in quiz_machine.batches(split="train", desc="train-entropy"):
 272     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 273         (0, 1)
 274     )
 275 token_probas = token_count / token_count.sum()
 276 entropy = -torch.xlogy(token_probas, token_probas).sum()
 277 train_set_perplexity = math.exp(entropy)
 278
 279 ######################################################################
 280 # A bit of paranoia never hurts
 281
 282 if args.max_percents_of_test_in_train >= 0:
 283
 284     def subsets_as_tuples(batches, cs):
 285         s = set()
 286         for batch in batches:
 287             for x in batch:
 288                 s.add(tuple([v.item() for v in x]))
 289                 if len(s) == cs:
 290                     yield s
 291                     s = set()
 292         yield s
 293
 294     nb_test, nb_in_train = 0, 0
 295     for test_subset in subsets_as_tuples(
 296         quiz_machine.batches(split="test", desc="test-check"), 25000
 297     ):
 298         in_train = set()
 299         for train_subset in subsets_as_tuples(
 300             quiz_machine.batches(split="train", desc="train-check"), 25000
 301         ):
 302             in_train.update(test_subset.intersection(train_subset))
 303         nb_in_train += len(in_train)
 304         nb_test += len(test_subset)
 305
 306     log_string(
 307         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 308     )
 309
 310     assert (
 311         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 312     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 313
 314 ##############################
 315
 316
 317 def one_epoch(model, quiz_machine):
 318     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 319
 320     model.train()
 321
 322     nb_train_samples, acc_train_loss = 0, 0.0
 323
 324     for input in quiz_machine.batches(split="train"):
 325         input = input.to(device)
 326
 327         if nb_train_samples % args.batch_size == 0:
 328             optimizer.zero_grad()
 329
 330         output = model(mygpt.BracketedSequence(input)).x
 331         loss = F.cross_entropy(output.transpose(1, 2), input)
 332         acc_train_loss += loss.item() * input.size(0)
 333
 334         nb_train_samples += input.size(0)
 335
 336         loss.backward()
 337
 338         if nb_train_samples % args.batch_size == 0:
 339             optimizer.step()
 340
 341     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 342
 343     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 344
 345
 346 ######################################################################
 347
 348
 349 def run_tests(model, quiz_machine, deterministic_synthesis):
 350     with torch.autograd.no_grad():
 351         model.eval()
 352
 353         nb_test_samples, acc_test_loss = 0, 0.0
 354         nb_samples_accumulated = 0
 355
 356         for input in quiz_machine.batches(split="test"):
 357             input = input.to(device)
 358
 359             bs = model(mygpt.BracketedSequence(input))
 360             output = bs.x
 361
 362             loss = F.cross_entropy(output.transpose(1, 2), input)
 363
 364             acc_test_loss += loss.item() * input.size(0)
 365
 366             nb_test_samples += input.size(0)
 367
 368         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 369
 370         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 371
 372         model.main_test_accuracy = quiz_machine.produce_results(
 373             n_epoch=n_epoch,
 374             model=model,
 375             result_dir=args.result_dir,
 376             deterministic_synthesis=deterministic_synthesis,
 377         )
 378
 379
 380 ######################################################################
 381
 382
 383 def valid_c_quizzes(recorded, criteria):
 384     result = [q[criteria(c)] for q, c in recorded]
 385     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 386
 387
 388 ######################################################################
 389
 390
 391 def create_c_quizzes(
 392     models,
 393     quiz_machine,
 394     nb_for_train=1000,
 395     nb_for_test=100,
 396 ):
 397     quizzes_and_nb_correct_records = []
 398
 399     nb_to_create = nb_for_train + nb_for_test
 400
 401     # ------------------------------------------------------------
 402
 403     standard_validity = lambda nb_correct: torch.logical_and(
 404         nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
 405     )
 406
 407     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 408
 409     with open(file_name, "w") as logp_file:
 410         while (
 411             valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
 412             < nb_to_create
 413         ):
 414             # Select a model at random to generate the new quizzes
 415
 416             model_for_generation = models[torch.randint(len(models), (1,))]
 417
 418             c_quizzes = quiz_machine.generate_quizzes(
 419                 nb_to_create,
 420                 model_for_generation=model_for_generation,
 421                 temperature=args.generation_temperature,
 422             )
 423
 424             nb_correct, seq_logproba = quiz_machine.compute_correctness(
