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 ##############################
  86 # filetask
  87
  88 parser.add_argument("--filetask_train_file", type=str, default=None)
  89
  90 parser.add_argument("--filetask_test_file", type=str, default=None)
  91
  92 ##############################
  93 # rpl options
  94
  95 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
  96
  97 parser.add_argument("--rpl_max_input", type=int, default=9)
  98
  99 parser.add_argument("--rpl_prog_len", type=int, default=8)
 100
 101 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 102
 103 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 104
 105 ##############################
 106 # grid options
 107
 108 parser.add_argument("--grid_size", type=int, default=6)
 109
 110 parser.add_argument("--grid_fraction_play", type=float, default=0)
 111
 112 ##############################
 113 # picoclvr options
 114
 115 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 116
 117 parser.add_argument("--picoclvr_height", type=int, default=12)
 118
 119 parser.add_argument("--picoclvr_width", type=int, default=16)
 120
 121 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 122
 123 ##############################
 124 # Maze options
 125
 126 parser.add_argument("--maze_height", type=int, default=13)
 127
 128 parser.add_argument("--maze_width", type=int, default=21)
 129
 130 parser.add_argument("--maze_nb_walls", type=int, default=15)
 131
 132 ##############################
 133 # Snake options
 134
 135 parser.add_argument("--snake_height", type=int, default=9)
 136
 137 parser.add_argument("--snake_width", type=int, default=12)
 138
 139 parser.add_argument("--snake_nb_colors", type=int, default=5)
 140
 141 parser.add_argument("--snake_length", type=int, default=200)
 142
 143 ##############################
 144 # ByHeart options
 145
 146 parser.add_argument("--byheart_separation", type=int, default=1)
 147
 148 ##############################
 149 # Stack options
 150
 151 parser.add_argument("--stack_nb_steps", type=int, default=100)
 152
 153 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 154
 155 parser.add_argument("--stack_nb_digits", type=int, default=3)
 156
 157 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 158
 159 ##############################
 160 # Expr options
 161
 162 parser.add_argument("--expr_nb_variables", type=int, default=5)
 163
 164 parser.add_argument("--expr_sequence_length", type=int, default=40)
 165
 166 parser.add_argument("--expr_operand_max", type=int, default=9)
 167
 168 parser.add_argument("--expr_result_max", type=int, default=99)
 169
 170 parser.add_argument("--expr_input_file", type=str, default=None)
 171
 172 ##############################
 173 # Mixing
 174
 175 parser.add_argument("--mixing_hard", action="store_true", default=False)
 176
 177 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 178
 179 ##############################
 180 # greed options
 181
 182 parser.add_argument("--greed_height", type=int, default=5)
 183
 184 parser.add_argument("--greed_width", type=int, default=7)
 185
 186 parser.add_argument("--greed_T", type=int, default=25)
 187
 188 parser.add_argument("--greed_nb_walls", type=int, default=5)
 189
 190 parser.add_argument("--greed_nb_coins", type=int, default=2)
 191
 192 ######################################################################
 193
 194 args = parser.parse_args()
 195
 196 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 197
 198 if args.result_dir is None:
 199     args.result_dir = f"results_{args.task}"
 200
 201 ######################################################################
 202
 203 default_task_args = {
 204     "world": {
 205         "model": "37M",
 206         "batch_size": 100,
 207         "nb_train_samples": 250000,
 208         "nb_test_samples": 10000,
 209     },
 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     "byheart": {
 223         "model": "37M",
 224         "batch_size": 25,
 225         "nb_train_samples": 50000,
 226         "nb_test_samples": 10000,
 227     },
 228     "expr": {
 229         "model": "352M",
 230         "batch_size": 25,
 231         "nb_train_samples": 2500000,
 232         "nb_test_samples": 10000,
 233     },
 234     "grid": {
 235         "model": "37M",
 236         "batch_size": 25,
 237         "nb_train_samples": 250000,
 238         "nb_test_samples": 10000,
