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 nb_new_c_quizzes_for_train = 1000
  23 nb_new_c_quizzes_for_test = 100
  24
  25 ######################################################################
  26
  27 if torch.cuda.is_available():
  28     device = torch.device("cuda")
  29     torch.backends.cuda.matmul.allow_tf32 = True
  30 else:
  31     device = torch.device("cpu")
  32
  33 ######################################################################
  34
  35 parser = argparse.ArgumentParser(
  36     description="An implementation of GPT with cache.",
  37     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  38 )
  39
  40 parser.add_argument("--log_filename", type=str, default="train.log")
  41
  42 parser.add_argument("--result_dir", type=str, default=None)
  43
  44 parser.add_argument("--seed", type=int, default=0)
  45
  46 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  47
  48 ########################################
  49
  50 parser.add_argument("--nb_epochs", type=int, default=10000)
  51
  52 parser.add_argument("--batch_size", type=int, default=None)
  53
  54 parser.add_argument("--physical_batch_size", type=int, default=None)
  55
  56 parser.add_argument("--nb_train_samples", type=int, default=None)
  57
  58 parser.add_argument("--nb_test_samples", type=int, default=None)
  59
  60 parser.add_argument("--learning_rate", type=float, default=5e-4)
  61
  62 ########################################
  63
  64 parser.add_argument("--model", type=str, default=None)
  65
  66 parser.add_argument("--dim_model", type=int, default=None)
  67
  68 parser.add_argument("--dim_keys", type=int, default=None)
  69
  70 parser.add_argument("--dim_hidden", type=int, default=None)
  71
  72 parser.add_argument("--nb_heads", type=int, default=None)
  73
  74 parser.add_argument("--nb_blocks", type=int, default=None)
  75
  76 parser.add_argument("--dropout", type=float, default=0.1)
  77
  78 ########################################
  79
  80 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  81
  82 parser.add_argument("--problem", type=str, default="grids")
  83
  84 parser.add_argument("--nb_gpts", type=int, default=5)
  85
  86 parser.add_argument("--min_to_validate", type=int, default=None)
  87
  88 parser.add_argument("--max_to_validate", type=int, default=None)
  89
  90 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  91
  92 parser.add_argument("--generation_temperature", type=float, default=2.0)
  93
  94 parser.add_argument("--deterministic_validation", action="store_true", default=False)
  95
  96 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
  97
  98 parser.add_argument("--dirty_debug", action="store_true", default=False)
  99
 100 ######################################################################
 101
 102 parser.add_argument("--sky_height", type=int, default=6)
 103
 104 parser.add_argument("--sky_width", type=int, default=8)
 105
 106 parser.add_argument("--sky_nb_birds", type=int, default=3)
 107
 108 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 109
 110 parser.add_argument("--sky_speed", type=int, default=3)
 111
 112 ######################################################################
 113
 114 args = parser.parse_args()
 115
 116 if args.min_to_validate is None:
 117     args.min_to_validate = args.nb_gpts - 1
 118
 119 if args.max_to_validate is None:
 120     args.max_to_validate = args.nb_gpts - 1
 121
 122 if args.result_dir is None:
 123     args.result_dir = f"results_culture"
 124
 125 ######################################################################
 126
 127 if args.dirty_debug:
 128     args.accuracy_to_make_c_quizzes = 0.0
 129     args.nb_gpts = 2
 130     nb_new_c_quizzes_for_train = 100
 131     nb_new_c_quizzes_for_test = 10
 132
 133 ######################################################################
 134
 135 default_args = {
 136     "model": "37M",
 137     "batch_size": 100,
 138     "nb_train_samples": 100000,
 139     "nb_test_samples": 10000,
 140 }
 141
 142 for k, v in default_args.items():
 143     if getattr(args, k) is None:
 144         setattr(args, k, v)
 145
 146 ######################################################################
 147
 148 default_model_args = {
 149     "17K": {
 150         "dim_model": 32,
 151         "dim_keys": 32,
 152         "dim_hidden": 32,
 153         "nb_heads": 2,
 154         "nb_blocks": 2,
 155     },
 156     "4M": {
 157         "dim_model": 256,
