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(
  33     "--task",
  34     type=str,
  35     default="world",
  36     help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
  37 )
  38
  39 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  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=1e-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 ##############################
  82 # filetask
  83
  84 parser.add_argument("--filetask_train_file", type=str, default=None)
  85
  86 parser.add_argument("--filetask_test_file", type=str, default=None)
  87
  88 ##############################
  89 # rpl options
  90
  91 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
  92
  93 parser.add_argument("--rpl_max_input", type=int, default=9)
  94
  95 parser.add_argument("--rpl_prog_len", type=int, default=8)
  96
  97 parser.add_argument("--rpl_nb_runs", type=int, default=5)
  98
  99 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 100
 101 ##############################
 102 # grid options
 103
 104 parser.add_argument("--grid_size", type=int, default=6)
 105
 106 parser.add_argument("--grid_fraction_play", type=float, default=0)
 107
 108 ##############################
 109 # picoclvr options
 110
 111 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 112
 113 parser.add_argument("--picoclvr_height", type=int, default=12)
 114
 115 parser.add_argument("--picoclvr_width", type=int, default=16)
 116
 117 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 118
 119 ##############################
 120 # Maze options
 121
 122 parser.add_argument("--maze_height", type=int, default=13)
 123
 124 parser.add_argument("--maze_width", type=int, default=21)
 125
 126 parser.add_argument("--maze_nb_walls", type=int, default=15)
 127
 128 ##############################
 129 # Snake options
 130
 131 parser.add_argument("--snake_height", type=int, default=9)
 132
 133 parser.add_argument("--snake_width", type=int, default=12)
 134
 135 parser.add_argument("--snake_nb_colors", type=int, default=5)
 136
 137 parser.add_argument("--snake_length", type=int, default=200)
 138
 139 ##############################
 140 # ByHeart options
 141
 142 parser.add_argument("--byheart_separation", type=int, default=1)
 143
 144 ##############################
 145 # Stack options
 146
 147 parser.add_argument("--stack_nb_steps", type=int, default=100)
 148
 149 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 150
 151 parser.add_argument("--stack_nb_digits", type=int, default=3)
 152
 153 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 154
 155 ##############################
 156 # Expr options
 157
 158 parser.add_argument("--expr_nb_variables", type=int, default=5)
 159
 160 parser.add_argument("--expr_sequence_length", type=int, default=40)
 161
 162 parser.add_argument("--expr_operand_max", type=int, default=9)
 163
 164 parser.add_argument("--expr_result_max", type=int, default=99)
 165
 166 parser.add_argument("--expr_input_file", type=str, default=None)
 167
 168 ##############################
 169 # Mixing
 170
 171 parser.add_argument("--mixing_hard", action="store_true", default=False)
 172
 173 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 174
 175 ##############################
 176 # greed options
 177
 178 parser.add_argument("--greed_height", type=int, default=5)
 179
 180 parser.add_argument("--greed_width", type=int, default=7)
 181
 182 parser.add_argument("--greed_T", type=int, default=25)
 183
 184 parser.add_argument("--greed_nb_walls", type=int, default=5)
 185
 186 parser.add_argument("--greed_nb_coins", type=int, default=2)
 187
 188 ######################################################################
 189
 190 args = parser.parse_args()
 191
 192 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 193
 194 if args.result_dir is None:
 195     args.result_dir = f"results_{args.task}"
 196
 197 ######################################################################
 198
 199 default_task_args = {
 200     "world": {
 201         "model": "37M",
 202         "batch_size": 100,
 203         "nb_train_samples": 250000,
 204         "nb_test_samples": 10000,
 205     },
 206     "file": {
 207         "model": "37M",
 208         "batch_size": 25,
 209         "nb_train_samples": 250000,
