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("--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("--proba_understands", type=float, default=0.99)
  88
  89 parser.add_argument("--proba_not_understands", type=float, default=0.5)
  90
  91 parser.add_argument("--generation_temperature", type=float, default=2.0)
  92
  93 parser.add_argument("--dirty_debug", action="store_true", default=False)
  94
  95 ######################################################################
  96
  97 grids_tasks = ", ".join(
  98     [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
  99 )
 100
 101 parser.add_argument(
 102     "--grids_tasks",
 103     type=str,
 104     default=None,
 105     help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
 106 )
 107
 108 ######################################################################
 109
 110 parser.add_argument("--sky_height", type=int, default=6)
 111
 112 parser.add_argument("--sky_width", type=int, default=8)
 113
 114 parser.add_argument("--sky_nb_birds", type=int, default=3)
 115
 116 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 117
 118 parser.add_argument("--sky_speed", type=int, default=3)
 119
 120 ######################################################################
 121
 122 args = parser.parse_args()
 123
 124 if args.min_to_validate is None:
 125     args.min_to_validate = args.nb_gpts - 1
 126
 127 if args.max_to_validate is None:
 128     args.max_to_validate = args.nb_gpts - 1
 129
 130 if args.result_dir is None:
 131     args.result_dir = f"results_culture"
 132
 133 ######################################################################
 134
 135 default_args = {
 136     "model": "37M",
 137     "batch_size": 25,
 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 if args.resume:
 196     assert os.path.isdir(args.result_dir)
 197
 198 else:
 199     try:
 200         os.mkdir(args.result_dir)
 201     except FileExistsError:
 202         print(f"result directory {args.result_dir} already exists")
 203         exit(1)
 204
 205 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 206
 207 if args.seed >= 0:
 208     # torch.backends.cudnn.deterministic = True
 209     # torch.backends.cudnn.benchmark = False
 210     # torch.use_deterministic_algorithms(True)
 211     torch.manual_seed(args.seed)
 212     if torch.cuda.is_available():
 213         torch.cuda.manual_seed_all(args.seed)
 214
 215 ######################################################################
 216
 217
 218 def log_string(s):
 219     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 220
 221     if log_file is not None:
 222         log_file.write(t + s + "\n")
 223         log_file.flush()
 224
 225     print(t + s)
 226     sys.stdout.flush()
 227
 228
 229 log_string(f"argv {' '.join(sys.argv)}")
 230
 231 for n in vars(args):
 232     log_string(f"args.{n} {getattr(args, n)}")
 233
 234
 235 ######################################################################
 236
 237 if args.gpus == "all":
 238     gpus_idx = range(torch.cuda.device_count())
 239 else:
 240     gpus_idx = [int(k) for k in args.gpus.split(",")]
 241
 242 gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
 243
 244 if torch.cuda.is_available():
 245     main_device = gpus[0]
 246 else:
 247     assert len(gpus) == 0
 248     main_device = torch.device("cpu")
 249
 250 if args.dirty_debug:
 251     args.nb_train_samples = 2500
 252     args.nb_test_samples = 100
 253
 254 if args.physical_batch_size is None:
 255     args.physical_batch_size = args.batch_size
 256 else:
 257     assert args.batch_size % args.physical_batch_size == 0
 258
 259 assert args.nb_train_samples % args.batch_size == 0
 260 assert args.nb_test_samples % args.batch_size == 0
 261
 262 if args.problem == "sky":
 263     problem = sky.Sky(
 264         height=args.sky_height,
 265         width=args.sky_width,
 266         nb_birds=args.sky_nb_birds,
 267         nb_iterations=args.sky_nb_iterations,
 268         speed=args.sky_speed,
 269         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 270         chunk_size=100,
 271         nb_threads=args.nb_threads,
 272     )
 273     back_accuracy = False
 274 elif args.problem == "grids":
 275     problem = grids.Grids(
 276         max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
 277         chunk_size=100,
 278         nb_threads=args.nb_threads,
 279         tasks=args.grids_tasks,
 280     )
 281     back_accuracy = True
 282 else:
 283     raise ValueError
 284
 285 problem.save_some_examples(args.result_dir)
 286
 287 quiz_machine = quiz_machine.QuizMachine(
 288     problem=problem,
 289     nb_train_samples=args.nb_train_samples,
 290     nb_test_samples=args.nb_test_samples,
 291     back_accuracy=back_accuracy,
 292     batch_size=args.physical_batch_size,
 293     result_dir=args.result_dir,
 294     logger=log_string,
 295     device=main_device,
 296 )
 297
 298 ######################################################################
 299
 300 log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
 301
 302 vocabulary_size = quiz_machine.vocabulary_size()
 303
 304 log_string(f"vocabulary_size {vocabulary_size}")
 305
 306 ######################################################################
 307
 308
 309 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
 310     with torch.autograd.no_grad():
 311         model.eval().to(local_device)
 312
 313         nb_test_samples, acc_test_loss = 0, 0.0
 314         nb_samples_accumulated = 0
 315
 316         for input in quiz_machine.batches(model, split="test"):
 317             input = input.to(local_device)
 318
 319             bs = model(mygpt.BracketedSequence(input))
 320             output = bs.x
 321
 322             loss = F.cross_entropy(output.transpose(1, 2), input)
 323
 324             acc_test_loss += loss.item() * input.size(0)
 325
 326             nb_test_samples += input.size(0)
 327