 158         "dim_keys": 32,
 159         "dim_hidden": 1024,
 160         "nb_heads": 4,
 161         "nb_blocks": 6,
 162     },
 163     "37M": {
 164         "dim_model": 512,
 165         "dim_keys": 64,
 166         "dim_hidden": 2048,
 167         "nb_heads": 8,
 168         "nb_blocks": 12,
 169     },
 170     "122M": {
 171         "dim_model": 768,
 172         "dim_keys": 64,
 173         "dim_hidden": 2048,
 174         "nb_heads": 8,
 175         "nb_blocks": 24,
 176     },
 177     "352M": {
 178         "dim_model": 1024,
 179         "dim_keys": 64,
 180         "dim_hidden": 2048,
 181         "nb_heads": 8,
 182         "nb_blocks": 48,
 183     },
 184 }
 185
 186 if args.model in default_model_args:
 187     for k, v in default_model_args[args.model].items():
 188         if getattr(args, k) is None:
 189             setattr(args, k, v)
 190 else:
 191     raise ValueError(f"Unknown model {args.model}")
 192
 193 ######################################################################
 194
 195 try:
 196     os.mkdir(args.result_dir)
 197 except FileExistsError:
 198     print(f"result directory {args.result_dir} already exists")
 199     exit(1)
 200
 201 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 202
 203 if args.seed >= 0:
 204     # torch.backends.cudnn.deterministic = True
 205     # torch.backends.cudnn.benchmark = False
 206     # torch.use_deterministic_algorithms(True)
 207     torch.manual_seed(args.seed)
 208     if torch.cuda.is_available():
 209         torch.cuda.manual_seed_all(args.seed)
 210
 211 ######################################################################
 212
 213
 214 def log_string(s):
 215     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 216
 217     if log_file is not None:
 218         log_file.write(t + s + "\n")
 219         log_file.flush()
 220
 221     print(t + s)
 222     sys.stdout.flush()
 223
 224
 225 log_string(f"argv {' '.join(sys.argv)}")
 226
 227 for n in vars(args):
 228     log_string(f"args.{n} {getattr(args, n)}")
 229
 230
 231 ######################################################################
 232
 233 if args.dirty_debug:
 234     args.nb_train_samples = 2500
 235     args.nb_test_samples = 100
 236
 237 if args.physical_batch_size is None:
 238     args.physical_batch_size = args.batch_size
 239 else:
 240     assert args.batch_size % args.physical_batch_size == 0
 241
 242 assert args.nb_train_samples % args.batch_size == 0
 243 assert args.nb_test_samples % args.batch_size == 0
 244
 245 if args.problem == "sky":
 246     problem = sky.Sky(
 247         height=args.sky_height,
 248         width=args.sky_width,
 249         nb_birds=args.sky_nb_birds,
 250         nb_iterations=args.sky_nb_iterations,
 251         speed=args.sky_speed,
 252     )
 253     back_accuracy = False
 254 elif args.problem == "grids":
 255     problem = grids.Grids(device=device)
 256     back_accuracy = True
 257 else:
 258     raise ValueError
 259
 260 quiz_machine = quiz_machine.QuizMachine(
 261     problem=problem,
 262     nb_train_samples=args.nb_train_samples,
 263     nb_test_samples=args.nb_test_samples,
 264     back_accuracy=back_accuracy,
 265     batch_size=args.physical_batch_size,
 266     result_dir=args.result_dir,
 267     logger=log_string,
 268     device=device,
 269 )
 270
 271 ######################################################################
 272
 273 log_string(f"device {device}")
 274
 275 vocabulary_size = quiz_machine.vocabulary_size()
 276
 277 log_string(f"vocabulary_size {vocabulary_size}")
 278
 279 ######################################################################
 280
 281 # Compute the entropy of the training tokens
 282
 283 token_count = 0
 284 for input in quiz_machine.batches(split="train", desc="train-entropy"):
 285     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 286         (0, 1)
 287     )
 288 token_probas = token_count / token_count.sum()
 289 entropy = -torch.xlogy(token_probas, token_probas).sum()
 290 train_set_perplexity = math.exp(entropy)
 291
 292 ######################################################################
 293 # A bit of paranoia never hurts
 294
 295 if args.max_percents_of_test_in_train >= 0:
 296
 297     def subsets_as_tuples(batches, cs):
 298         s = set()
 299         for batch in batches:
 300             for x in batch:
 301                 s.add(tuple([v.item() for v in x]))
 302                 if len(s) == cs:
 303                     yield s