 210         "nb_test_samples": 10000,
 211     },
 212     "addition": {
 213         "model": "352M",
 214         "batch_size": 25,
 215         "nb_train_samples": 250000,
 216         "nb_test_samples": 10000,
 217     },
 218     "byheart": {
 219         "model": "37M",
 220         "batch_size": 25,
 221         "nb_train_samples": 50000,
 222         "nb_test_samples": 10000,
 223     },
 224     "expr": {
 225         "model": "352M",
 226         "batch_size": 25,
 227         "nb_train_samples": 2500000,
 228         "nb_test_samples": 10000,
 229     },
 230     "grid": {
 231         "model": "37M",
 232         "batch_size": 25,
 233         "nb_train_samples": 250000,
 234         "nb_test_samples": 10000,
 235     },
 236     "qmlp": {
 237         "model": "37M",
 238         "batch_size": 10,
 239         "nb_train_samples": 100000,
 240         "nb_test_samples": 1000,
 241     },
 242     "guessop": {
 243         "model": "352M",
 244         "batch_size": 25,
 245         "nb_train_samples": 1000000,
 246         "nb_test_samples": 10000,
 247     },
 248     "learnop": {
 249         "model": "37M",
 250         "batch_size": 25,
 251         "nb_train_samples": 50000,
 252         "nb_test_samples": 10000,
 253     },
 254     "maze": {
 255         "model": "37M",
 256         "batch_size": 5,
 257         "nb_train_samples": 100000,
 258         "nb_test_samples": 10000,
 259     },
 260     "picoclvr": {
 261         "model": "37M",
 262         "batch_size": 25,
 263         "nb_train_samples": 250000,
 264         "nb_test_samples": 10000,
 265     },
 266     "rpl": {
 267         "model": "352M",
 268         "batch_size": 5,
 269         "nb_train_samples": 2500000,
 270         "nb_test_samples": 10000,
 271     },
 272     "snake": {
 273         "model": "37M",
 274         "batch_size": 25,
 275         "nb_train_samples": 250000,
 276         "nb_test_samples": 10000,
 277     },
 278     "stack": {
 279         "model": "37M",
 280         "batch_size": 25,
 281         "nb_train_samples": 100000,
 282         "nb_test_samples": 1000,
 283     },
 284     "twotargets": {
 285         "model": "37M",
 286         "batch_size": 25,
 287         "nb_train_samples": 50000,
 288         "nb_test_samples": 10000,
 289     },
 290     "memory": {
 291         "model": "37M",
 292         "batch_size": 100,
 293         "nb_train_samples": 25000,
 294         "nb_test_samples": 1000,
 295     },
 296     "mixing": {
 297         "model": "37M",
 298         "batch_size": 25,
 299         "nb_train_samples": 250000,
 300         "nb_test_samples": 10000,
 301     },
 302     "mnist": {
 303         "model": "37M",
 304         "batch_size": 10,
 305         "nb_train_samples": 60000,
 306         "nb_test_samples": 10000,
 307     },
 308     "greed": {
 309         "model": "37M",
 310         "batch_size": 25,
 311         "nb_train_samples": 25000,
 312         "nb_test_samples": 10000,
 313     },
 314 }
 315
 316 if args.task in default_task_args:
 317     for k, v in default_task_args[args.task].items():
 318         if getattr(args, k) is None:
 319             setattr(args, k, v)
 320
 321 ######################################################################
 322
 323 default_model_args = {
 324     "17K": {
 325         "dim_model": 32,
 326         "dim_keys": 32,
 327         "dim_hidden": 32,
 328         "nb_heads": 2,
 329         "nb_blocks": 2,
 330     },
 331     "4M": {
 332         "dim_model": 256,
 333         "dim_keys": 32,
 334         "dim_hidden": 1024,
 335         "nb_heads": 4,
 336         "nb_blocks": 6,
 337     },
 338     "37M": {
 339         "dim_model": 512,
 340         "dim_keys": 64,
 341         "dim_hidden": 2048,
 342         "nb_heads": 8,
 343         "nb_blocks": 12,
 344     },
 345     "122M": {
 346         "dim_model": 768,
 347         "dim_keys": 64,
 348         "dim_hidden": 2048,
 349         "nb_heads": 8,
 350         "nb_blocks": 24,
 351     },
 352     "352M": {
 353         "dim_model": 1024,
 354         "dim_keys": 64,
 355         "dim_hidden": 2048,
 356         "nb_heads": 8,
 357         "nb_blocks": 48,
 358     },
 359 }
 360
 361 if args.model in default_model_args:
