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="twotargets",
  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=50)
  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("--optim", type=str, default="adam")
  60
  61 parser.add_argument("--learning_rate", type=float, default=1e-4)
  62
  63 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  64
  65 ########################################
  66
  67 parser.add_argument("--model", type=str, default=None)
  68
  69 parser.add_argument("--dim_model", type=int, default=None)
  70
  71 parser.add_argument("--dim_keys", type=int, default=None)
  72
  73 parser.add_argument("--dim_hidden", type=int, default=None)
  74
  75 parser.add_argument("--nb_heads", type=int, default=None)
  76
  77 parser.add_argument("--nb_blocks", type=int, default=None)
  78
  79 parser.add_argument("--dropout", type=float, default=0.1)
  80
  81 ########################################
  82
  83 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  84
  85 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  86
  87 parser.add_argument("--resume", action="store_true", default=False)
  88
  89 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  90
  91 ##############################
  92 # filetask
  93
  94 parser.add_argument("--filetask_train_file", type=str, default=None)
  95
  96 parser.add_argument("--filetask_test_file", type=str, default=None)
  97
  98 ##############################
  99 # rpl options
 100
 101 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 102
 103 parser.add_argument("--rpl_max_input", type=int, default=9)
 104
 105 parser.add_argument("--rpl_prog_len", type=int, default=8)
 106
 107 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 108
 109 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 110
 111 ##############################
 112 # grid options
 113
 114 parser.add_argument("--grid_size", type=int, default=6)
 115
 116 parser.add_argument("--grid_fraction_play", type=float, default=0)
 117
 118 ##############################
 119 # picoclvr options
 120
 121 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 122
 123 parser.add_argument("--picoclvr_height", type=int, default=12)
 124
 125 parser.add_argument("--picoclvr_width", type=int, default=16)
 126
 127 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 128
 129 ##############################
 130 # Maze options
 131
 132 parser.add_argument("--maze_height", type=int, default=13)
 133
 134 parser.add_argument("--maze_width", type=int, default=21)
 135
 136 parser.add_argument("--maze_nb_walls", type=int, default=15)
 137
 138 ##############################
 139 # Snake options
 140
 141 parser.add_argument("--snake_height", type=int, default=9)
 142
 143 parser.add_argument("--snake_width", type=int, default=12)
 144
 145 parser.add_argument("--snake_nb_colors", type=int, default=5)
 146
 147 parser.add_argument("--snake_length", type=int, default=200)
 148
 149 ##############################
 150 # ByHeart options
 151
 152 parser.add_argument("--byheart_separation", type=int, default=1)
 153
 154 ##############################
 155 # Stack options
 156
 157 parser.add_argument("--stack_nb_steps", type=int, default=100)
 158
 159 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 160
 161 parser.add_argument("--stack_nb_digits", type=int, default=3)
 162
 163 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 164
 165 ##############################
 166 # Expr options
 167
 168 parser.add_argument("--expr_nb_variables", type=int, default=5)
 169
 170 parser.add_argument("--expr_sequence_length", type=int, default=40)
 171
 172 parser.add_argument("--expr_operand_max", type=int, default=9)
 173
 174 parser.add_argument("--expr_result_max", type=int, default=99)
 175
 176 parser.add_argument("--expr_input_file", type=str, default=None)
 177
 178 ##############################
 179 # Mixing
 180
 181 parser.add_argument("--mixing_hard", action="store_true", default=False)
 182
 183 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 184
 185 ##############################
 186 # greed options
 187
 188 parser.add_argument("--greed_height", type=int, default=5)
 189
 190 parser.add_argument("--greed_width", type=int, default=7)
 191
 192 parser.add_argument("--greed_T", type=int, default=25)
 193
 194 parser.add_argument("--greed_nb_walls", type=int, default=5)
 195
 196 parser.add_argument("--greed_nb_coins", type=int, default=2)
 197
 198 ######################################################################
 199
 200 args = parser.parse_args()
 201