 328         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 329
 330         log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
 331
 332         model.main_test_accuracy = quiz_machine.produce_results(
 333             n_epoch=n_epoch,
 334             model=model,
 335             result_dir=args.result_dir,
 336             deterministic_synthesis=deterministic_synthesis,
 337         )
 338
 339
 340 def one_epoch(model, quiz_machine, local_device=main_device):
 341     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 342
 343     model.to(local_device).train()
 344
 345     nb_train_samples, acc_train_loss = 0, 0.0
 346
 347     for input in quiz_machine.batches(model, split="train"):
 348         input = input.to(local_device)
 349
 350         if nb_train_samples % args.batch_size == 0:
 351             optimizer.zero_grad()
 352
 353         output = model(mygpt.BracketedSequence(input)).x
 354         loss = F.cross_entropy(output.transpose(1, 2), input)
 355         acc_train_loss += loss.item() * input.size(0)
 356
 357         nb_train_samples += input.size(0)
 358
 359         loss.backward()
 360
 361         if nb_train_samples % args.batch_size == 0:
 362             optimizer.step()
 363
 364     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 365
 366     log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
 367
 368     run_tests(model, quiz_machine, deterministic_synthesis=False)
 369
 370     model.to(main_device)
 371
 372
 373 ######################################################################
 374
 375
 376 def standard_validity(logproba):
 377     l = logproba.sort(dim=-1).values
 378     return (l[:, 0] < math.log(args.proba_not_understands)) & (
 379         l[:, 1] > math.log(args.proba_understands)
 380     )
 381
 382
 383 def valid_c_quizzes(recorded, criteria):
 384     result = [q[criteria(lp)] for q, lp 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_logproba_records = []
 398
 399     nb_to_create = nb_for_train + nb_for_test
 400
 401     # ------------------------------------------------------------
 402
 403     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 404
 405     with open(file_name, "w") as logp_file:
 406         while (
 407             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 408             < nb_to_create
 409         ):
 410             # Select a model at random to generate the new quizzes
 411
 412             model_for_generation = models[torch.randint(len(models), (1,))]
 413
 414             c_quizzes = quiz_machine.generate_quizzes(
 415                 nb_to_create,
 416                 model_for_generation=model_for_generation,
 417                 temperature=args.generation_temperature,
 418             )
 419
 420             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 421
 422             if c_quizzes.size(0) > 0:
 423                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 424                 for l in logproba:
 425                     s = " ".join([str(x.item()) for x in l])
 426                     logp_file.write(s + "\n")
 427                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 428
 429             nb_validated = valid_c_quizzes(
 430                 quizzes_and_logproba_records, standard_validity
 431             ).size(0)
 432
 433             log_string(
 434                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 435             )
 436
 437     # store the new c_quizzes which have been validated
 438
 439     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 440
 441     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 442
 443     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 444     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 445
 446     # save images
 447
 448     q = new_c_quizzes[:72]
 449
 450     if q.size(0) > 0:
 451         quiz_machine.save_quiz_illustrations(
 452             args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q
 453         )
 454
 455
 456 ######################################################################
 457
 458 models = []
 459
 460 for k in range(args.nb_gpts):
 461     log_string(f"creating model {k} and its w_quizzes")
 462     model = mygpt.MyGPT(
 463         vocabulary_size=vocabulary_size,
 464         dim_model=args.dim_model,
 465         dim_keys=args.dim_keys,
 466         dim_hidden=args.dim_hidden,
 467         nb_heads=args.nb_heads,
 468         nb_blocks=args.nb_blocks,
 469         causal=True,
 470         dropout=args.dropout,
 471     ).to(main_device)
 472
 473     model.main_test_accuracy = 0.0
 474     model.id = k
 475
 476     model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
 477     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 478     model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
 479     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 480
 481     models.append(model)
 482
 483 ######################################################################
 484
 485 if args.resume:
 486     try:
 487         for model in models:
 488             filename = f"gpt_{model.id:03d}.pth"
 489
 490             try:
 491                 model.load_state_dict(
 492                     torch.load(os.path.join(args.result_dir, filename))
 493                 )
 494                 log_string(f"successfully loaded {filename}")
 495             except FileNotFoundError:
 496                 log_string(f"cannot find {filename}")
 497                 pass
 498
 499         try:
 500             filename = "c_quizzes.pth"
 501             quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
 502             log_string(f"successfully loaded {filename}")
 503         except FileNotFoundError:
 504             log_string(f"cannot find {filename}")
 505             pass
 506
 507     except:
 508         log_string(f"error when loading {filename}.")