 239     },
 240     "qmlp": {
 241         "model": "37M",
 242         "batch_size": 10,
 243         "nb_train_samples": 100000,
 244         "nb_test_samples": 1000,
 245     },
 246     "guessop": {
 247         "model": "352M",
 248         "batch_size": 25,
 249         "nb_train_samples": 1000000,
 250         "nb_test_samples": 10000,
 251     },
 252     "learnop": {
 253         "model": "37M",
 254         "batch_size": 25,
 255         "nb_train_samples": 50000,
 256         "nb_test_samples": 10000,
 257     },
 258     "maze": {
 259         "model": "37M",
 260         "batch_size": 5,
 261         "nb_train_samples": 100000,
 262         "nb_test_samples": 10000,
 263     },
 264     "picoclvr": {
 265         "model": "37M",
 266         "batch_size": 25,
 267         "nb_train_samples": 250000,
 268         "nb_test_samples": 10000,
 269     },
 270     "rpl": {
 271         "model": "352M",
 272         "batch_size": 5,
 273         "nb_train_samples": 2500000,
 274         "nb_test_samples": 10000,
 275     },
 276     "snake": {
 277         "model": "37M",
 278         "batch_size": 25,
 279         "nb_train_samples": 250000,
 280         "nb_test_samples": 10000,
 281     },
 282     "stack": {
 283         "model": "37M",
 284         "batch_size": 25,
 285         "nb_train_samples": 100000,
 286         "nb_test_samples": 1000,
 287     },
 288     "twotargets": {
 289         "model": "37M",
 290         "batch_size": 25,
 291         "nb_train_samples": 50000,
 292         "nb_test_samples": 10000,
 293     },
 294     "memory": {
 295         "model": "37M",
 296         "batch_size": 100,
 297         "nb_train_samples": 25000,
 298         "nb_test_samples": 1000,
 299     },
 300     "mixing": {
 301         "model": "37M",
 302         "batch_size": 25,
 303         "nb_train_samples": 250000,
 304         "nb_test_samples": 10000,
 305     },
 306     "mnist": {
 307         "model": "37M",
 308         "batch_size": 10,
 309         "nb_train_samples": 60000,
 310         "nb_test_samples": 10000,
 311     },
 312     "greed": {
 313         "model": "37M",
 314         "batch_size": 25,
 315         "nb_train_samples": 25000,
 316         "nb_test_samples": 10000,
 317     },
 318 }
 319
 320 if args.task in default_task_args:
 321     for k, v in default_task_args[args.task].items():
 322         if getattr(args, k) is None:
 323             setattr(args, k, v)
 324
 325 ######################################################################
 326
 327 default_model_args = {
 328     "17K": {
 329         "dim_model": 32,
 330         "dim_keys": 32,
 331         "dim_hidden": 32,
 332         "nb_heads": 2,
 333         "nb_blocks": 2,
 334     },
 335     "4M": {
 336         "dim_model": 256,
 337         "dim_keys": 32,
 338         "dim_hidden": 1024,
 339         "nb_heads": 4,
 340         "nb_blocks": 6,
 341     },
 342     "37M": {
 343         "dim_model": 512,
 344         "dim_keys": 64,
 345         "dim_hidden": 2048,
 346         "nb_heads": 8,
 347         "nb_blocks": 12,
 348     },
 349     "122M": {
 350         "dim_model": 768,
 351         "dim_keys": 64,
 352         "dim_hidden": 2048,
 353         "nb_heads": 8,
 354         "nb_blocks": 24,
 355     },
 356     "352M": {
 357         "dim_model": 1024,
 358         "dim_keys": 64,
 359         "dim_hidden": 2048,
 360         "nb_heads": 8,
 361         "nb_blocks": 48,
 362     },
 363 }
 364
 365 if args.model in default_model_args:
 366     for k, v in default_model_args[args.model].items():
 367         if getattr(args, k) is None:
 368             setattr(args, k, v)
 369 else:
 370     raise ValueError(f"Unknown model {args.model}")
 371
 372 ######################################################################
 373
 374 try:
 375     os.mkdir(args.result_dir)
 376 except FileExistsError:
 377     if not args.resume:
 378         print(f"result directory {args.result_dir} already exists")
 379         exit(1)
 380
 381 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 382
 383 if args.seed >= 0:
 384     # torch.backends.cudnn.deterministic = True
 385     # torch.backends.cudnn.benchmark = False
 386     # torch.use_deterministic_algorithms(True)
 387     torch.manual_seed(args.seed)
 388     if torch.cuda.is_available():
 389         torch.cuda.manual_seed_all(args.seed)
 390
 391 ######################################################################
