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=100)
  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=None)
  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     print(f"result directory {args.result_dir} already exists")
 378     exit(1)
 379
 380 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 381
 382 if args.seed >= 0:
 383     # torch.backends.cudnn.deterministic = True
 384     # torch.backends.cudnn.benchmark = False
 385     # torch.use_deterministic_algorithms(True)
 386     torch.manual_seed(args.seed)
 387     if torch.cuda.is_available():
 388         torch.cuda.manual_seed_all(args.seed)
 389
 390 ######################################################################
 391
 392
 393 def log_string(s):
 394     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 395
 396     if log_file is not None:
 397         log_file.write(t + s + "\n")
 398         log_file.flush()
 399
 400     print(t + s)
 401     sys.stdout.flush()
 402
 403
 404 log_string(f"argv {' '.join(sys.argv)}")
 405
 406 for n in vars(args):
 407     log_string(f"args.{n} {getattr(args, n)}")
 408
 409
 410 ######################################################################
 411
 412
 413 def picoclvr_pruner_horizontal_green(p):
 414     return not ("green" in p and ("left" in p or "right" in p))
 415
 416
 417 picoclvr_pruner_train = (
 418     picoclvr_pruner_horizontal_green
 419     if args.picocvlr_prune_properties in {"train+eval"}
 420     else None
 421 )
 422
 423 picoclvr_pruner_eval = (
 424     (lambda p: not picoclvr_pruner_horizontal_green(p))
 425     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 426     else None
 427 )
 428
 429 ######################################################################
 430
 431 if args.physical_batch_size is None:
 432     args.physical_batch_size = args.batch_size
 433 else:
 434     assert args.batch_size % args.physical_batch_size == 0
 435
 436 assert args.nb_train_samples % args.batch_size == 0
 437 assert args.nb_test_samples % args.batch_size == 0
 438
 439 if args.task == "file":
 440     assert (
 441         args.filetask_train_file is not None and args.filetask_test_file is not None
 442     ), "You have to specify the task train and test files"
 443     task = tasks.TaskFromFile(
 444         args.filetask_train_file,
 445         args.filetask_test_file,
 446         nb_train_samples=args.nb_train_samples,
 447         nb_test_samples=args.nb_test_samples,
 448         batch_size=args.physical_batch_size,
 449         shuffle=True,
 450         device=device,
 451     )
 452     args.max_percents_of_test_in_train = 0
 453
 454 elif args.task == "byheart":
 455     task = tasks.SandBox(
 456         problem=problems.ProblemByHeart(separation=args.byheart_separation),
 457         nb_train_samples=args.nb_train_samples,
 458         nb_test_samples=args.nb_test_samples,
 459         batch_size=args.physical_batch_size,
 460         logger=log_string,
 461         device=device,
 462     )
 463     args.max_percents_of_test_in_train = -1
 464
 465 elif args.task == "world":
 466     task = tasks.World(
 467         nb_train_samples=args.nb_train_samples,
 468         nb_test_samples=args.nb_test_samples,
 469         batch_size=args.physical_batch_size,
 470         result_dir=args.result_dir,
 471         logger=log_string,
 472         device=device,
 473     )
 474     args.max_percents_of_test_in_train = -1
 475
 476 elif args.task == "learnop":
 477     task = tasks.SandBox(
 478         problem=problems.ProblemLearnOperator(),
 479         nb_train_samples=args.nb_train_samples,
 480         nb_test_samples=args.nb_test_samples,
 481         batch_size=args.physical_batch_size,
 482         logger=log_string,
 483         device=device,
 484     )
 485
 486
 487 elif args.task == "guessop":
 488     task = tasks.SandBox(
 489         problem=problems.ProblemGuessOperator(),
 490         nb_train_samples=args.nb_train_samples,
 491         nb_test_samples=args.nb_test_samples,
 492         batch_size=args.physical_batch_size,
 493         logger=log_string,
 494         device=device,
 495     )
 496
 497
 498 elif args.task == "twotargets":
 499     task = tasks.SandBox(
 500         problem=problems.ProblemTwoTargets(),
 501         nb_train_samples=args.nb_train_samples,
 502         nb_test_samples=args.nb_test_samples,
 503         batch_size=args.physical_batch_size,
 504         logger=log_string,
 505         device=device,
 506     )
 507
 508 elif args.task == "memory":
 509     task = tasks.SandBox(