 202 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 203
 204 if args.result_dir is None:
 205     args.result_dir = f"results_{args.task}"
 206
 207 ######################################################################
 208
 209 default_task_args = {
 210     "file": {
 211         "model": "37M",
 212         "batch_size": 25,
 213         "nb_train_samples": 250000,
 214         "nb_test_samples": 10000,
 215     },
 216     "addition": {
 217         "model": "352M",
 218         "batch_size": 25,
 219         "nb_train_samples": 250000,
 220         "nb_test_samples": 10000,
 221     },
 222     "world": {
 223         "model": "37M",
 224         "batch_size": 25,
 225         "nb_train_samples": 50000,
 226         "nb_test_samples": 10000,
 227     },
 228     "byheart": {
 229         "model": "37M",
 230         "batch_size": 25,
 231         "nb_train_samples": 50000,
 232         "nb_test_samples": 10000,
 233     },
 234     "expr": {
 235         "model": "352M",
 236         "batch_size": 25,
 237         "nb_train_samples": 2500000,
 238         "nb_test_samples": 10000,
 239     },
 240     "grid": {
 241         "model": "37M",
 242         "batch_size": 25,
 243         "nb_train_samples": 250000,
 244         "nb_test_samples": 10000,
 245     },
 246     "qmlp": {
 247         "model": "37M",
 248         "batch_size": 10,
 249         "nb_train_samples": 100000,
 250         "nb_test_samples": 1000,
 251     },
 252     "guessop": {
 253         "model": "352M",
 254         "batch_size": 25,
 255         "nb_train_samples": 1000000,
 256         "nb_test_samples": 10000,
 257     },
 258     "learnop": {
 259         "model": "37M",
 260         "batch_size": 25,
 261         "nb_train_samples": 50000,
 262         "nb_test_samples": 10000,
 263     },
 264     "maze": {
 265         "model": "37M",
 266         "batch_size": 5,
 267         "nb_train_samples": 100000,
 268         "nb_test_samples": 10000,
 269     },
 270     "picoclvr": {
 271         "model": "37M",
 272         "batch_size": 25,
 273         "nb_train_samples": 250000,
 274         "nb_test_samples": 10000,
 275     },
 276     "rpl": {
 277         "model": "352M",
 278         "batch_size": 5,
 279         "nb_train_samples": 2500000,
 280         "nb_test_samples": 10000,
 281     },
 282     "snake": {
 283         "model": "37M",
 284         "batch_size": 25,
 285         "nb_train_samples": 250000,
 286         "nb_test_samples": 10000,
 287     },
 288     "stack": {
 289         "model": "37M",
 290         "batch_size": 25,
 291         "nb_train_samples": 100000,
 292         "nb_test_samples": 1000,
 293     },
 294     "twotargets": {
 295         "model": "37M",
 296         "batch_size": 25,
 297         "nb_train_samples": 50000,
 298         "nb_test_samples": 10000,
 299     },
 300     "memory": {
 301         "model": "37M",
 302         "batch_size": 100,
 303         "nb_train_samples": 25000,
 304         "nb_test_samples": 1000,
 305     },
 306     "mixing": {
 307         "model": "37M",
 308         "batch_size": 25,
 309         "nb_train_samples": 250000,
 310         "nb_test_samples": 10000,
 311     },
 312     "mnist": {
 313         "model": "37M",
 314         "batch_size": 10,
 315         "nb_train_samples": 60000,
 316         "nb_test_samples": 10000,
 317     },
 318     "greed": {
 319         "model": "37M",
 320         "batch_size": 25,
 321         "nb_train_samples": 25000,
 322         "nb_test_samples": 10000,
 323     },
 324 }
 325
 326 if args.task in default_task_args:
 327     for k, v in default_task_args[args.task].items():
 328         if getattr(args, k) is None:
 329             setattr(args, k, v)
 330
 331 ######################################################################
 332
 333 default_model_args = {
 334     "17K": {
 335         "dim_model": 32,
 336         "dim_keys": 32,
 337         "dim_hidden": 32,
 338         "nb_heads": 2,
 339         "nb_blocks": 2,
 340     },
 341     "4M": {
 342         "dim_model": 256,
 343         "dim_keys": 32,
 344         "dim_hidden": 1024,
 345         "nb_heads": 4,
 346         "nb_blocks": 6,
 347     },
 348     "37M": {
 349         "dim_model": 512,
 350         "dim_keys": 64,
 351         "dim_hidden": 2048,
 352         "nb_heads": 8,
 353         "nb_blocks": 12,
 354     },
 355     "122M": {
 356         "dim_model": 768,
 357         "dim_keys": 64,
 358         "dim_hidden": 2048,
 359         "nb_heads": 8,
 360         "nb_blocks": 24,
 361     },
 362     "352M": {
 363         "dim_model": 1024,
 364         "dim_keys": 64,