 392
 393
 394 def log_string(s):
 395     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 396
 397     if log_file is not None:
 398         log_file.write(t + s + "\n")
 399         log_file.flush()
 400
 401     print(t + s)
 402     sys.stdout.flush()
 403
 404
 405 log_string(f"argv {' '.join(sys.argv)}")
 406
 407 for n in vars(args):
 408     log_string(f"args.{n} {getattr(args, n)}")
 409
 410
 411 ######################################################################
 412
 413
 414 def picoclvr_pruner_horizontal_green(p):
 415     return not ("green" in p and ("left" in p or "right" in p))
 416
 417
 418 picoclvr_pruner_train = (
 419     picoclvr_pruner_horizontal_green
 420     if args.picocvlr_prune_properties in {"train+eval"}
 421     else None
 422 )
 423
 424 picoclvr_pruner_eval = (
 425     (lambda p: not picoclvr_pruner_horizontal_green(p))
 426     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 427     else None
 428 )
 429
 430 ######################################################################
 431
 432 if args.physical_batch_size is None:
 433     args.physical_batch_size = args.batch_size
 434 else:
 435     assert args.batch_size % args.physical_batch_size == 0
 436
 437 assert args.nb_train_samples % args.batch_size == 0
 438 assert args.nb_test_samples % args.batch_size == 0
 439
 440 if args.task == "file":
 441     assert (
 442         args.filetask_train_file is not None and args.filetask_test_file is not None
 443     ), "You have to specify the task train and test files"
 444     task = tasks.TaskFromFile(
 445         args.filetask_train_file,
 446         args.filetask_test_file,
 447         nb_train_samples=args.nb_train_samples,
 448         nb_test_samples=args.nb_test_samples,
 449         batch_size=args.physical_batch_size,
 450         shuffle=True,
 451         device=device,
 452     )
 453     args.max_percents_of_test_in_train = 0
 454
 455 elif args.task == "byheart":
 456     task = tasks.SandBox(
 457         problem=problems.ProblemByHeart(separation=args.byheart_separation),
 458         nb_train_samples=args.nb_train_samples,
 459         nb_test_samples=args.nb_test_samples,
 460         batch_size=args.physical_batch_size,
 461         logger=log_string,
 462         device=device,
 463     )
 464     args.max_percents_of_test_in_train = -1
 465
 466 elif args.task == "world":
 467     task = tasks.World(
 468         nb_train_samples=args.nb_train_samples,
 469         nb_test_samples=args.nb_test_samples,
 470         batch_size=args.physical_batch_size,
 471         result_dir=args.result_dir,
 472         logger=log_string,
 473         device=device,
 474     )
 475     args.max_percents_of_test_in_train = -1
 476
 477 elif args.task == "learnop":
 478     task = tasks.SandBox(
 479         problem=problems.ProblemLearnOperator(),
 480         nb_train_samples=args.nb_train_samples,
 481         nb_test_samples=args.nb_test_samples,
 482         batch_size=args.physical_batch_size,
 483         logger=log_string,
 484         device=device,
 485     )
 486
 487
 488 elif args.task == "guessop":
 489     task = tasks.SandBox(
 490         problem=problems.ProblemGuessOperator(),
 491         nb_train_samples=args.nb_train_samples,
 492         nb_test_samples=args.nb_test_samples,
 493         batch_size=args.physical_batch_size,
 494         logger=log_string,
 495         device=device,
 496     )
 497
 498
 499 elif args.task == "twotargets":
 500     task = tasks.SandBox(
 501         problem=problems.ProblemTwoTargets(),
 502         nb_train_samples=args.nb_train_samples,
 503         nb_test_samples=args.nb_test_samples,
 504         batch_size=args.physical_batch_size,
 505         logger=log_string,
 506         device=device,
 507     )
 508
 509 elif args.task == "memory":
 510     task = tasks.SandBox(
 511         problem=problems.ProblemMemory(),
 512         nb_train_samples=args.nb_train_samples,
 513         nb_test_samples=args.nb_test_samples,
 514         batch_size=args.physical_batch_size,
 515         logger=log_string,
 516         device=device,
 517     )
 518
 519 elif args.task == "mixing":
 520     task = tasks.SandBox(
 521         problem=problems.ProblemMixing(
 522             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 523         ),
 524         nb_train_samples=args.nb_train_samples,
