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 ######################################################################
  24
  25 parser = argparse.ArgumentParser(
  26     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  27 )
  28
  29 parser.add_argument("--log_filename", type=str, default="train.log")
  30
  31 parser.add_argument("--result_dir", type=str, default=None)
  32
  33 parser.add_argument("--seed", type=int, default=0)
  34
  35 parser.add_argument("--resume", action="store_true", default=False)
  36
  37 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
  38
  39 ########################################
  40
  41 parser.add_argument("--nb_epochs", type=int, default=10000)
  42
  43 parser.add_argument("--batch_size", type=int, default=None)
  44
  45 parser.add_argument("--physical_batch_size", type=int, default=None)
  46
  47 parser.add_argument("--nb_train_samples", type=int, default=None)
  48
  49 parser.add_argument("--nb_test_samples", type=int, default=None)
  50
  51 parser.add_argument("--learning_rate", type=float, default=5e-4)
  52
  53 ########################################
  54
  55 parser.add_argument("--model", type=str, default=None)
  56
  57 parser.add_argument("--dim_model", type=int, default=None)
  58
  59 parser.add_argument("--dim_keys", type=int, default=None)
  60
  61 parser.add_argument("--dim_hidden", type=int, default=None)
  62
  63 parser.add_argument("--nb_heads", type=int, default=None)
  64
  65 parser.add_argument("--nb_blocks", type=int, default=None)
  66
  67 parser.add_argument("--dropout", type=float, default=0.1)
  68
  69 ########################################
  70
  71 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  72
  73 parser.add_argument("--problem", type=str, default="grids")
  74
  75 parser.add_argument("--nb_threads", type=int, default=1)
  76
  77 parser.add_argument("--gpus", type=str, default="all")
  78
  79 parser.add_argument("--nb_gpts", type=int, default=5)
  80
  81 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  82
  83 parser.add_argument("--proba_understands", type=float, default=0.99)
  84
  85 parser.add_argument("--proba_not_understands", type=float, default=0.5)
  86
  87 parser.add_argument("--generation_temperature", type=float, default=2.0)
  88
  89 parser.add_argument("--dirty_debug", action="store_true", default=False)
  90
  91 ######################################################################
  92
  93 grids_tasks = ", ".join(
  94     [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
  95 )
  96
  97 parser.add_argument(
  98     "--grids_tasks",
  99     type=str,
 100     default=None,
 101     help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
 102 )
 103
 104 ######################################################################
 105
 106 parser.add_argument("--sky_height", type=int, default=6)
 107
 108 parser.add_argument("--sky_width", type=int, default=8)
 109
 110 parser.add_argument("--sky_nb_birds", type=int, default=3)
 111
 112 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 113
 114 parser.add_argument("--sky_speed", type=int, default=3)
 115
 116 ######################################################################
 117
 118 args = parser.parse_args()
 119
 120 if args.result_dir is None:
 121     args.result_dir = f"results_culture"
 122
 123 ######################################################################
 124
 125 default_args = {
 126     "model": "37M",
 127     "batch_size": 25,
 128     "nb_train_samples": 100000,
 129     "nb_test_samples": 10000,
 130 }
 131
 132 for k, v in default_args.items():
 133     if getattr(args, k) is None:
 134         setattr(args, k, v)
 135
 136 ######################################################################
 137
 138 default_model_args = {
 139     "17K": {
 140         "dim_model": 32,
 141         "dim_keys": 32,
 142         "dim_hidden": 32,
 143         "nb_heads": 2,
 144         "nb_blocks": 2,
 145     },
 146     "4M": {
 147         "dim_model": 256,
 148         "dim_keys": 32,
 149         "dim_hidden": 1024,
 150         "nb_heads": 4,
 151         "nb_blocks": 6,
 152     },
 153     "37M": {
 154         "dim_model": 512,
 155         "dim_keys": 64,
 156         "dim_hidden": 2048,
 157         "nb_heads": 8,
 158         "nb_blocks": 12,
 159     },
 160     "122M": {
 161         "dim_model": 768,
 162         "dim_keys": 64,
 163         "dim_hidden": 2048,
 164         "nb_heads": 8,
 165         "nb_blocks": 24,
 166     },
 167     "352M": {
 168         "dim_model": 1024,
 169         "dim_keys": 64,
 170         "dim_hidden": 2048,
 171         "nb_heads": 8,
 172         "nb_blocks": 48,
 173     },
 174 }
 175
 176 if args.model in default_model_args:
 177     for k, v in default_model_args[args.model].items():
 178         if getattr(args, k) is None:
 179             setattr(args, k, v)
 180 else:
 181     raise ValueError(f"Unknown model {args.model}")