 304                     s = set()
 305         yield s
 306
 307     nb_test, nb_in_train = 0, 0
 308     for test_subset in subsets_as_tuples(
 309         quiz_machine.batches(split="test", desc="test-check"), 25000
 310     ):
 311         in_train = set()
 312         for train_subset in subsets_as_tuples(
 313             quiz_machine.batches(split="train", desc="train-check"), 25000
 314         ):
 315             in_train.update(test_subset.intersection(train_subset))
 316         nb_in_train += len(in_train)
 317         nb_test += len(test_subset)
 318
 319     log_string(
 320         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 321     )
 322
 323     assert (
 324         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 325     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 326
 327 ##############################
 328
 329
 330 def one_epoch(model, quiz_machine):
 331     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 332
 333     model.train()
 334
 335     nb_train_samples, acc_train_loss = 0, 0.0
 336
 337     for input in quiz_machine.batches(split="train"):
 338         input = input.to(device)
 339
 340         if nb_train_samples % args.batch_size == 0:
 341             optimizer.zero_grad()
 342
 343         output = model(mygpt.BracketedSequence(input)).x
 344         loss = F.cross_entropy(output.transpose(1, 2), input)
 345         acc_train_loss += loss.item() * input.size(0)
 346
 347         nb_train_samples += input.size(0)
 348
 349         loss.backward()
 350
 351         if nb_train_samples % args.batch_size == 0:
 352             optimizer.step()
 353
 354     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 355
 356     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 357
 358
 359 ######################################################################
 360
 361
 362 def run_tests(model, quiz_machine, deterministic_synthesis):
 363     with torch.autograd.no_grad():
 364         model.eval()
 365
 366         nb_test_samples, acc_test_loss = 0, 0.0
 367         nb_samples_accumulated = 0
 368
 369         for input in quiz_machine.batches(split="test"):
 370             input = input.to(device)
 371
 372             bs = model(mygpt.BracketedSequence(input))
 373             output = bs.x
 374
 375             loss = F.cross_entropy(output.transpose(1, 2), input)
 376
 377             acc_test_loss += loss.item() * input.size(0)
 378
 379             nb_test_samples += input.size(0)
 380
 381         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 382
 383         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 384
 385         model.main_test_accuracy = quiz_machine.produce_results(
 386             n_epoch=n_epoch,
 387             model=model,
 388             result_dir=args.result_dir,
 389             deterministic_synthesis=deterministic_synthesis,
 390         )
 391
 392
 393 ######################################################################
 394
 395
 396 def valid_c_quizzes(recorded, criteria):
 397     result = [q[criteria(c)] for q, c in recorded]
 398     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 399
 400
 401 ######################################################################
 402
 403
 404 def create_c_quizzes(
 405     models,
 406     quiz_machine,
 407     nb_for_train=1000,
 408     nb_for_test=100,
 409 ):
 410     quizzes_and_nb_correct_records = []
 411
 412     nb_to_create = nb_for_train + nb_for_test
 413
 414     # ------------------------------------------------------------
 415
 416     standard_validity = lambda nb_correct: torch.logical_and(
 417         nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
 418     )
 419
 420     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 421
 422     with open(file_name, "w") as logp_file:
 423         while (
 424             valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
 425             < nb_to_create
 426         ):
 427             # Select a model at random to generate the new quizzes
 428
 429             model_for_generation = models[torch.randint(len(models), (1,))]
 430
 431             c_quizzes = quiz_machine.generate_quizzes(
 432                 nb_to_create,
 433                 model_for_generation=model_for_generation,
 434                 temperature=args.generation_temperature,
 435             )
 436
 437             nb_correct, seq_logproba = quiz_machine.compute_correctness(