 525         nb_test_samples=args.nb_test_samples,
 526         batch_size=args.physical_batch_size,
 527         logger=log_string,
 528         device=device,
 529     )
 530
 531 elif args.task == "addition":
 532     task = tasks.SandBox(
 533         problem=problems.ProblemAddition(),
 534         nb_train_samples=args.nb_train_samples,
 535         nb_test_samples=args.nb_test_samples,
 536         batch_size=args.physical_batch_size,
 537         logger=log_string,
 538         device=device,
 539     )
 540
 541 elif args.task == "picoclvr":
 542     task = tasks.PicoCLVR(
 543         nb_train_samples=args.nb_train_samples,
 544         nb_test_samples=args.nb_test_samples,
 545         batch_size=args.physical_batch_size,
 546         height=args.picoclvr_height,
 547         width=args.picoclvr_width,
 548         nb_colors=args.picoclvr_nb_colors,
 549         logger=log_string,
 550         device=device,
 551         pruner_train=picoclvr_pruner_train,
 552         pruner_eval=picoclvr_pruner_eval,
 553     )
 554
 555 elif args.task == "mnist":
 556     task = tasks.MNIST(
 557         nb_train_samples=args.nb_train_samples,
 558         nb_test_samples=args.nb_test_samples,
 559         batch_size=args.physical_batch_size,
 560         device=device,
 561     )
 562
 563 elif args.task == "maze":
 564     task = tasks.Maze(
 565         nb_train_samples=args.nb_train_samples,
 566         nb_test_samples=args.nb_test_samples,
 567         batch_size=args.physical_batch_size,
 568         height=args.maze_height,
 569         width=args.maze_width,
 570         nb_walls=args.maze_nb_walls,
 571         device="cpu",
 572     )
 573
 574 elif args.task == "snake":
 575     task = tasks.Snake(
 576         nb_train_samples=args.nb_train_samples,
 577         nb_test_samples=args.nb_test_samples,
 578         batch_size=args.physical_batch_size,
 579         height=args.snake_height,
 580         width=args.snake_width,
 581         nb_colors=args.snake_nb_colors,
 582         length=args.snake_length,
 583         prompt_length=args.snake_length // 2,
 584         device=device,
 585     )
 586
 587 elif args.task == "stack":
 588     task = tasks.Stack(
 589         nb_train_samples=args.nb_train_samples,
 590         nb_test_samples=args.nb_test_samples,
 591         batch_size=args.physical_batch_size,
 592         logger=log_string,
 593         nb_steps=args.stack_nb_steps,
 594         nb_stacks=args.stack_nb_stacks,
 595         nb_digits=args.stack_nb_digits,
 596         fraction_values_for_train=args.stack_fraction_values_for_train,
 597         device=device,
 598     )
 599
 600 elif args.task == "expr":
 601     task = tasks.Expr(
 602         nb_train_samples=args.nb_train_samples,
 603         nb_test_samples=args.nb_test_samples,
 604         nb_variables=args.expr_nb_variables,
 605         sequence_length=args.expr_sequence_length,
 606         operand_max=args.expr_operand_max,
 607         result_max=args.expr_result_max,
 608         batch_size=args.physical_batch_size,
 609         device=device,
 610     )
 611
 612 elif args.task == "rpl":
 613     task = tasks.RPL(
 614         nb_train_samples=args.nb_train_samples,
 615         nb_test_samples=args.nb_test_samples,
 616         batch_size=args.physical_batch_size,
 617         nb_starting_values=args.rpl_nb_starting_values,
 618         max_input=args.rpl_max_input,
 619         prog_len=args.rpl_prog_len,
 620         nb_runs=args.rpl_nb_runs,
 621         no_prog=args.rpl_no_prog,
 622         logger=log_string,
 623         device=device,
 624     )
 625
 626 elif args.task == "grid":
 627     task = tasks.Grid(
 628         nb_train_samples=args.nb_train_samples,
 629         nb_test_samples=args.nb_test_samples,
 630         batch_size=args.physical_batch_size,
 631         size=args.grid_size,
 632         fraction_play=args.grid_fraction_play,
 633         logger=log_string,
 634         device=device,
 635     )
 636
 637 elif args.task == "qmlp":
 638     task = tasks.QMLP(
 639         nb_train_samples=args.nb_train_samples,
 640         nb_test_samples=args.nb_test_samples,
 641         batch_size=args.physical_batch_size,
 642         result_dir=args.result_dir,
 643         logger=log_string,
 644         device=device,
 645     )
 646
 647 elif args.task == "greed":
 648     task = tasks.Greed(
 649         nb_train_samples=args.nb_train_samples,
