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, tasks, problems
  16
  17 ######################################################################
  18
  19 if torch.cuda.is_available():
  20     device = torch.device("cuda")
  21     torch.backends.cuda.matmul.allow_tf32 = True
  22 else:
  23     device = torch.device("cpu")
  24
  25 ######################################################################
  26
  27 parser = argparse.ArgumentParser(
  28     description="An implementation of GPT with cache.",
  29     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  30 )
  31
  32 parser.add_argument("--task", type=str, default="world", help="world")
  33
  34 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  35
  36 parser.add_argument("--result_dir", type=str, default=None)
  37
  38 parser.add_argument("--seed", type=int, default=0)
  39
  40 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  41
  42 ########################################
  43
  44 parser.add_argument("--nb_epochs", type=int, default=10000)
  45
  46 parser.add_argument("--batch_size", type=int, default=None)
  47
  48 parser.add_argument("--physical_batch_size", type=int, default=None)
  49
  50 parser.add_argument("--nb_train_samples", type=int, default=None)
  51
  52 parser.add_argument("--nb_test_samples", type=int, default=None)
  53
  54 parser.add_argument("--learning_rate", type=float, default=1e-4)
  55
  56 ########################################
  57
  58 parser.add_argument("--model", type=str, default=None)
  59
  60 parser.add_argument("--dim_model", type=int, default=None)
  61
  62 parser.add_argument("--dim_keys", type=int, default=None)
  63
  64 parser.add_argument("--dim_hidden", type=int, default=None)
  65
  66 parser.add_argument("--nb_heads", type=int, default=None)
  67
  68 parser.add_argument("--nb_blocks", type=int, default=None)
  69
  70 parser.add_argument("--dropout", type=float, default=0.1)
  71
  72 ########################################
  73
  74 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  75
  76 parser.add_argument("--check", action="store_true", default=False)
  77
  78 ######################################################################
  79
  80 args = parser.parse_args()
  81
  82 if args.result_dir is None:
  83     args.result_dir = f"results_{args.task}"
  84
  85 ######################################################################
  86
  87 default_task_args = {
  88     "world": {
  89         "model": "37M",
  90         "batch_size": 100,
  91         "nb_train_samples": 250000,
  92         "nb_test_samples": 10000,
  93     },
  94 }
  95
  96 if args.task in default_task_args:
  97     for k, v in default_task_args[args.task].items():
  98         if getattr(args, k) is None:
  99             setattr(args, k, v)
 100
 101 ######################################################################
 102
 103 default_model_args = {
 104     "17K": {
 105         "dim_model": 32,
 106         "dim_keys": 32,
 107         "dim_hidden": 32,
 108         "nb_heads": 2,
 109         "nb_blocks": 2,
 110     },
 111     "4M": {
 112         "dim_model": 256,
 113         "dim_keys": 32,
 114         "dim_hidden": 1024,
 115         "nb_heads": 4,
 116         "nb_blocks": 6,
 117     },
 118     "37M": {
 119         "dim_model": 512,
 120         "dim_keys": 64,
 121         "dim_hidden": 2048,
 122         "nb_heads": 8,
 123         "nb_blocks": 12,
 124     },
 125     "122M": {
 126         "dim_model": 768,
 127         "dim_keys": 64,
 128         "dim_hidden": 2048,
 129         "nb_heads": 8,
 130         "nb_blocks": 24,
 131     },
 132     "352M": {
 133         "dim_model": 1024,
 134         "dim_keys": 64,
 135         "dim_hidden": 2048,
 136         "nb_heads": 8,
 137         "nb_blocks": 48,
 138     },
 139 }
 140
 141 if args.model in default_model_args:
 142     for k, v in default_model_args[args.model].items():
 143         if getattr(args, k) is None:
 144             setattr(args, k, v)
 145 else:
 146     raise ValueError(f"Unknown model {args.model}")
 147
 148 ######################################################################
 149