 182
 183 ######################################################################
 184
 185 if args.resume:
 186     assert os.path.isdir(args.result_dir)
 187
 188 else:
 189     try:
 190         os.mkdir(args.result_dir)
 191     except FileExistsError:
 192         print(f"result directory {args.result_dir} already exists")
 193         exit(1)
 194
 195 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 196
 197 if args.seed >= 0:
 198     # torch.backends.cudnn.deterministic = True
 199     # torch.backends.cudnn.benchmark = False
 200     # torch.use_deterministic_algorithms(True)
 201     torch.manual_seed(args.seed)
 202     if torch.cuda.is_available():
 203         torch.cuda.manual_seed_all(args.seed)
 204
 205 ######################################################################
 206
 207
 208 def log_string(s):
 209     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 210
 211     if log_file is not None:
 212         log_file.write(t + s + "\n")
 213         log_file.flush()
 214
 215     print(t + s)
 216     sys.stdout.flush()
 217
 218
 219 log_string(f"argv {' '.join(sys.argv)}")
 220
 221 for n in vars(args):
 222     log_string(f"args.{n} {getattr(args, n)}")
 223
 224
 225 ######################################################################
 226
 227 if args.gpus == "all":
 228     gpus_idx = range(torch.cuda.device_count())
 229 else:
 230     gpus_idx = [int(k) for k in args.gpus.split(",")]
 231
 232 gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
 233
 234 if torch.cuda.is_available():
 235     main_device = gpus[0]
 236 else:
 237     assert len(gpus) == 0
 238     main_device = torch.device("cpu")
 239
 240 if args.dirty_debug:
 241     args.nb_train_samples = 2500
 242     args.nb_test_samples = 100
 243
 244 if args.physical_batch_size is None:
 245     args.physical_batch_size = args.batch_size
 246 else:
 247     assert args.batch_size % args.physical_batch_size == 0
 248
 249 assert args.nb_train_samples % args.batch_size == 0
 250 assert args.nb_test_samples % args.batch_size == 0
 251
 252 if args.problem == "sky":
 253     problem = sky.Sky(
 254         height=args.sky_height,
 255         width=args.sky_width,
 256         nb_birds=args.sky_nb_birds,
 257         nb_iterations=args.sky_nb_iterations,
 258         speed=args.sky_speed,
 259         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 260         chunk_size=100,
 261         nb_threads=args.nb_threads,
 262     )
 263     back_accuracy = False
 264 elif args.problem == "grids":
 265     problem = grids.Grids(
 266         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 267         chunk_size=100,
 268         nb_threads=args.nb_threads,
 269         tasks=args.grids_tasks,
 270     )
 271     back_accuracy = True
 272 else:
 273     raise ValueError
 274
 275 problem.save_some_examples(args.result_dir)
 276
 277 quiz_machine = quiz_machine.QuizMachine(
 278     problem=problem,
 279     nb_train_samples=args.nb_train_samples,
 280     nb_test_samples=args.nb_test_samples,
 281     back_accuracy=back_accuracy,
 282     batch_size=args.physical_batch_size,
 283     result_dir=args.result_dir,
 284     logger=log_string,
 285     device=main_device,
 286 )
 287
 288 ######################################################################
 289
 290 log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
 291
 292 vocabulary_size = quiz_machine.vocabulary_size()
 293
 294 log_string(f"vocabulary_size {vocabulary_size}")
 295
 296 ######################################################################
 297
 298
 299 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
 300     with torch.autograd.no_grad():
 301         model.eval().to(local_device)
 302
 303         nb_test_samples, acc_test_loss = 0, 0.0
 304         nb_samples_accumulated = 0
 305
 306         for input in quiz_machine.batches(model, split="test"):
 307             input = input.to(local_device)
 308
 309             bs = model(mygpt.BracketedSequence(input))
 310             output = bs.x
 311
 312             loss = F.cross_entropy(output.transpose(1, 2), input)
 313
 314             acc_test_loss += loss.item() * input.size(0)
 315
 316             nb_test_samples += input.size(0)
 317
 318         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 319
 320         log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
 321
 322         model.main_test_accuracy = quiz_machine.produce_results(
 323             n_epoch=n_epoch,
 324             model=model,
 325             result_dir=args.result_dir,
 326             deterministic_synthesis=deterministic_synthesis,
 327         )
 328
 329
 330 def one_epoch(model, quiz_machine, local_device=main_device):
 331     model.to(local_device).train()
 332
 333     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 334
 335     nb_train_samples, acc_train_loss = 0, 0.0
 336
 337     for input in quiz_machine.batches(model, split="train"):