 650         nb_test_samples=args.nb_test_samples,
 651         batch_size=args.physical_batch_size,
 652         height=args.greed_height,
 653         width=args.greed_width,
 654         T=args.greed_T,
 655         nb_walls=args.greed_nb_walls,
 656         nb_coins=args.greed_nb_coins,
 657         logger=log_string,
 658         device=device,
 659     )
 660
 661 else:
 662     raise ValueError(f"Unknown task {args.task}")
 663
 664 ######################################################################
 665
 666 log_string(f"device {device}")
 667
 668 vocabulary_size = task.vocabulary_size()
 669
 670 log_string(f"vocabulary_size {vocabulary_size}")
 671
 672 ##############################
 673
 674 models = []
 675
 676 for k in range(2):
 677     models.append(
 678         mygpt.MyGPT(
 679             vocabulary_size=vocabulary_size,
 680             dim_model=args.dim_model,
 681             dim_keys=args.dim_keys,
 682             dim_hidden=args.dim_hidden,
 683             nb_heads=args.nb_heads,
 684             nb_blocks=args.nb_blocks,
 685             causal=True,
 686             dropout=args.dropout,
 687         ).to(device)
 688     )
 689
 690
 691 nb_parameters = sum(p.numel() for p in models[0].parameters())
 692 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 693
 694 ######################################################################
 695
 696 # Compute the entropy of the training tokens
 697
 698 token_count = 0
 699 for input in task.batches(split="train", desc="train-entropy"):
 700     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 701 token_probas = token_count / token_count.sum()
 702 entropy = -torch.xlogy(token_probas, token_probas).sum()
 703 train_set_perplexity = math.exp(entropy)
 704
 705 ######################################################################
 706 # A bit of paranoia never hurts
 707
 708 if args.max_percents_of_test_in_train >= 0:
 709
 710     def subsets_as_tuples(batches, cs):
 711         s = set()
 712         for batch in batches:
 713             for x in batch:
 714                 s.add(tuple([v.item() for v in x]))
 715                 if len(s) == cs:
 716                     yield s
 717                     s = set()
 718         yield s
 719
 720     nb_test, nb_in_train = 0, 0
 721     for test_subset in subsets_as_tuples(
 722         task.batches(split="test", desc="test-check"), 25000
 723     ):
 724         in_train = set()
 725         for train_subset in subsets_as_tuples(
 726             task.batches(split="train", desc="train-check"), 25000
 727         ):
 728             in_train.update(test_subset.intersection(train_subset))
 729         nb_in_train += len(in_train)
 730         nb_test += len(test_subset)
 731
 732     log_string(
 733         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 734     )
 735
 736     assert (
 737         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 738     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 739
 740 ##############################
 741
 742 if args.learning_rate_schedule == "cos":
 743     learning_rate_schedule = {}
 744     for n_epoch in range(args.nb_epochs):
 745         u = n_epoch / args.nb_epochs * math.pi
 746         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 747 else:
 748     u = {
 749         int(k): float(v)
 750         for k, v in [
 751             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 752         ]
 753     }
 754
 755     learning_rate_schedule = {}
 756     learning_rate = args.learning_rate
 757     for n_epoch in range(args.nb_epochs):
 758         if n_epoch in u:
 759             learning_rate = u[n_epoch]
 760         learning_rate_schedule[n_epoch] = learning_rate
 761
 762 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 763
 764 time_pred_result = None
 765
 766 ######################################################################
 767
 768
 769 def one_epoch(model, task, learning_rate):
 770     log_string(f"learning_rate {learning_rate}")
 771
 772     if args.optim == "sgd":
 773         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 774     elif args.optim == "adam":
 775         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 776     elif args.optim == "adamw":
 777         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 778     else:
 779         raise ValueError(f"Unknown optimizer {args.optim}.")