 150 try:
 151     os.mkdir(args.result_dir)
 152 except FileExistsError:
 153     print(f"result directory {args.result_dir} already exists")
 154     exit(1)
 155
 156 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 157
 158 if args.seed >= 0:
 159     # torch.backends.cudnn.deterministic = True
 160     # torch.backends.cudnn.benchmark = False
 161     # torch.use_deterministic_algorithms(True)
 162     torch.manual_seed(args.seed)
 163     if torch.cuda.is_available():
 164         torch.cuda.manual_seed_all(args.seed)
 165
 166 ######################################################################
 167
 168
 169 def log_string(s):
 170     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 171
 172     if log_file is not None:
 173         log_file.write(t + s + "\n")
 174         log_file.flush()
 175
 176     print(t + s)
 177     sys.stdout.flush()
 178
 179
 180 log_string(f"argv {' '.join(sys.argv)}")
 181
 182 for n in vars(args):
 183     log_string(f"args.{n} {getattr(args, n)}")
 184
 185
 186 ######################################################################
 187
 188 if args.test:
 189     args.nb_train_samples = 1000
 190     args.nb_test_samples = 25
 191
 192 if args.physical_batch_size is None:
 193     args.physical_batch_size = args.batch_size
 194 else:
 195     assert args.batch_size % args.physical_batch_size == 0
 196
 197 assert args.nb_train_samples % args.batch_size == 0
 198 assert args.nb_test_samples % args.batch_size == 0
 199
 200 if args.task == "file":
 201     assert (
 202         args.filetask_train_file is not None and args.filetask_test_file is not None
 203     ), "You have to specify the task train and test files"
 204     task = tasks.TaskFromFile(
 205         args.filetask_train_file,
 206         args.filetask_test_file,
 207         nb_train_samples=args.nb_train_samples,
 208         nb_test_samples=args.nb_test_samples,
 209         batch_size=args.physical_batch_size,
 210         shuffle=True,
 211         device=device,
 212     )
 213     args.max_percents_of_test_in_train = 0
 214
 215 elif args.task == "byheart":
 216     task = tasks.SandBox(
 217         problem=problems.ProblemByHeart(separation=args.byheart_separation),
 218         nb_train_samples=args.nb_train_samples,
 219         nb_test_samples=args.nb_test_samples,
 220         batch_size=args.physical_batch_size,
 221         logger=log_string,
 222         device=device,
 223     )
 224     args.max_percents_of_test_in_train = -1
 225
 226 elif args.task == "world":
 227     task = tasks.World(
 228         nb_train_samples=args.nb_train_samples,
 229         nb_test_samples=args.nb_test_samples,
 230         batch_size=args.physical_batch_size,
 231         result_dir=args.result_dir,
 232         logger=log_string,
 233         device=device,
 234     )
 235     args.max_percents_of_test_in_train = -1
 236
 237 elif args.task == "learnop":
 238     task = tasks.SandBox(
 239         problem=problems.ProblemLearnOperator(),
 240         nb_train_samples=args.nb_train_samples,
 241         nb_test_samples=args.nb_test_samples,
 242         batch_size=args.physical_batch_size,
 243         logger=log_string,
 244         device=device,
 245     )
 246
 247
 248 elif args.task == "guessop":
 249     task = tasks.SandBox(
 250         problem=problems.ProblemGuessOperator(),
 251         nb_train_samples=args.nb_train_samples,
 252         nb_test_samples=args.nb_test_samples,
 253         batch_size=args.physical_batch_size,
 254         logger=log_string,
 255         device=device,
 256     )
 257
 258
 259 elif args.task == "twotargets":
 260     task = tasks.SandBox(
 261         problem=problems.ProblemTwoTargets(),
 262         nb_train_samples=args.nb_train_samples,
 263         nb_test_samples=args.nb_test_samples,
 264         batch_size=args.physical_batch_size,
 265         logger=log_string,
 266         device=device,
 267     )
 268
 269 elif args.task == "memory":
 270     task = tasks.SandBox(
 271         problem=problems.ProblemMemory(),
 272         nb_train_samples=args.nb_train_samples,
 273         nb_test_samples=args.nb_test_samples,
 274         batch_size=args.physical_batch_size,