 338         input = input.to(local_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} model {model.id} {train_perplexity}")
 357
 358     run_tests(model, quiz_machine, deterministic_synthesis=False)
 359
 360     model.to(main_device)
 361
 362
 363 ######################################################################
 364
 365
 366 def standard_validity(logproba):
 367     l = logproba.sort(dim=-1).values
 368     return (l[:, 0] < math.log(args.proba_not_understands)) & (
 369         l[:, 1] > math.log(args.proba_understands)
 370     )
 371
 372
 373 def valid_c_quizzes(recorded, criteria):
 374     result = [q[criteria(lp)] for q, lp in recorded]
 375     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 376
 377
 378 ######################################################################
 379
 380
 381 def create_c_quizzes(
 382     models,
 383     quiz_machine,
 384     nb_for_train=1000,
 385     nb_for_test=100,
 386 ):
 387     quizzes_and_logproba_records = []
 388
 389     nb_to_create = nb_for_train + nb_for_test
 390
 391     # ------------------------------------------------------------
 392
 393     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 394
 395     with open(file_name, "w") as logp_file:
 396         while (
 397             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 398             < nb_to_create
 399         ):
 400             # Select a model at random to generate the new quizzes
 401
 402             model_for_generation = models[torch.randint(len(models), (1,))]
 403
 404             c_quizzes = quiz_machine.generate_quizzes(
 405                 nb_to_create,
 406                 model_for_generation=model_for_generation,
 407                 temperature=args.generation_temperature,
 408             )
 409
 410             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 411
 412             if c_quizzes.size(0) > 0:
 413                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 414                 for l in logproba:
 415                     s = " ".join([str(x.item()) for x in l])
 416                     logp_file.write(s + "\n")
 417                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 418
 419             nb_validated = valid_c_quizzes(
 420                 quizzes_and_logproba_records, standard_validity
 421             ).size(0)
 422
 423             log_string(
 424                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 425             )
 426
 427     # store the new c_quizzes which have been validated
 428
 429     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 430
 431     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 432
 433     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 434     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 435
 436     # save images
 437
 438     q = new_c_quizzes[:72]
 439
 440     if q.size(0) > 0:
 441         quiz_machine.save_quiz_illustrations(
 442             args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q
 443         )
 444
 445
 446 ######################################################################
 447
 448 models = []
 449
 450 for k in range(args.nb_gpts):
 451     log_string(f"creating model {k} and its w_quizzes")
 452     model = mygpt.MyGPT(
 453         vocabulary_size=vocabulary_size,
 454         dim_model=args.dim_model,
 455         dim_keys=args.dim_keys,
 456         dim_hidden=args.dim_hidden,
 457         nb_heads=args.nb_heads,
 458         nb_blocks=args.nb_blocks,
 459         causal=True,
 460         dropout=args.dropout,
 461     ).to(main_device)
 462
 463     model.main_test_accuracy = 0.0
 464     model.id = k
 465
 466     model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
 467     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 468     model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
 469     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 470
 471     models.append(model)
 472
 473 ######################################################################
 474
 475 if args.resume:
 476     try:
 477         for model in models:
 478             filename = f"gpt_{model.id:03d}.pth"
 479
 480             try:
 481                 model.load_state_dict(
 482                     torch.load(os.path.join(args.result_dir, filename))
 483                 )
 484                 log_string(f"successfully loaded {filename}")
 485             except FileNotFoundError:
 486                 log_string(f"cannot find {filename}")
 487                 pass
 488
 489         try:
 490             filename = "c_quizzes.pth"
 491             quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
 492             log_string(f"successfully loaded {filename}")
 493         except FileNotFoundError:
 494             log_string(f"cannot find {filename}")
 495             pass
 496
 497     except:
 498         log_string(f"error when loading {filename}.")
