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=0.975)
  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 quiz_machine = quiz_machine.QuizMachine(
 257     problem=problem,
 258     nb_train_samples=args.nb_train_samples,
 259     nb_test_samples=args.nb_test_samples,
 260     back_accuracy=back_accuracy,
 261     batch_size=args.physical_batch_size,
 262     result_dir=args.result_dir,
 263     logger=log_string,
 264     device=device,
 265 )
 266
 267 ######################################################################
 268
 269 log_string(f"device {device}")
 270
 271 vocabulary_size = quiz_machine.vocabulary_size()
 272
 273 log_string(f"vocabulary_size {vocabulary_size}")
 274
 275 ######################################################################
 276
 277
 278 ######################################################################
 279
 280
 281 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
 282     if local_device is None:
 283         local_device = device
 284
 285     with torch.autograd.no_grad():
 286         model.eval().to(local_device)
 287
 288         nb_test_samples, acc_test_loss = 0, 0.0
 289         nb_samples_accumulated = 0
 290
 291         for input in quiz_machine.batches(model, split="test"):
 292             input = input.to(local_device)
 293
 294             bs = model(mygpt.BracketedSequence(input))
 295             output = bs.x
 296
 297             loss = F.cross_entropy(output.transpose(1, 2), input)
 298
 299             acc_test_loss += loss.item() * input.size(0)
 300
 301             nb_test_samples += input.size(0)
 302
 303         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 304
 305         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 306
 307         model.main_test_accuracy = quiz_machine.produce_results(
 308             n_epoch=n_epoch,
 309             model=model,
 310             result_dir=args.result_dir,
 311             deterministic_synthesis=deterministic_synthesis,
 312         )
 313
 314
 315 def one_epoch(model, quiz_machine, local_device=None):
 316     if local_device is None:
 317         local_device = device
 318
 319     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 320
 321     model.to(local_device).train()
 322
 323     nb_train_samples, acc_train_loss = 0, 0.0
 324
 325     for input in quiz_machine.batches(model, split="train"):
 326         input = input.to(local_device)
 327
 328         if nb_train_samples % args.batch_size == 0:
 329             optimizer.zero_grad()
 330
 331         output = model(mygpt.BracketedSequence(input)).x
 332         loss = F.cross_entropy(output.transpose(1, 2), input)
 333         acc_train_loss += loss.item() * input.size(0)
 334
 335         nb_train_samples += input.size(0)
 336
 337         loss.backward()
 338
 339         if nb_train_samples % args.batch_size == 0:
 340             optimizer.step()
 341
 342     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 343
 344     log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
 345
 346     run_tests(model, quiz_machine, deterministic_synthesis=False)
 347
 348     model.TRAINING_LOCK.release()
 349
 350
 351 ######################################################################
 352
 353
 354 def standard_validity(logproba):
 355     l = logproba.sort(dim=-1).values
 356     return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
 357
 358
 359 def valid_c_quizzes(recorded, criteria):
 360     result = [q[criteria(lp)] for q, lp in recorded]
 361     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 362
 363
 364 ######################################################################
 365
 366
 367 def create_c_quizzes(
 368     models,
 369     quiz_machine,
 370     nb_for_train=1000,
 371     nb_for_test=100,
 372 ):
 373     quizzes_and_logproba_records = []
 374
 375     nb_to_create = nb_for_train + nb_for_test
 376
 377     # ------------------------------------------------------------
 378
 379     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 380
 381     with open(file_name, "w") as logp_file:
 382         while (
 383             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 384             < nb_to_create
 385         ):
 386             # Select a model at random to generate the new quizzes
 387
 388             model_for_generation = models[torch.randint(len(models), (1,))]
 389
 390             c_quizzes = quiz_machine.generate_quizzes(
 391                 nb_to_create,
 392                 model_for_generation=model_for_generation,
 393                 temperature=args.generation_temperature,
 394             )
 395
 396             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 397
 398             if c_quizzes.size(0) > 0:
 399                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 400                 for l in logproba:
 401                     s = " ".join([str(x.item()) for x in l])
 402                     logp_file.write(s + "\n")
 403                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 404
 405             nb_validated = valid_c_quizzes(
 406                 quizzes_and_logproba_records, standard_validity
 407             ).size(0)
 408
 409             log_string(
 410                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 411             )
 412
 413     # store the new c_quizzes which have been validated
 414
 415     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 416
 417     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 418
 419     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 420     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 421