 425                 c_quizzes,
 426                 models,
 427                 bidirectional_validation=args.bidirectional_validation,
 428                 deterministic_validation=args.deterministic_validation,
 429             )
 430
 431             for n, l in zip(nb_correct, seq_logproba):
 432                 s = " ".join([str(x.item()) for x in l])
 433                 logp_file.write(f"{n} {s}\n")
 434
 435             if args.dirty_debug:
 436                 nb_correct = torch.randint(
 437                     len(models) + 1, nb_correct.size(), device=c_quizzes.device
 438                 )
 439
 440             quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 441
 442             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 443             nv = " ".join([str(x.item()) for x in nv])
 444
 445             nb_validated = valid_c_quizzes(
 446                 quizzes_and_nb_correct_records, standard_validity
 447             ).size(0)
 448
 449             log_string(
 450                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 451             )
 452
 453     # store the new c_quizzes which have been validated
 454
 455     new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
 456
 457     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 458
 459     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 460     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 461
 462     # save a bunch of images to investigate what quizzes with a
 463     # certain nb of correct predictions look like
 464
 465     for n in range(len(models) + 1):
 466         s = (
 467             "_validated"
 468             if n >= args.min_to_validate and n <= args.max_to_validate
 469             else ""
 470         )
 471
 472         q = valid_c_quizzes(
 473             quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
 474         )[:72]
 475
 476         quiz_machine.reverse_random_half_in_place(q)
 477
 478         if q.size(0) > 0:
 479             quiz_machine.save_quizzes(
 480                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 481             )
 482
 483
 484 ######################################################################
 485
 486 models = []
 487
 488 for k in range(args.nb_gpts):
 489     model = mygpt.MyGPT(
 490         vocabulary_size=vocabulary_size,
 491         dim_model=args.dim_model,
 492         dim_keys=args.dim_keys,
 493         dim_hidden=args.dim_hidden,
 494         nb_heads=args.nb_heads,
 495         nb_blocks=args.nb_blocks,
 496         causal=True,
 497         dropout=args.dropout,
 498     ).to(device)
 499
 500     model.main_test_accuracy = 0.0
 501     model.id = k
 502
 503     models.append(model)
 504
 505
 506 nb_parameters = sum(p.numel() for p in models[0].parameters())
 507 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 508
 509 ######################################################################
 510
 511 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 512 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 513
 514 log_string(
 515     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}"
 516 )
 517
 518 ######################################################################
 519
 520 if args.dirty_debug:
 521     args.accuracy_to_make_c_quizzes = 0.0
 522     args.nb_gpts = 2
 523     nb_new_c_quizzes_for_train = 100
 524     nb_new_c_quizzes_for_test = 10
 525
 526 ######################################################################
 527
 528 for n_epoch in range(args.nb_epochs):
 529     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 530
 531     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 532     log_string(f"current_test_accuracies {cta}")
 533
 534     ##################################################
 535     # Select, improve, and eval the worst model
 536
 537     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 538
 539     log_string(
 540         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 541     )
 542
 543     one_epoch(weakest_model, quiz_machine)
 544
 545     log_string(
 546         f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 547     )
 548
 549     run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
 550
 551     log_string(
 552         f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 553     )
 554
 555     ##################################################
 556     # Replace a fraction of the w_quizzes with fresh ones
 557
 558     quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 559
 560     ##################################################
 561     # If all the models are good enough, generate new quizzes and
 562     # re-compute the test errors
 563
 564     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 565         create_c_quizzes(
 566             models,
 567             quiz_machine,
 568             nb_for_train=nb_new_c_quizzes_for_train,
 569             nb_for_test=nb_new_c_quizzes_for_test,
 570         )
 571
 572         for model in models:
 573             run_tests(model, quiz_machine, deterministic_synthesis=False)
 574
 575
 576 ######################################################################