 499         exit(1)
 500
 501 ######################################################################
 502
 503 nb_parameters = sum(p.numel() for p in models[0].parameters())
 504 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 505
 506 ######################################################################
 507
 508 # Compute the entropy of the training tokens
 509
 510 token_count = 0
 511 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 512     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 513         (0, 1)
 514     )
 515 token_probas = token_count / token_count.sum()
 516 entropy = -torch.xlogy(token_probas, token_probas).sum()
 517 train_set_perplexity = math.exp(entropy)
 518
 519 ######################################################################
 520 # A bit of paranoia never hurts
 521
 522 if args.max_percents_of_test_in_train >= 0:
 523
 524     def subsets_as_tuples(batches, cs):
 525         s = set()
 526         for batch in batches:
 527             for x in batch:
 528                 s.add(tuple([v.item() for v in x]))
 529                 if len(s) == cs:
 530                     yield s
 531                     s = set()
 532         yield s
 533
 534     nb_test, nb_in_train = 0, 0
 535     for test_subset in subsets_as_tuples(
 536         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 537     ):
 538         in_train = set()
 539         for train_subset in subsets_as_tuples(
 540             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 541         ):
 542             in_train.update(test_subset.intersection(train_subset))
 543         nb_in_train += len(in_train)
 544         nb_test += len(test_subset)
 545
 546     log_string(
 547         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 548     )
 549
 550     assert (
 551         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 552     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 553
 554 ######################################################################
 555
 556 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 557 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 558
 559 log_string(
 560     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}"
 561 )
 562
 563 ######################################################################
 564
 565 if args.dirty_debug:
 566     args.accuracy_to_make_c_quizzes = 0.0
 567     args.nb_gpts = 2
 568     nb_new_c_quizzes_for_train = 100
 569     nb_new_c_quizzes_for_test = 10
 570
 571     def standard_validity(logproba):
 572         l = logproba.sort(dim=-1).values
 573         return l[:, 0] < math.log(0.5)
 574
 575
 576 ######################################################################
 577
 578 for n_epoch in range(args.nb_epochs):
 579     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 580
 581     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 582     log_string(f"current_test_accuracies {cta}")
 583
 584     ##################################################
 585     # Select, improve, and eval the worst model
 586
 587     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 588
 589     weakest_models = ranked_models[: len(gpus)]
 590
 591     threads = []
 592
 593     for gpu, model in zip(gpus, weakest_models):
 594         log_string(f"training model {model.id}")
 595
 596         t = threading.Thread(
 597             target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
 598         )
 599
 600         threads.append(t)
 601
 602         t.start()
 603
 604     for t in threads:
 605         t.join()
 606
 607     for model in weakest_models:
 608         filename = f"gpt_{model.id:03d}.pth"
 609         torch.save(model.state_dict(), os.path.join(args.result_dir, filename))
 610         log_string(f"wrote {filename}")
 611
 612     ##################################################
 613     # Replace a fraction of the w_quizzes with fresh ones
 614
 615     log_string(
 616         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 617     )
 618
 619     # Renew entirely the train set
 620
 621     for model in weakest_models:
 622         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 623
 624     ##################################################
 625     # If all the models are good enough, generate new quizzes and
 626     # re-compute the test errors
 627
 628     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 629         create_c_quizzes(
 630             models,
 631             quiz_machine,
 632             nb_for_train=nb_new_c_quizzes_for_train,
 633             nb_for_test=nb_new_c_quizzes_for_test,
 634         )
 635
 636         quiz_machine.save_c_quizzes(os.path.join(args.result_dir, "c_quizzes.pth"))
 637
 638 ######################################################################