 780
 781     model.train()
 782
 783     nb_train_samples, acc_train_loss = 0, 0.0
 784
 785     for input in task.batches(split="train"):
 786         input = input.to(device)
 787
 788         if nb_train_samples % args.batch_size == 0:
 789             optimizer.zero_grad()
 790
 791         output = model(mygpt.BracketedSequence(input)).x
 792         loss = F.cross_entropy(output.transpose(1, 2), input)
 793         acc_train_loss += loss.item() * input.size(0)
 794
 795         nb_train_samples += input.size(0)
 796
 797         loss.backward()
 798
 799         if nb_train_samples % args.batch_size == 0:
 800             optimizer.step()
 801
 802     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 803
 804     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 805
 806
 807 ######################################################################
 808
 809
 810 def run_tests(model, task, deterministic_synthesis):
 811     with torch.autograd.no_grad():
 812         model.eval()
 813
 814         nb_test_samples, acc_test_loss = 0, 0.0
 815         nb_samples_accumulated = 0
 816
 817         for input in task.batches(split="test"):
 818             input = input.to(device)
 819
 820             bs = model(mygpt.BracketedSequence(input))
 821             output = bs.x
 822
 823             loss = F.cross_entropy(output.transpose(1, 2), input)
 824
 825             acc_test_loss += loss.item() * input.size(0)
 826
 827             nb_test_samples += input.size(0)
 828
 829         main_test_accuracy = task.produce_results(
 830             n_epoch=n_epoch,
 831             model=model,
 832             result_dir=args.result_dir,
 833             logger=log_string,
 834             deterministic_synthesis=deterministic_synthesis,
 835         )
 836
 837         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 838
 839         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 840
 841     return main_test_accuracy
 842
 843
 844 ######################################################################
 845
 846
 847 def create_quizzes(
 848     other_models,
 849     task,
 850     nb_for_train=1000,
 851     nb_for_test=100,
 852     nb_runs=10,
 853     nb_min_correct=9,
 854     nb_max_correct=9,
 855 ):
 856     kept = []
 857
 858     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 859         new_quizzes, nb_correct = task.create_new_quizzes(
 860             n_epoch=n_epoch,
 861             result_dir=args.result_dir,
 862             logger=log_string,
 863             nb=4 * (nb_for_train + nb_for_test),
 864             models=other_models,
 865             nb_runs=nb_runs,
 866         )
 867
 868         to_keep = new_quizzes[
 869             torch.logical_and(
 870                 nb_correct >= nb_min_correct, nb_correct <= nb_max_correct
 871             )
 872         ]
 873         log_string(f"keep {to_keep.size(0)} quizzes")
 874         kept.append(to_keep)
 875
 876     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 877
 878     task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
 879     task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
 880
 881     task.save_image(
 882         new_quizzes[:96],
 883         args.result_dir,
 884         f"world_new_{n_epoch:04d}.png",
 885         log_string,
 886     )
 887
 888
 889 ######################################################################
 890
 891 accuracy_to_make_quizzes = 0.95
 892
 893 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 894     learning_rate = learning_rate_schedule[n_epoch]
 895
 896     for m in models:
 897         one_epoch(m, task, learning_rate)
 898         test_accuracy = run_tests(m, task, deterministic_synthesis=False)
 899
 900     if test_accuracy >= accuracy_to_make_quizzes:
 901         other_models = models.copy()
 902         other_models.remove(model)
 903         create_quizzes(other_models, task)
 904
 905     # --------------------------------------------
 906
 907     time_current_result = datetime.datetime.now()
 908     if time_pred_result is not None:
 909         log_string(
 910             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 911         )
 912     time_pred_result = time_current_result
 913
 914 ######################################################################