 438                 c_quizzes,
 439                 models,
 440                 bidirectional_validation=args.bidirectional_validation,
 441                 deterministic_validation=args.deterministic_validation,
 442             )
 443
 444             for n, l in zip(nb_correct, seq_logproba):
 445                 s = " ".join([str(x.item()) for x in l])
 446                 logp_file.write(f"{n} {s}\n")
 447
 448             if args.dirty_debug:
 449                 nb_correct = torch.randint(
 450                     len(models) + 1, nb_correct.size(), device=c_quizzes.device
 451                 )
 452
 453             quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 454
 455             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 456             nv = " ".join([str(x.item()) for x in nv])
 457
 458             nb_validated = valid_c_quizzes(
 459                 quizzes_and_nb_correct_records, standard_validity
 460             ).size(0)
 461
 462             log_string(
 463                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 464             )
 465
 466     # store the new c_quizzes which have been validated
 467
 468     new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
 469
 470     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 471
 472     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 473     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 474
 475     # save a bunch of images to investigate what quizzes with a
 476     # certain nb of correct predictions look like
 477
 478     for n in range(len(models) + 1):
 479         s = (
 480             "_validated"
 481             if n >= args.min_to_validate and n <= args.max_to_validate
 482             else ""
 483         )
 484
 485         q = valid_c_quizzes(
 486             quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
 487         )[:72]
 488
 489         quiz_machine.reverse_random_half_in_place(q)
 490
 491         if q.size(0) > 0:
 492             quiz_machine.save_quizzes(
 493                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 494             )
 495
 496
 497 ######################################################################
 498
 499 models = []
 500
 501 for k in range(args.nb_gpts):
 502     model = mygpt.MyGPT(
 503         vocabulary_size=vocabulary_size,
 504         dim_model=args.dim_model,
 505         dim_keys=args.dim_keys,
 506         dim_hidden=args.dim_hidden,
 507         nb_heads=args.nb_heads,
 508         nb_blocks=args.nb_blocks,
 509         causal=True,
 510         dropout=args.dropout,
 511     ).to(device)
 512
 513     model.main_test_accuracy = 0.0
 514     model.id = k
 515
 516     models.append(model)
 517
 518
 519 nb_parameters = sum(p.numel() for p in models[0].parameters())
 520 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 521
 522 ######################################################################
 523
 524 for n_epoch in range(args.nb_epochs):
 525     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 526
 527     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 528     log_string(f"current_test_accuracies {cta}")
 529
 530     # Select, improve, and eval the worst model
 531
 532     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 533
 534     log_string(
 535         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 536     )
 537
 538     one_epoch(weakest_model, quiz_machine)
 539
 540     log_string(
 541         f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 542     )
 543
 544     run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
 545
 546     log_string(
 547         f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
 548     )
 549
 550     # Replace a fraction of the w_quizzes with fresh ones
 551
 552     quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 553
 554     # If all the models are good enough, generate new quizzes and
 555     # re-compute the test errors
 556
 557     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 558         create_c_quizzes(
 559             models,
 560             quiz_machine,
 561             nb_for_train=nb_new_c_quizzes_for_train,
 562             nb_for_test=nb_new_c_quizzes_for_test,
 563         )
 564
 565         for model in models:
 566             run_tests(model, quiz_machine, deterministic_synthesis=False)
 567
 568
 569 ######################################################################