 362     for k, v in default_model_args[args.model].items():
 363         if getattr(args, k) is None:
 364             setattr(args, k, v)
 365 else:
 366     raise ValueError(f"Unknown model {args.model}")
 367
 368 ######################################################################
 369
 370 try:
 371     os.mkdir(args.result_dir)
 372 except FileExistsError:
 373     print(f"result directory {args.result_dir} already exists")
 374     exit(1)
 375
 376 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 377
 378 if args.seed >= 0:
 379     # torch.backends.cudnn.deterministic = True
 380     # torch.backends.cudnn.benchmark = False
 381     # torch.use_deterministic_algorithms(True)
 382     torch.manual_seed(args.seed)
 383     if torch.cuda.is_available():
 384         torch.cuda.manual_seed_all(args.seed)
 385
 386 ######################################################################
 387
 388
 389 def log_string(s):
 390     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 391
 392     if log_file is not None:
 393         log_file.write(t + s + "\n")
 394         log_file.flush()
 395
 396     print(t + s)
 397     sys.stdout.flush()
 398
 399
 400 log_string(f"argv {' '.join(sys.argv)}")
 401
 402 for n in vars(args):
 403     log_string(f"args.{n} {getattr(args, n)}")
 404
 405
 406 ######################################################################
 407
 408
 409 def picoclvr_pruner_horizontal_green(p):
 410     return not ("green" in p and ("left" in p or "right" in p))
 411
 412
 413 picoclvr_pruner_train = (
 414     picoclvr_pruner_horizontal_green
 415     if args.picocvlr_prune_properties in {"train+eval"}
 416     else None
 417 )
 418
 419 picoclvr_pruner_eval = (
 420     (lambda p: not picoclvr_pruner_horizontal_green(p))
 421     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 422     else None
 423 )
 424
 425 ######################################################################
 426
 427 if args.physical_batch_size is None:
 428     args.physical_batch_size = args.batch_size
 429 else:
 430     assert args.batch_size % args.physical_batch_size == 0
 431
 432 assert args.nb_train_samples % args.batch_size == 0
 433 assert args.nb_test_samples % args.batch_size == 0
 434
 435 if args.task == "file":
 436     assert (
 437         args.filetask_train_file is not None and args.filetask_test_file is not None
 438     ), "You have to specify the task train and test files"
 439     task = tasks.TaskFromFile(
 440         args.filetask_train_file,
 441         args.filetask_test_file,
 442         nb_train_samples=args.nb_train_samples,
 443         nb_test_samples=args.nb_test_samples,
 444         batch_size=args.physical_batch_size,
 445         shuffle=True,
 446         device=device,
 447     )
 448     args.max_percents_of_test_in_train = 0
 449
 450 elif args.task == "byheart":
 451     task = tasks.SandBox(
 452         problem=problems.ProblemByHeart(separation=args.byheart_separation),
 453         nb_train_samples=args.nb_train_samples,
 454         nb_test_samples=args.nb_test_samples,
 455         batch_size=args.physical_batch_size,
 456         logger=log_string,
 457         device=device,
 458     )
 459     args.max_percents_of_test_in_train = -1
 460
 461 elif args.task == "world":
 462     task = tasks.World(
 463         nb_train_samples=args.nb_train_samples,
 464         nb_test_samples=args.nb_test_samples,
 465         batch_size=args.physical_batch_size,
 466         result_dir=args.result_dir,
 467         logger=log_string,
 468         device=device,
 469     )
 470     args.max_percents_of_test_in_train = -1
 471
 472 elif args.task == "learnop":
 473     task = tasks.SandBox(
 474         problem=problems.ProblemLearnOperator(),
 475         nb_train_samples=args.nb_train_samples,
 476         nb_test_samples=args.nb_test_samples,
 477         batch_size=args.physical_batch_size,
 478         logger=log_string,
 479         device=device,
 480     )
 481
 482
 483 elif args.task == "guessop":
 484     task = tasks.SandBox(
 485         problem=problems.ProblemGuessOperator(),
 486         nb_train_samples=args.nb_train_samples,
 487         nb_test_samples=args.nb_test_samples,
 488         batch_size=args.physical_batch_size,
 489         logger=log_string,