 365         "dim_hidden": 2048,
 366         "nb_heads": 8,
 367         "nb_blocks": 48,
 368     },
 369 }
 370
 371 if args.model in default_model_args:
 372     for k, v in default_model_args[args.model].items():
 373         if getattr(args, k) is None:
 374             setattr(args, k, v)
 375 else:
 376     raise ValueError(f"Unknown model {args.model}")
 377
 378 ######################################################################
 379
 380 try:
 381     os.mkdir(args.result_dir)
 382 except FileExistsError:
 383     if not args.resume:
 384         print(f"result directory {args.result_dir} already exists")
 385         exit(1)
 386
 387 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 388
 389 if args.seed >= 0:
 390     # torch.backends.cudnn.deterministic = True
 391     # torch.backends.cudnn.benchmark = False
 392     # torch.use_deterministic_algorithms(True)
 393     torch.manual_seed(args.seed)
 394     if torch.cuda.is_available():
 395         torch.cuda.manual_seed_all(args.seed)
 396
 397 ######################################################################
 398
 399
 400 def log_string(s):
 401     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 402
 403     if log_file is not None:
 404         log_file.write(t + s + "\n")
 405         log_file.flush()
 406
 407     print(t + s)
 408     sys.stdout.flush()
 409
 410
 411 log_string(f"argv {' '.join(sys.argv)}")
 412
 413 for n in vars(args):
 414     log_string(f"args.{n} {getattr(args, n)}")
 415
 416
 417 ######################################################################
 418
 419
 420 def picoclvr_pruner_horizontal_green(p):
 421     return not ("green" in p and ("left" in p or "right" in p))
 422
 423
 424 picoclvr_pruner_train = (
 425     picoclvr_pruner_horizontal_green
 426     if args.picocvlr_prune_properties in {"train+eval"}
 427     else None
 428 )
 429
 430 picoclvr_pruner_eval = (
 431     (lambda p: not picoclvr_pruner_horizontal_green(p))
 432     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 433     else None
 434 )
 435
 436 ######################################################################
 437
 438 if args.physical_batch_size is None:
 439     args.physical_batch_size = args.batch_size
 440 else:
 441     assert args.batch_size % args.physical_batch_size == 0
 442
 443 assert args.nb_train_samples % args.batch_size == 0
 444 assert args.nb_test_samples % args.batch_size == 0
 445
 446 if args.task == "file":
 447     assert (
 448         args.filetask_train_file is not None and args.filetask_test_file is not None
 449     ), "You have to specify the task train and test files"
 450     task = tasks.TaskFromFile(
 451         args.filetask_train_file,
 452         args.filetask_test_file,
 453         nb_train_samples=args.nb_train_samples,
 454         nb_test_samples=args.nb_test_samples,
 455         batch_size=args.physical_batch_size,
 456         shuffle=True,
 457         device=device,
 458     )
 459     args.max_percents_of_test_in_train = 0
 460
 461 elif args.task == "byheart":
 462     task = tasks.SandBox(
 463         problem=problems.ProblemByHeart(separation=args.byheart_separation),
 464         nb_train_samples=args.nb_train_samples,
 465         nb_test_samples=args.nb_test_samples,
 466         batch_size=args.physical_batch_size,
 467         logger=log_string,
 468         device=device,
 469     )
 470     args.max_percents_of_test_in_train = -1
 471
 472 elif args.task == "world":
 473     task = tasks.World(
 474         nb_train_samples=args.nb_train_samples,
 475         nb_test_samples=args.nb_test_samples,
 476         batch_size=args.physical_batch_size,
 477         logger=log_string,
 478         device=device,
 479     )
 480     args.max_percents_of_test_in_train = -1
 481
 482 elif args.task == "learnop":
 483     task = tasks.SandBox(
 484         problem=problems.ProblemLearnOperator(),
 485         nb_train_samples=args.nb_train_samples,
 486         nb_test_samples=args.nb_test_samples,
 487         batch_size=args.physical_batch_size,
 488         logger=log_string,
 489         device=device,
 490     )
 491
 492
 493 elif args.task == "guessop":
 494     task = tasks.SandBox(
 495         problem=problems.ProblemGuessOperator(),
 496         nb_train_samples=args.nb_train_samples,
 497         nb_test_samples=args.nb_test_samples,
 498         batch_size=args.physical_batch_size,
 499         logger=log_string,
 500         device=device,
 501     )
 502
 503
 504 elif args.task == "twotargets":
 505     task = tasks.SandBox(
 506         problem=problems.ProblemTwoTargets(),