 422     # save a bunch of images to investigate what quizzes with a
 423     # certain nb of correct predictions look like
 424
 425     q = new_c_quizzes[:72]
 426
 427     if q.size(0) > 0:
 428         quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
 429
 430
 431 ######################################################################
 432
 433 models = []
 434
 435 for k in range(args.nb_gpts):
 436     log_string(f"creating model {k} and its w_quizzes")
 437     model = mygpt.MyGPT(
 438         vocabulary_size=vocabulary_size,
 439         dim_model=args.dim_model,
 440         dim_keys=args.dim_keys,
 441         dim_hidden=args.dim_hidden,
 442         nb_heads=args.nb_heads,
 443         nb_blocks=args.nb_blocks,
 444         causal=True,
 445         dropout=args.dropout,
 446     ).to(device)
 447
 448     model.main_test_accuracy = 0.0
 449     model.id = k
 450     model.TRAINING_LOCK = threading.Lock()
 451
 452     model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
 453     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 454     model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
 455     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 456
 457     models.append(model)
 458
 459
 460 nb_parameters = sum(p.numel() for p in models[0].parameters())
 461 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 462
 463 ######################################################################
 464
 465 # Compute the entropy of the training tokens
 466
 467 token_count = 0
 468 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 469     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 470         (0, 1)
 471     )
 472 token_probas = token_count / token_count.sum()
 473 entropy = -torch.xlogy(token_probas, token_probas).sum()
 474 train_set_perplexity = math.exp(entropy)
 475
 476 ######################################################################
 477 # A bit of paranoia never hurts
 478
 479 if args.max_percents_of_test_in_train >= 0:
 480
 481     def subsets_as_tuples(batches, cs):
 482         s = set()
 483         for batch in batches:
 484             for x in batch:
 485                 s.add(tuple([v.item() for v in x]))
 486                 if len(s) == cs:
 487                     yield s
 488                     s = set()
 489         yield s
 490
 491     nb_test, nb_in_train = 0, 0
 492     for test_subset in subsets_as_tuples(
 493         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 494     ):
 495         in_train = set()
 496         for train_subset in subsets_as_tuples(
 497             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 498         ):
 499             in_train.update(test_subset.intersection(train_subset))
 500         nb_in_train += len(in_train)
 501         nb_test += len(test_subset)
 502
 503     log_string(
 504         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 505     )
 506
 507     assert (
 508         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 509     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 510
 511 ######################################################################
 512
 513 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 514 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 515
 516 log_string(
 517     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}"
 518 )
 519
 520 ######################################################################
 521
 522 if args.dirty_debug:
 523     args.accuracy_to_make_c_quizzes = 0.0
 524     args.nb_gpts = 2
 525     nb_new_c_quizzes_for_train = 100
 526     nb_new_c_quizzes_for_test = 10
 527
 528     def standard_validity(logproba):
 529         l = logproba.sort(dim=-1).values
 530         return l[:, 0] < math.log(0.5)
 531
 532
 533 ######################################################################
 534
 535 for n_epoch in range(args.nb_epochs):
 536     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 537
 538     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 539     log_string(f"current_test_accuracies {cta}")
 540
 541     ##################################################
 542     # Select, improve, and eval the worst model
 543
 544     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 545
 546     weakest_models = ranked_models[: args.nb_gpus]
 547
 548     for gpu_id, model in enumerate(weakest_models):
 549         model.TRAINING_LOCK.acquire()
 550
 551         log_string(
 552             f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 553         )
 554
 555         threading.Thread(
 556             target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
 557         ).start()
 558
 559     for model in weakest_models:
 560         model.TRAINING_LOCK.acquire()
 561         model.TRAINING_LOCK.release()
 562
 563     ##################################################
 564     # Replace a fraction of the w_quizzes with fresh ones
 565
 566     log_string(
 567         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 568     )
 569
 570     # Renew entirely the train set
 571
 572     for model in weakest_models:
 573         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 574
 575     ##################################################
 576     # If all the models are good enough, generate new quizzes and
 577     # re-compute the test errors
 578
 579     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 580         create_c_quizzes(
 581             models,
 582             quiz_machine,
 583             nb_for_train=nb_new_c_quizzes_for_train,
 584             nb_for_test=nb_new_c_quizzes_for_test,
 585         )
 586
 587 ######################################################################