 490         device=device,
 491     )
 492
 493
 494 elif args.task == "twotargets":
 495     task = tasks.SandBox(
 496         problem=problems.ProblemTwoTargets(),
 497         nb_train_samples=args.nb_train_samples,
 498         nb_test_samples=args.nb_test_samples,
 499         batch_size=args.physical_batch_size,
 500         logger=log_string,
 501         device=device,
 502     )
 503
 504 elif args.task == "memory":
 505     task = tasks.SandBox(
 506         problem=problems.ProblemMemory(),
 507         nb_train_samples=args.nb_train_samples,
 508         nb_test_samples=args.nb_test_samples,
 509         batch_size=args.physical_batch_size,
 510         logger=log_string,
 511         device=device,
 512     )
 513
 514 elif args.task == "mixing":
 515     task = tasks.SandBox(
 516         problem=problems.ProblemMixing(
 517             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 518         ),
 519         nb_train_samples=args.nb_train_samples,
 520         nb_test_samples=args.nb_test_samples,
 521         batch_size=args.physical_batch_size,
 522         logger=log_string,
 523         device=device,
 524     )
 525
 526 elif args.task == "addition":
 527     task = tasks.SandBox(
 528         problem=problems.ProblemAddition(),
 529         nb_train_samples=args.nb_train_samples,
 530         nb_test_samples=args.nb_test_samples,
 531         batch_size=args.physical_batch_size,
 532         logger=log_string,
 533         device=device,
 534     )
 535
 536 elif args.task == "picoclvr":
 537     task = tasks.PicoCLVR(
 538         nb_train_samples=args.nb_train_samples,
 539         nb_test_samples=args.nb_test_samples,
 540         batch_size=args.physical_batch_size,
 541         height=args.picoclvr_height,
 542         width=args.picoclvr_width,
 543         nb_colors=args.picoclvr_nb_colors,
 544         logger=log_string,
 545         device=device,
 546         pruner_train=picoclvr_pruner_train,
 547         pruner_eval=picoclvr_pruner_eval,
 548     )
 549
 550 elif args.task == "mnist":
 551     task = tasks.MNIST(
 552         nb_train_samples=args.nb_train_samples,
 553         nb_test_samples=args.nb_test_samples,
 554         batch_size=args.physical_batch_size,
 555         device=device,
 556     )
 557
 558 elif args.task == "maze":
 559     task = tasks.Maze(
 560         nb_train_samples=args.nb_train_samples,
 561         nb_test_samples=args.nb_test_samples,
 562         batch_size=args.physical_batch_size,
 563         height=args.maze_height,
 564         width=args.maze_width,
 565         nb_walls=args.maze_nb_walls,
 566         device="cpu",
 567     )
 568
 569 elif args.task == "snake":
 570     task = tasks.Snake(
 571         nb_train_samples=args.nb_train_samples,
 572         nb_test_samples=args.nb_test_samples,
 573         batch_size=args.physical_batch_size,
 574         height=args.snake_height,
 575         width=args.snake_width,
 576         nb_colors=args.snake_nb_colors,
 577         length=args.snake_length,
 578         prompt_length=args.snake_length // 2,
 579         device=device,
 580     )
 581
 582 elif args.task == "stack":
 583     task = tasks.Stack(
 584         nb_train_samples=args.nb_train_samples,
 585         nb_test_samples=args.nb_test_samples,
 586         batch_size=args.physical_batch_size,
 587         logger=log_string,
 588         nb_steps=args.stack_nb_steps,
 589         nb_stacks=args.stack_nb_stacks,
 590         nb_digits=args.stack_nb_digits,
 591         fraction_values_for_train=args.stack_fraction_values_for_train,
 592         device=device,
 593     )
 594
 595 elif args.task == "expr":
 596     task = tasks.Expr(
 597         nb_train_samples=args.nb_train_samples,
 598         nb_test_samples=args.nb_test_samples,
 599         nb_variables=args.expr_nb_variables,
 600         sequence_length=args.expr_sequence_length,
 601         operand_max=args.expr_operand_max,
 602         result_max=args.expr_result_max,
 603         batch_size=args.physical_batch_size,
 604         device=device,
 605     )
 606
 607 elif args.task == "rpl":
 608     task = tasks.RPL(
 609         nb_train_samples=args.nb_train_samples,
 610         nb_test_samples=args.nb_test_samples,
 611         batch_size=args.physical_batch_size,
