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