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 import torch.multiprocessing as mp
  22
  23 # world quizzes vs. culture quizzes
  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     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  37 )
  38
  39 parser.add_argument("--log_filename", type=str, default="train.log")
  40
  41 parser.add_argument("--result_dir", type=str, default=None)
  42
  43 parser.add_argument("--seed", type=int, default=0)
  44
  45 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
  46
  47 ########################################
  48
  49 parser.add_argument("--nb_epochs", type=int, default=10000)
  50
  51 parser.add_argument("--batch_size", type=int, default=None)
  52
  53 parser.add_argument("--physical_batch_size", type=int, default=None)
  54
  55 parser.add_argument("--nb_train_samples", type=int, default=None)
  56
  57 parser.add_argument("--nb_test_samples", type=int, default=None)
  58
  59 parser.add_argument("--learning_rate", type=float, default=5e-4)
  60
  61 ########################################
  62
  63 parser.add_argument("--model", type=str, default=None)
  64
  65 parser.add_argument("--dim_model", type=int, default=None)
  66
  67 parser.add_argument("--dim_keys", type=int, default=None)
  68
  69 parser.add_argument("--dim_hidden", type=int, default=None)
  70
  71 parser.add_argument("--nb_heads", type=int, default=None)
  72
  73 parser.add_argument("--nb_blocks", type=int, default=None)
  74
  75 parser.add_argument("--dropout", type=float, default=0.1)
  76
  77 ########################################
  78
  79 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  80
  81 parser.add_argument("--problem", type=str, default="grids")
  82
  83 parser.add_argument("--nb_threads", type=int, default=1)
  84
  85 parser.add_argument("--nb_gpus", type=int, default=1)
  86
  87 parser.add_argument("--nb_gpts", type=int, default=5)
  88
  89 parser.add_argument("--min_to_validate", type=int, default=None)
  90
  91 parser.add_argument("--max_to_validate", type=int, default=None)
  92
  93 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  94
  95 parser.add_argument("--generation_temperature", type=float, default=2.0)
  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": 25,
 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         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 244         chunk_size=100,
 245         nb_threads=args.nb_threads,
 246     )
 247     back_accuracy = False
 248 elif args.problem == "grids":
 249     problem = grids.Grids(
 250         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 251         chunk_size=100,
 252         nb_threads=args.nb_threads,
 253     )
 254     back_accuracy = True
 255 else:
 256     raise ValueError
 257
 258 problem.save_some_examples(args.result_dir)
 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
 282 ######################################################################
 283
 284
 285 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
 286     if local_device is None:
 287         local_device = device
 288
 289     with torch.autograd.no_grad():
 290         model.eval().to(local_device)
 291
 292         nb_test_samples, acc_test_loss = 0, 0.0
 293         nb_samples_accumulated = 0
 294
 295         for input in quiz_machine.batches(model, split="test"):
 296             input = input.to(local_device)
 297
 298             bs = model(mygpt.BracketedSequence(input))
 299             output = bs.x
 300
 301             loss = F.cross_entropy(output.transpose(1, 2), input)
 302
 303             acc_test_loss += loss.item() * input.size(0)
 304
 305             nb_test_samples += input.size(0)
 306
 307         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 308
 309         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 310
 311         model.main_test_accuracy = quiz_machine.produce_results(
 312             n_epoch=n_epoch,
 313             model=model,
 314             result_dir=args.result_dir,
 315             deterministic_synthesis=deterministic_synthesis,
 316         )
 317
 318
 319 def one_epoch(model, quiz_machine, local_device=None):
 320     if local_device is None:
 321         local_device = device
 322
 323     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 324
 325     model.to(local_device).train()
 326
 327     nb_train_samples, acc_train_loss = 0, 0.0
 328
 329     for input in quiz_machine.batches(model, split="train"):
 330         input = input.to(local_device)
 331
 332         if nb_train_samples % args.batch_size == 0:
 333             optimizer.zero_grad()
 334
 335         output = model(mygpt.BracketedSequence(input)).x
 336         loss = F.cross_entropy(output.transpose(1, 2), input)
 337         acc_train_loss += loss.item() * input.size(0)
 338
 339         nb_train_samples += input.size(0)
 340
 341         loss.backward()
 342
 343         if nb_train_samples % args.batch_size == 0:
 344             optimizer.step()
 345
 346     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 347
 348     log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
 349
 350     run_tests(model, quiz_machine, deterministic_synthesis=False)
 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
 360
 361 def valid_c_quizzes(recorded, criteria):
 362     result = [q[criteria(lp)] for q, lp in recorded]
 363     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 364
 365
 366 ######################################################################
 367
 368
 369 def create_c_quizzes(
 370     models,
 371     quiz_machine,
 372     nb_for_train=1000,
 373     nb_for_test=100,
 374 ):
 375     quizzes_and_logproba_records = []
 376
 377     nb_to_create = nb_for_train + nb_for_test
 378
 379     # ------------------------------------------------------------
 380
 381     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 382
 383     with open(file_name, "w") as logp_file:
 384         while (
 385             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 386             < nb_to_create
 387         ):
 388             # Select a model at random to generate the new quizzes
 389
 390             model_for_generation = models[torch.randint(len(models), (1,))]
 391
 392             c_quizzes = quiz_machine.generate_quizzes(
 393                 nb_to_create,
 394                 model_for_generation=model_for_generation,
 395                 temperature=args.generation_temperature,
 396             )
 397
 398             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 399