 275         logger=log_string,
 276         device=device,
 277     )
 278
 279 elif args.task == "mixing":
 280     task = tasks.SandBox(
 281         problem=problems.ProblemMixing(
 282             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 283         ),
 284         nb_train_samples=args.nb_train_samples,
 285         nb_test_samples=args.nb_test_samples,
 286         batch_size=args.physical_batch_size,
 287         logger=log_string,
 288         device=device,
 289     )
 290
 291 elif args.task == "addition":
 292     task = tasks.SandBox(
 293         problem=problems.ProblemAddition(),
 294         nb_train_samples=args.nb_train_samples,
 295         nb_test_samples=args.nb_test_samples,
 296         batch_size=args.physical_batch_size,
 297         logger=log_string,
 298         device=device,
 299     )
 300
 301 elif args.task == "picoclvr":
 302     task = tasks.PicoCLVR(
 303         nb_train_samples=args.nb_train_samples,
 304         nb_test_samples=args.nb_test_samples,
 305         batch_size=args.physical_batch_size,
 306         height=args.picoclvr_height,
 307         width=args.picoclvr_width,
 308         nb_colors=args.picoclvr_nb_colors,
 309         logger=log_string,
 310         device=device,
 311         pruner_train=picoclvr_pruner_train,
 312         pruner_eval=picoclvr_pruner_eval,
 313     )
 314
 315 elif args.task == "mnist":
 316     task = tasks.MNIST(
 317         nb_train_samples=args.nb_train_samples,
 318         nb_test_samples=args.nb_test_samples,
 319         batch_size=args.physical_batch_size,
 320         device=device,
 321     )
 322
 323 elif args.task == "maze":
 324     task = tasks.Maze(
 325         nb_train_samples=args.nb_train_samples,
 326         nb_test_samples=args.nb_test_samples,
 327         batch_size=args.physical_batch_size,
 328         height=args.maze_height,
 329         width=args.maze_width,
 330         nb_walls=args.maze_nb_walls,
 331         device="cpu",
 332     )
 333
 334 elif args.task == "snake":
 335     task = tasks.Snake(
 336         nb_train_samples=args.nb_train_samples,
 337         nb_test_samples=args.nb_test_samples,
 338         batch_size=args.physical_batch_size,
 339         height=args.snake_height,
 340         width=args.snake_width,
 341         nb_colors=args.snake_nb_colors,
 342         length=args.snake_length,
 343         prompt_length=args.snake_length // 2,
 344         device=device,
 345     )
 346
 347 elif args.task == "stack":
 348     task = tasks.Stack(
 349         nb_train_samples=args.nb_train_samples,
 350         nb_test_samples=args.nb_test_samples,
 351         batch_size=args.physical_batch_size,
 352         logger=log_string,
 353         nb_steps=args.stack_nb_steps,
 354         nb_stacks=args.stack_nb_stacks,
 355         nb_digits=args.stack_nb_digits,
 356         fraction_values_for_train=args.stack_fraction_values_for_train,
 357         device=device,
 358     )
 359
 360 elif args.task == "expr":
 361     task = tasks.Expr(
 362         nb_train_samples=args.nb_train_samples,
 363         nb_test_samples=args.nb_test_samples,
 364         nb_variables=args.expr_nb_variables,
 365         sequence_length=args.expr_sequence_length,
 366         operand_max=args.expr_operand_max,
 367         result_max=args.expr_result_max,
 368         batch_size=args.physical_batch_size,
 369         device=device,
 370     )
 371
 372 elif args.task == "rpl":
 373     task = tasks.RPL(
 374         nb_train_samples=args.nb_train_samples,
 375         nb_test_samples=args.nb_test_samples,
 376         batch_size=args.physical_batch_size,
 377         nb_starting_values=args.rpl_nb_starting_values,
 378         max_input=args.rpl_max_input,
 379         prog_len=args.rpl_prog_len,
 380         nb_runs=args.rpl_nb_runs,
 381         no_prog=args.rpl_no_prog,
 382         logger=log_string,
 383         device=device,
 384     )
 385
 386 elif args.task == "grid":
 387     task = tasks.Grid(
 388         nb_train_samples=args.nb_train_samples,
 389         nb_test_samples=args.nb_test_samples,
 390         batch_size=args.physical_batch_size,