 509         exit(1)
 510
 511 ######################################################################
 512
 513 nb_parameters = sum(p.numel() for p in models[0].parameters())
 514 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 515
 516 ######################################################################
 517
 518 # Compute the entropy of the training tokens
 519
 520 token_count = 0
 521 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 522     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 523         (0, 1)
 524     )
 525 token_probas = token_count / token_count.sum()
 526 entropy = -torch.xlogy(token_probas, token_probas).sum()
 527 train_set_perplexity = math.exp(entropy)
 528
 529 ######################################################################
 530 # A bit of paranoia never hurts
 531
 532 if args.max_percents_of_test_in_train >= 0:
 533
 534     def subsets_as_tuples(batches, cs):
 535         s = set()
 536         for batch in batches:
 537             for x in batch:
 538                 s.add(tuple([v.item() for v in x]))
 539                 if len(s) == cs:
 540                     yield s
 541                     s = set()
 542         yield s
 543
 544     nb_test, nb_in_train = 0, 0
 545     for test_subset in subsets_as_tuples(
 546         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 547     ):
 548         in_train = set()
 549         for train_subset in subsets_as_tuples(
 550             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 551         ):
 552             in_train.update(test_subset.intersection(train_subset))
 553         nb_in_train += len(in_train)
 554         nb_test += len(test_subset)
 555
 556     log_string(
 557         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 558     )
 559
 560     assert (
 561         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 562     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 563
 564 ######################################################################
 565
 566 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 567 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 568
 569 log_string(
 570     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}"
 571 )
 572
 573 ######################################################################
 574
 575 if args.dirty_debug:
 576     args.accuracy_to_make_c_quizzes = 0.0
 577     args.nb_gpts = 2
 578     nb_new_c_quizzes_for_train = 100
 579     nb_new_c_quizzes_for_test = 10
 580
 581     def standard_validity(logproba):
 582         l = logproba.sort(dim=-1).values
 583         return l[:, 0] < math.log(0.5)
 584
 585
 586 ######################################################################
 587
 588 for n_epoch in range(args.nb_epochs):
 589     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 590
 591     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 592     log_string(f"current_test_accuracies {cta}")
 593
 594     ##################################################
 595     # Select, improve, and eval the worst model
 596
 597     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 598
 599     weakest_models = ranked_models[: len(gpus)]
 600
 601     threads = []
 602
 603     for gpu, model in zip(gpus, weakest_models):
 604         log_string(f"training model {model.id}")
 605
 606         t = threading.Thread(
 607             target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
 608         )
 609
 610         threads.append(t)
 611
 612         t.start()
 613
 614     for t in threads:
 615         t.join()
 616
 617     for model in weakest_models:
 618         filename = f"gpt_{model.id:03d}.pth"
 619         torch.save(model.state_dict(), os.path.join(args.result_dir, filename))
 620         log_string(f"wrote {filename}")
 621
 622     ##################################################
 623     # Replace a fraction of the w_quizzes with fresh ones
 624
 625     log_string(
 626         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 627     )
 628
 629     # Renew entirely the train set
 630
 631     for model in weakest_models:
 632         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 633
 634     ##################################################
 635     # If all the models are good enough, generate new quizzes and
 636     # re-compute the test errors
 637
 638     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 639         create_c_quizzes(
 640             models,
 641             quiz_machine,
 642             nb_for_train=nb_new_c_quizzes_for_train,
 643             nb_for_test=nb_new_c_quizzes_for_test,
 644         )
 645
 646         quiz_machine.save_c_quizzes(os.path.join(args.result_dir, "c_quizzes.pth"))
 647
 648 ######################################################################