 612         nb_starting_values=args.rpl_nb_starting_values,
 613         max_input=args.rpl_max_input,
 614         prog_len=args.rpl_prog_len,
 615         nb_runs=args.rpl_nb_runs,
 616         no_prog=args.rpl_no_prog,
 617         logger=log_string,
 618         device=device,
 619     )
 620
 621 elif args.task == "grid":
 622     task = tasks.Grid(
 623         nb_train_samples=args.nb_train_samples,
 624         nb_test_samples=args.nb_test_samples,
 625         batch_size=args.physical_batch_size,
 626         size=args.grid_size,
 627         fraction_play=args.grid_fraction_play,
 628         logger=log_string,
 629         device=device,
 630     )
 631
 632 elif args.task == "qmlp":
 633     task = tasks.QMLP(
 634         nb_train_samples=args.nb_train_samples,
 635         nb_test_samples=args.nb_test_samples,
 636         batch_size=args.physical_batch_size,
 637         result_dir=args.result_dir,
 638         logger=log_string,
 639         device=device,
 640     )
 641
 642 elif args.task == "greed":
 643     task = tasks.Greed(
 644         nb_train_samples=args.nb_train_samples,
 645         nb_test_samples=args.nb_test_samples,
 646         batch_size=args.physical_batch_size,
 647         height=args.greed_height,
 648         width=args.greed_width,
 649         T=args.greed_T,
 650         nb_walls=args.greed_nb_walls,
 651         nb_coins=args.greed_nb_coins,
 652         logger=log_string,
 653         device=device,
 654     )
 655
 656 else:
 657     raise ValueError(f"Unknown task {args.task}")
 658
 659 ######################################################################
 660
 661 log_string(f"device {device}")
 662
 663 vocabulary_size = task.vocabulary_size()
 664
 665 log_string(f"vocabulary_size {vocabulary_size}")
 666
 667 ######################################################################
 668
 669 # Compute the entropy of the training tokens
 670
 671 token_count = 0
 672 for input in task.batches(split="train", desc="train-entropy"):
 673     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 674 token_probas = token_count / token_count.sum()
 675 entropy = -torch.xlogy(token_probas, token_probas).sum()
 676 train_set_perplexity = math.exp(entropy)
 677
 678 ######################################################################
 679 # A bit of paranoia never hurts
 680
 681 if args.max_percents_of_test_in_train >= 0:
 682
 683     def subsets_as_tuples(batches, cs):
 684         s = set()
 685         for batch in batches:
 686             for x in batch:
 687                 s.add(tuple([v.item() for v in x]))
 688                 if len(s) == cs:
 689                     yield s
 690                     s = set()
 691         yield s
 692
 693     nb_test, nb_in_train = 0, 0
 694     for test_subset in subsets_as_tuples(
 695         task.batches(split="test", desc="test-check"), 25000
 696     ):
 697         in_train = set()
 698         for train_subset in subsets_as_tuples(
 699             task.batches(split="train", desc="train-check"), 25000
 700         ):
 701             in_train.update(test_subset.intersection(train_subset))
 702         nb_in_train += len(in_train)
 703         nb_test += len(test_subset)
 704
 705     log_string(
 706         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 707     )
 708
 709     assert (
 710         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 711     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 712
 713 ##############################
 714
 715
 716 def one_epoch(model, task):
 717     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 718
 719     model.train()
 720
 721     nb_train_samples, acc_train_loss = 0, 0.0
 722
 723     for input in task.batches(split="train"):
 724         input = input.to(device)
 725
 726         if nb_train_samples % args.batch_size == 0:
 727             optimizer.zero_grad()
 728
 729         output = model(mygpt.BracketedSequence(input)).x
 730         loss = F.cross_entropy(output.transpose(1, 2), input)
 731         acc_train_loss += loss.item() * input.size(0)
 732
 733         nb_train_samples += input.size(0)
 734
 735         loss.backward()
 736
 737         if nb_train_samples % args.batch_size == 0:
 738             optimizer.step()