 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 == "memory":
 515     task = tasks.SandBox(
 516         problem=problems.ProblemMemory(),
 517         nb_train_samples=args.nb_train_samples,
 518         nb_test_samples=args.nb_test_samples,
 519         batch_size=args.physical_batch_size,
 520         logger=log_string,
 521         device=device,
 522     )
 523
 524 elif args.task == "mixing":
 525     task = tasks.SandBox(
 526         problem=problems.ProblemMixing(
 527             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 528         ),
 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 == "addition":
 537     task = tasks.SandBox(
 538         problem=problems.ProblemAddition(),
 539         nb_train_samples=args.nb_train_samples,
 540         nb_test_samples=args.nb_test_samples,
 541         batch_size=args.physical_batch_size,
 542         logger=log_string,
 543         device=device,
 544     )
 545
 546 elif args.task == "picoclvr":
 547     task = tasks.PicoCLVR(
 548         nb_train_samples=args.nb_train_samples,
 549         nb_test_samples=args.nb_test_samples,
 550         batch_size=args.physical_batch_size,
 551         height=args.picoclvr_height,
 552         width=args.picoclvr_width,
 553         nb_colors=args.picoclvr_nb_colors,
 554         logger=log_string,
 555         device=device,
 556         pruner_train=picoclvr_pruner_train,
 557         pruner_eval=picoclvr_pruner_eval,
 558     )
 559
 560 elif args.task == "mnist":
 561     task = tasks.MNIST(
 562         nb_train_samples=args.nb_train_samples,
 563         nb_test_samples=args.nb_test_samples,
 564         batch_size=args.physical_batch_size,
 565         device=device,
 566     )
 567
 568 elif args.task == "maze":
 569     task = tasks.Maze(
 570         nb_train_samples=args.nb_train_samples,
 571         nb_test_samples=args.nb_test_samples,
 572         batch_size=args.physical_batch_size,
 573         height=args.maze_height,
 574         width=args.maze_width,
 575         nb_walls=args.maze_nb_walls,
 576         device="cpu",
 577     )
 578
 579 elif args.task == "snake":
 580     task = tasks.Snake(
 581         nb_train_samples=args.nb_train_samples,
 582         nb_test_samples=args.nb_test_samples,
 583         batch_size=args.physical_batch_size,
 584         height=args.snake_height,
 585         width=args.snake_width,
 586         nb_colors=args.snake_nb_colors,
 587         length=args.snake_length,
 588         prompt_length=args.snake_length // 2,
 589         device=device,
 590     )
 591
 592 elif args.task == "stack":
 593     task = tasks.Stack(
 594         nb_train_samples=args.nb_train_samples,
 595         nb_test_samples=args.nb_test_samples,
 596         batch_size=args.physical_batch_size,
 597         logger=log_string,
 598         nb_steps=args.stack_nb_steps,
 599         nb_stacks=args.stack_nb_stacks,
 600         nb_digits=args.stack_nb_digits,
 601         fraction_values_for_train=args.stack_fraction_values_for_train,
 602         device=device,
 603     )
 604
 605 elif args.task == "expr":
 606     task = tasks.Expr(
 607         nb_train_samples=args.nb_train_samples,
 608         nb_test_samples=args.nb_test_samples,
 609         nb_variables=args.expr_nb_variables,
 610         sequence_length=args.expr_sequence_length,
 611         operand_max=args.expr_operand_max,
 612         result_max=args.expr_result_max,
 613         batch_size=args.physical_batch_size,
 614         device=device,
 615     )
 616
 617 elif args.task == "rpl":
 618     task = tasks.RPL(
 619         nb_train_samples=args.nb_train_samples,
 620         nb_test_samples=args.nb_test_samples,
 621         batch_size=args.physical_batch_size,
 622         nb_starting_values=args.rpl_nb_starting_values,
 623         max_input=args.rpl_max_input,
 624         prog_len=args.rpl_prog_len,
 625         nb_runs=args.rpl_nb_runs,
 626         no_prog=args.rpl_no_prog,
 627         logger=log_string,
 628         device=device,
 629     )
 630
 631 elif args.task == "grid":
 632     task = tasks.Grid(
 633         nb_train_samples=args.nb_train_samples,
 634         nb_test_samples=args.nb_test_samples,
 635         batch_size=args.physical_batch_size,
 636         size=args.grid_size,
 637         fraction_play=args.grid_fraction_play,
 638         logger=log_string,
 639         device=device,
 640     )
 641
 642 elif args.task == "qmlp":
 643     task = tasks.QMLP(
 644         nb_train_samples=args.nb_train_samples,
 645         nb_test_samples=args.nb_test_samples,
 646         batch_size=args.physical_batch_size,