 391         size=args.grid_size,
 392         fraction_play=args.grid_fraction_play,
 393         logger=log_string,
 394         device=device,
 395     )
 396
 397 elif args.task == "qmlp":
 398     task = tasks.QMLP(
 399         nb_train_samples=args.nb_train_samples,
 400         nb_test_samples=args.nb_test_samples,
 401         batch_size=args.physical_batch_size,
 402         result_dir=args.result_dir,
 403         logger=log_string,
 404         device=device,
 405     )
 406
 407 elif args.task == "greed":
 408     task = tasks.Greed(
 409         nb_train_samples=args.nb_train_samples,
 410         nb_test_samples=args.nb_test_samples,
 411         batch_size=args.physical_batch_size,
 412         height=args.greed_height,
 413         width=args.greed_width,
 414         T=args.greed_T,
 415         nb_walls=args.greed_nb_walls,
 416         nb_coins=args.greed_nb_coins,
 417         logger=log_string,
 418         device=device,
 419     )
 420
 421 else:
 422     raise ValueError(f"Unknown task {args.task}")
 423
 424 ######################################################################
 425
 426 log_string(f"device {device}")
 427
 428 vocabulary_size = task.vocabulary_size()
 429
 430 log_string(f"vocabulary_size {vocabulary_size}")
 431
 432 ######################################################################
 433
 434 # Compute the entropy of the training tokens
 435
 436 token_count = 0
 437 for input in task.batches(split="train", desc="train-entropy"):
 438     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 439 token_probas = token_count / token_count.sum()
 440 entropy = -torch.xlogy(token_probas, token_probas).sum()
 441 train_set_perplexity = math.exp(entropy)
 442
 443 ######################################################################
 444 # A bit of paranoia never hurts
 445
 446 if args.max_percents_of_test_in_train >= 0:
 447
 448     def subsets_as_tuples(batches, cs):
 449         s = set()
 450         for batch in batches:
 451             for x in batch:
 452                 s.add(tuple([v.item() for v in x]))
 453                 if len(s) == cs:
 454                     yield s
 455                     s = set()
 456         yield s
 457
 458     nb_test, nb_in_train = 0, 0
 459     for test_subset in subsets_as_tuples(
 460         task.batches(split="test", desc="test-check"), 25000
 461     ):
 462         in_train = set()
 463         for train_subset in subsets_as_tuples(
 464             task.batches(split="train", desc="train-check"), 25000
 465         ):
 466             in_train.update(test_subset.intersection(train_subset))
 467         nb_in_train += len(in_train)
 468         nb_test += len(test_subset)
 469
 470     log_string(
 471         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 472     )
 473
 474     assert (
 475         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 476     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 477
 478 ##############################
 479
 480
 481 def one_epoch(model, task):
 482     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 483
 484     model.train()
 485
 486     nb_train_samples, acc_train_loss = 0, 0.0
 487
 488     for input in task.batches(split="train"):
 489         input = input.to(device)
 490
 491         if nb_train_samples % args.batch_size == 0:
 492             optimizer.zero_grad()
 493
 494         output = model(mygpt.BracketedSequence(input)).x
 495         loss = F.cross_entropy(output.transpose(1, 2), input)
 496         acc_train_loss += loss.item() * input.size(0)
 497
 498         nb_train_samples += input.size(0)
 499
 500         loss.backward()
 501
 502         if nb_train_samples % args.batch_size == 0:
 503             optimizer.step()
 504
 505     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 506
 507     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 508
 509
 510 ######################################################################
 511
 512
 513 def run_tests(model, task, deterministic_synthesis):
 514     with torch.autograd.no_grad():
 515         model.eval()
 516
 517         nb_test_samples, acc_test_loss = 0, 0.0
 518         nb_samples_accumulated = 0