 739
 740     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 741
 742     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 743
 744
 745 ######################################################################
 746
 747
 748 def run_tests(model, task, deterministic_synthesis):
 749     with torch.autograd.no_grad():
 750         model.eval()
 751
 752         nb_test_samples, acc_test_loss = 0, 0.0
 753         nb_samples_accumulated = 0
 754
 755         for input in task.batches(split="test"):
 756             input = input.to(device)
 757
 758             bs = model(mygpt.BracketedSequence(input))
 759             output = bs.x
 760
 761             loss = F.cross_entropy(output.transpose(1, 2), input)
 762
 763             acc_test_loss += loss.item() * input.size(0)
 764
 765             nb_test_samples += input.size(0)
 766
 767         main_test_accuracy = task.produce_results(
 768             n_epoch=n_epoch,
 769             model=model,
 770             result_dir=args.result_dir,
 771             logger=log_string,
 772             deterministic_synthesis=deterministic_synthesis,
 773         )
 774
 775         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 776
 777         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 778
 779     model.main_test_accuracy = main_test_accuracy
 780
 781
 782 ######################################################################
 783
 784
 785 def create_quizzes(
 786     model,
 787     other_models,
 788     task,
 789     nb_for_train=1000,
 790     nb_for_test=100,
 791 ):
 792     kept = []
 793
 794     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 795         new_quizzes, nb_correct = task.create_new_quizzes(
 796             n_epoch=n_epoch,
 797             result_dir=args.result_dir,
 798             logger=log_string,
 799             nb=4 * (nb_for_train + nb_for_test),
 800             model=model,
 801             other_models=other_models,
 802         )
 803
 804         to_keep = new_quizzes[nb_correct == len(other_models) - 1]
 805         log_string(f"keep {to_keep.size(0)} quizzes")
 806         kept.append(to_keep)
 807
 808     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 809
 810     task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
 811     task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
 812
 813     task.save_image(
 814         new_quizzes[:96],
 815         args.result_dir,
 816         f"world_new_{n_epoch:04d}_{model.id:02d}.png",
 817         log_string,
 818     )
 819
 820
 821 ######################################################################
 822
 823 models = []
 824
 825 for k in range(5):
 826     model = mygpt.MyGPT(
 827         vocabulary_size=vocabulary_size,
 828         dim_model=args.dim_model,
 829         dim_keys=args.dim_keys,
 830         dim_hidden=args.dim_hidden,
 831         nb_heads=args.nb_heads,
 832         nb_blocks=args.nb_blocks,
 833         causal=True,
 834         dropout=args.dropout,
 835     ).to(device)
 836
 837     model.main_test_accuracy = 0.0
 838     model.id = k
 839
 840     models.append(model)
 841
 842
 843 nb_parameters = sum(p.numel() for p in models[0].parameters())
 844 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 845
 846 ######################################################################
 847
 848 accuracy_to_make_quizzes = 0.975
 849
 850 for n_epoch in range(args.nb_epochs):
 851     models.sort(key=lambda model: model.main_test_accuracy)
 852
 853     model = models[0]
 854
 855     log_string(
 856         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 857     )
 858
 859     one_epoch(model, task)
 860
 861     log_string(
 862         f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
 863     )
 864
 865     run_tests(model, task, deterministic_synthesis=False)
 866
 867     if model.main_test_accuracy >= accuracy_to_make_quizzes:
 868         other_models = models.copy()
 869         other_models.remove(model)
 870
 871         create_quizzes(
 872             model,
 873             other_models,
 874             task,
 875             nb_for_train=1000,
 876             nb_for_test=100,
 877         )
 878
 879
 880 ######################################################################