 647         result_dir=args.result_dir,
 648         logger=log_string,
 649         device=device,
 650     )
 651
 652 elif args.task == "greed":
 653     task = tasks.Greed(
 654         nb_train_samples=args.nb_train_samples,
 655         nb_test_samples=args.nb_test_samples,
 656         batch_size=args.physical_batch_size,
 657         height=args.greed_height,
 658         width=args.greed_width,
 659         T=args.greed_T,
 660         nb_walls=args.greed_nb_walls,
 661         nb_coins=args.greed_nb_coins,
 662         logger=log_string,
 663         device=device,
 664     )
 665
 666 else:
 667     raise ValueError(f"Unknown task {args.task}")
 668
 669 ######################################################################
 670
 671 log_string(f"device {device}")
 672
 673 vocabulary_size = task.vocabulary_size()
 674
 675 log_string(f"vocabulary_size {vocabulary_size}")
 676
 677 ##############################
 678
 679 model = mygpt.MyGPT(
 680     vocabulary_size=vocabulary_size,
 681     dim_model=args.dim_model,
 682     dim_keys=args.dim_keys,
 683     dim_hidden=args.dim_hidden,
 684     nb_heads=args.nb_heads,
 685     nb_blocks=args.nb_blocks,
 686     causal=True,
 687     dropout=args.dropout,
 688 )
 689
 690 model.to(device)
 691
 692 nb_parameters = sum(p.numel() for p in model.parameters())
 693 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 694
 695 ######################################################################
 696
 697 nb_epochs_finished = 0
 698
 699 if args.no_checkpoint:
 700     log_string(f"not trying to load checkpoint.")
 701
 702 else:
 703     try:
 704         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 705         checkpoint = torch.load(checkpoint_name)
 706         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 707         model.load_state_dict(checkpoint["model_state"])
 708         torch.set_rng_state(checkpoint["rng_state"])
 709         if torch.cuda.is_available():
 710             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 711
 712         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 713
 714     except FileNotFoundError:
 715         log_string("starting from scratch.")
 716
 717     except:
 718         log_string("error when loading the checkpoint.")
 719         exit(1)
 720
 721 ######################################################################
 722
 723 if args.task == "expr" and args.expr_input_file is not None:
 724     task.produce_results(
 725         n_epoch=nb_epochs_finished,
 726         model=model,
 727         result_dir=args.result_dir,
 728         logger=log_string,
 729         deterministic_synthesis=args.deterministic_synthesis,
 730         input_file=args.expr_input_file,
 731     )
 732
 733     exit(0)
 734
 735 ######################################################################
 736
 737 # Compute the entropy of the training tokens
 738
 739 token_count = 0
 740 for input in task.batches(split="train", desc="train-entropy"):
 741     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 742 token_probas = token_count / token_count.sum()
 743 entropy = -torch.xlogy(token_probas, token_probas).sum()
 744 train_set_perplexity = math.exp(entropy)
 745
 746 ######################################################################
 747 # A bit of paranoia never hurts
 748
 749 if args.max_percents_of_test_in_train >= 0:
 750
 751     def subsets_as_tuples(batches, cs):
 752         s = set()
 753         for batch in batches:
 754             for x in batch:
 755                 s.add(tuple([v.item() for v in x]))
 756                 if len(s) == cs:
 757                     yield s
 758                     s = set()
 759         yield s
 760
 761     nb_test, nb_in_train = 0, 0
 762     for test_subset in subsets_as_tuples(
 763         task.batches(split="test", desc="test-check"), 25000
 764     ):
 765         in_train = set()
 766         for train_subset in subsets_as_tuples(
 767             task.batches(split="train", desc="train-check"), 25000
 768         ):
 769             in_train.update(test_subset.intersection(train_subset))
 770         nb_in_train += len(in_train)
 771         nb_test += len(test_subset)
 772
 773     log_string(
 774         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 775     )
 776
 777     assert (
 778         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 779     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 780
 781 ##############################
 782
 783 if args.learning_rate_schedule == "cos":
 784     learning_rate_schedule = {}