 519
 520         for input in task.batches(split="test"):
 521             input = input.to(device)
 522
 523             bs = model(mygpt.BracketedSequence(input))
 524             output = bs.x
 525
 526             loss = F.cross_entropy(output.transpose(1, 2), input)
 527
 528             acc_test_loss += loss.item() * input.size(0)
 529
 530             nb_test_samples += input.size(0)
 531
 532         main_test_accuracy = task.produce_results(
 533             n_epoch=n_epoch,
 534             model=model,
 535             result_dir=args.result_dir,
 536             logger=log_string,
 537             deterministic_synthesis=deterministic_synthesis,
 538         )
 539
 540         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 541
 542         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 543
 544     model.main_test_accuracy = main_test_accuracy
 545
 546
 547 ######################################################################
 548
 549
 550 def create_quizzes(
 551     model,
 552     other_models,
 553     task,
 554     nb_for_train=1000,
 555     nb_for_test=100,
 556 ):
 557     kept = []
 558
 559     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 560         new_quizzes, nb_correct = task.create_new_quizzes(
 561             n_epoch=n_epoch,
 562             result_dir=args.result_dir,
 563             logger=log_string,
 564             nb=4 * (nb_for_train + nb_for_test),
 565             model=model,
 566             other_models=other_models,
 567         )
 568
 569         to_keep = new_quizzes[nb_correct == len(other_models) - 1]
 570         log_string(f"keep {to_keep.size(0)} quizzes")
 571         kept.append(to_keep)
 572
 573     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 574
 575     task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
 576     task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
 577
 578     task.save_image(
 579         new_quizzes[:96],
 580         args.result_dir,
 581         f"world_new_{n_epoch:04d}_{model.id:02d}.png",
 582         log_string,
 583     )
 584
 585
 586 ######################################################################
 587
 588 models = []
 589
 590 for k in range(5):
 591     model = mygpt.MyGPT(
 592         vocabulary_size=vocabulary_size,
 593         dim_model=args.dim_model,
 594         dim_keys=args.dim_keys,
 595         dim_hidden=args.dim_hidden,
 596         nb_heads=args.nb_heads,
 597         nb_blocks=args.nb_blocks,
 598         causal=True,
 599         dropout=args.dropout,
 600     ).to(device)
 601
 602     model.main_test_accuracy = 0.0
 603     model.id = k
 604
 605     models.append(model)
 606
 607
 608 nb_parameters = sum(p.numel() for p in models[0].parameters())
 609 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 610
 611 ######################################################################
 612
 613 accuracy_to_make_quizzes = 0.975
 614 nb_new_quizzes_for_train = 1000
 615 nb_new_quizzes_for_test = 100
 616
 617 if args.test:
 618     accuracy_to_make_quizzes = 0.0
 619     nb_new_quizzes_for_train = 10
 620     nb_new_quizzes_for_test = 10
 621
 622 for n_epoch in range(args.nb_epochs):
 623     # select the model with lowest accuracy
 624     models.sort(key=lambda model: model.main_test_accuracy)
 625     model = models[0]
 626
 627     log_string(
 628         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 629     )
 630
 631     # improve it
 632     one_epoch(model, task)
 633
 634     log_string(
 635         f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
 636     )
 637
 638     # test it
 639     run_tests(model, task, deterministic_synthesis=False)
 640
 641     if model.main_test_accuracy >= accuracy_to_make_quizzes:
 642         other_models = models.copy()
 643         other_models.remove(model)
 644
 645         create_quizzes(
 646             model,
 647             other_models,
 648             task,
 649             nb_for_train=nb_new_quizzes_for_train,
 650             nb_for_test=nb_new_quizzes_for_test,
 651         )
 652
 653
 654 ######################################################################