 400             if c_quizzes.size(0) > 0:
 401                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 402                 for l in logproba:
 403                     s = " ".join([str(x.item()) for x in l])
 404                     logp_file.write(s + "\n")
 405                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 406
 407             nb_validated = valid_c_quizzes(
 408                 quizzes_and_logproba_records, standard_validity
 409             ).size(0)
 410
 411             log_string(
 412                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 413             )
 414
 415     # store the new c_quizzes which have been validated
 416
 417     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 418
 419     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 420
 421     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 422     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 423
 424     # save a bunch of images to investigate what quizzes with a
 425     # certain nb of correct predictions look like
 426
 427     q = new_c_quizzes[:72]
 428
 429     if q.size(0) > 0:
 430         quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
 431
 432
 433 ######################################################################
 434
 435 models = []
 436
 437 for k in range(args.nb_gpts):
 438     log_string(f"creating model {k} and its w_quizzes")
 439     model = mygpt.MyGPT(
 440         vocabulary_size=vocabulary_size,
 441         dim_model=args.dim_model,
 442         dim_keys=args.dim_keys,
 443         dim_hidden=args.dim_hidden,
 444         nb_heads=args.nb_heads,
 445         nb_blocks=args.nb_blocks,
 446         causal=True,
 447         dropout=args.dropout,
 448     ).to(device)
 449
 450     model.main_test_accuracy = 0.0
 451     model.id = k
 452
 453     model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
 454     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 455     model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
 456     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 457
 458     models.append(model)
 459
 460
 461 nb_parameters = sum(p.numel() for p in models[0].parameters())
 462 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 463
 464 ######################################################################
 465
 466 # Compute the entropy of the training tokens
 467
 468 token_count = 0
 469 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 470     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 471         (0, 1)
 472     )
 473 token_probas = token_count / token_count.sum()
 474 entropy = -torch.xlogy(token_probas, token_probas).sum()
 475 train_set_perplexity = math.exp(entropy)
 476
 477 ######################################################################
 478 # A bit of paranoia never hurts
 479
 480 if args.max_percents_of_test_in_train >= 0:
 481
 482     def subsets_as_tuples(batches, cs):
 483         s = set()
 484         for batch in batches:
 485             for x in batch:
 486                 s.add(tuple([v.item() for v in x]))
 487                 if len(s) == cs:
 488                     yield s
 489                     s = set()
 490         yield s
 491
 492     nb_test, nb_in_train = 0, 0
 493     for test_subset in subsets_as_tuples(
 494         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 495     ):
 496         in_train = set()
 497         for train_subset in subsets_as_tuples(
 498             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 499         ):
 500             in_train.update(test_subset.intersection(train_subset))
 501         nb_in_train += len(in_train)
 502         nb_test += len(test_subset)
 503
 504     log_string(
 505         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 506     )
 507
 508     assert (
 509         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 510     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 511
 512 ######################################################################
 513
 514 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 515 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 516
 517 log_string(
 518     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}"
 519 )
 520
 521 ######################################################################
 522
 523 if args.dirty_debug:
 524     args.accuracy_to_make_c_quizzes = 0.0
 525     args.nb_gpts = 2
 526     nb_new_c_quizzes_for_train = 100
 527     nb_new_c_quizzes_for_test = 10
 528
 529     def standard_validity(logproba):
 530         l = logproba.sort(dim=-1).values
 531         return l[:, 0] < math.log(0.5)
 532
 533
 534 ######################################################################
 535
 536 for n_epoch in range(args.nb_epochs):
 537     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 538
 539     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 540     log_string(f"current_test_accuracies {cta}")
 541
 542     ##################################################
 543     # Select, improve, and eval the worst model
 544
 545     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 546
 547     weakest_models = ranked_models[: args.nb_gpus]
 548
 549     threads = []
 550
 551     for gpu_id, model in enumerate(weakest_models):
 552         log_string(f"training model {model.id}")
 553
 554         t = threading.Thread(
 555             target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
 556         )
 557
 558         threads.append(t)
 559
 560         t.start()
 561
 562     for t in threads:
 563         t.join()
 564
 565     ##################################################
 566     # Replace a fraction of the w_quizzes with fresh ones
 567
 568     log_string(
 569         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 570     )
 571
 572     # Renew entirely the train set
 573
 574     for model in weakest_models:
 575         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 576
 577     ##################################################
 578     # If all the models are good enough, generate new quizzes and
 579     # re-compute the test errors
 580
 581     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 582         create_c_quizzes(
 583             models,
 584             quiz_machine,
 585             nb_for_train=nb_new_c_quizzes_for_train,
 586             nb_for_test=nb_new_c_quizzes_for_test,
 587         )
 588
 589 ######################################################################