 785     for n_epoch in range(args.nb_epochs):
 786         u = n_epoch / args.nb_epochs * math.pi
 787         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 788 else:
 789     u = {
 790         int(k): float(v)
 791         for k, v in [
 792             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 793         ]
 794     }
 795
 796     learning_rate_schedule = {}
 797     learning_rate = args.learning_rate
 798     for n_epoch in range(args.nb_epochs):
 799         if n_epoch in u:
 800             learning_rate = u[n_epoch]
 801         learning_rate_schedule[n_epoch] = learning_rate
 802
 803 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 804
 805 ##############################
 806
 807 if nb_epochs_finished >= args.nb_epochs:
 808     task.produce_results(
 809         n_epoch=nb_epochs_finished,
 810         model=model,
 811         result_dir=args.result_dir,
 812         logger=log_string,
 813         deterministic_synthesis=args.deterministic_synthesis,
 814     )
 815
 816 time_pred_result = None
 817
 818 ######################################################################
 819
 820
 821 def one_epoch(model, task, learning_rate):
 822     log_string(f"learning_rate {learning_rate}")
 823
 824     if args.optim == "sgd":
 825         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 826     elif args.optim == "adam":
 827         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 828     elif args.optim == "adamw":
 829         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 830     else:
 831         raise ValueError(f"Unknown optimizer {args.optim}.")
 832
 833     model.train()
 834
 835     nb_train_samples, acc_train_loss = 0, 0.0
 836
 837     for input in task.batches(split="train"):
 838         input = input.to(device)
 839
 840         if nb_train_samples % args.batch_size == 0:
 841             optimizer.zero_grad()
 842
 843         output = model(mygpt.BracketedSequence(input)).x
 844         loss = F.cross_entropy(output.transpose(1, 2), input)
 845         acc_train_loss += loss.item() * input.size(0)
 846
 847         nb_train_samples += input.size(0)
 848
 849         loss.backward()
 850
 851         if nb_train_samples % args.batch_size == 0:
 852             optimizer.step()
 853
 854     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 855
 856     log_string(f"train)perplexity {n_epoch} {train_perplexity}")
 857
 858
 859 ######################################################################
 860
 861
 862 def run_tests(model, task):
 863     with torch.autograd.no_grad():
 864         model.eval()
 865
 866         nb_test_samples, acc_test_loss = 0, 0.0
 867         nb_samples_accumulated = 0
 868
 869         for input in task.batches(split="test"):
 870             input = input.to(device)
 871
 872             bs = model(mygpt.BracketedSequence(input))
 873             output = bs.x
 874
 875             loss = F.cross_entropy(output.transpose(1, 2), input)
 876
 877             acc_test_loss += loss.item() * input.size(0)
 878
 879             nb_test_samples += input.size(0)
 880
 881         task.produce_results(
 882             n_epoch=n_epoch,
 883             model=model,
 884             result_dir=args.result_dir,
 885             logger=log_string,
 886             deterministic_synthesis=args.deterministic_synthesis,
 887         )
 888
 889         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 890         log_string(f"test)perplexity {n_epoch} {test_perplexity}")
 891
 892
 893 ######################################################################
 894
 895 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 896     learning_rate = learning_rate_schedule[n_epoch]
 897
 898     one_epoch(model, task, learning_rate)
 899
 900     run_tests(model, task)
 901
 902     time_current_result = datetime.datetime.now()
 903     if time_pred_result is not None:
 904         log_string(
 905             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 906         )
 907     time_pred_result = time_current_result
 908
 909     checkpoint = {
 910         "nb_epochs_finished": n_epoch + 1,
 911         "model_state": model.state_dict(),
 912         "rng_state": torch.get_rng_state(),
 913     }
 914
 915     if torch.cuda.is_available():
 916         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 917
 918     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 919     torch.save(checkpoint, checkpoint_name)
 920     log_string(f"saved checkpoint {checkpoint_name}")
 921
 922 ######################################################################