 510         problem=problems.ProblemMemory(),
 511         nb_train_samples=args.nb_train_samples,
 512         nb_test_samples=args.nb_test_samples,
 513         batch_size=args.physical_batch_size,
 514         logger=log_string,
 515         device=device,
 516     )
 517
 518 elif args.task == "mixing":
 519     task = tasks.SandBox(
 520         problem=problems.ProblemMixing(
 521             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 522         ),
 523         nb_train_samples=args.nb_train_samples,
 524         nb_test_samples=args.nb_test_samples,
 525         batch_size=args.physical_batch_size,
 526         logger=log_string,
 527         device=device,
 528     )
 529
 530 elif args.task == "addition":
 531     task = tasks.SandBox(
 532         problem=problems.ProblemAddition(),
 533         nb_train_samples=args.nb_train_samples,
 534         nb_test_samples=args.nb_test_samples,
 535         batch_size=args.physical_batch_size,
 536         logger=log_string,
 537         device=device,
 538     )
 539
 540 elif args.task == "picoclvr":
 541     task = tasks.PicoCLVR(
 542         nb_train_samples=args.nb_train_samples,
 543         nb_test_samples=args.nb_test_samples,
 544         batch_size=args.physical_batch_size,
 545         height=args.picoclvr_height,
 546         width=args.picoclvr_width,
 547         nb_colors=args.picoclvr_nb_colors,
 548         logger=log_string,
 549         device=device,
 550         pruner_train=picoclvr_pruner_train,
 551         pruner_eval=picoclvr_pruner_eval,
 552     )
 553
 554 elif args.task == "mnist":
 555     task = tasks.MNIST(
 556         nb_train_samples=args.nb_train_samples,
 557         nb_test_samples=args.nb_test_samples,
 558         batch_size=args.physical_batch_size,
 559         device=device,
 560     )
 561
 562 elif args.task == "maze":
 563     task = tasks.Maze(
 564         nb_train_samples=args.nb_train_samples,
 565         nb_test_samples=args.nb_test_samples,
 566         batch_size=args.physical_batch_size,
 567         height=args.maze_height,
 568         width=args.maze_width,
 569         nb_walls=args.maze_nb_walls,
 570         device="cpu",
 571     )
 572
 573 elif args.task == "snake":
 574     task = tasks.Snake(
 575         nb_train_samples=args.nb_train_samples,
 576         nb_test_samples=args.nb_test_samples,
 577         batch_size=args.physical_batch_size,
 578         height=args.snake_height,
 579         width=args.snake_width,
 580         nb_colors=args.snake_nb_colors,
 581         length=args.snake_length,
 582         prompt_length=args.snake_length // 2,
 583         device=device,
 584     )
 585
 586 elif args.task == "stack":
 587     task = tasks.Stack(
 588         nb_train_samples=args.nb_train_samples,
 589         nb_test_samples=args.nb_test_samples,
 590         batch_size=args.physical_batch_size,
 591         logger=log_string,
 592         nb_steps=args.stack_nb_steps,
 593         nb_stacks=args.stack_nb_stacks,
 594         nb_digits=args.stack_nb_digits,
 595         fraction_values_for_train=args.stack_fraction_values_for_train,
 596         device=device,
 597     )
 598
 599 elif args.task == "expr":
 600     task = tasks.Expr(
 601         nb_train_samples=args.nb_train_samples,
 602         nb_test_samples=args.nb_test_samples,
 603         nb_variables=args.expr_nb_variables,
 604         sequence_length=args.expr_sequence_length,
 605         operand_max=args.expr_operand_max,
 606         result_max=args.expr_result_max,
 607         batch_size=args.physical_batch_size,
 608         device=device,
 609     )
 610
 611 elif args.task == "rpl":
 612     task = tasks.RPL(
 613         nb_train_samples=args.nb_train_samples,
 614         nb_test_samples=args.nb_test_samples,
 615         batch_size=args.physical_batch_size,
 616         nb_starting_values=args.rpl_nb_starting_values,
 617         max_input=args.rpl_max_input,
 618         prog_len=args.rpl_prog_len,
 619         nb_runs=args.rpl_nb_runs,
 620         no_prog=args.rpl_no_prog,
 621         logger=log_string,
 622         device=device,
 623     )
 624
 625 elif args.task == "grid":
 626     task = tasks.Grid(
 627         nb_train_samples=args.nb_train_samples,
 628         nb_test_samples=args.nb_test_samples,
 629         batch_size=args.physical_batch_size,
 630         size=args.grid_size,
 631         fraction_play=args.grid_fraction_play,
 632         logger=log_string,
 633         device=device,
 634     )
 635
 636 elif args.task == "qmlp":
 637     task = tasks.QMLP(
 638         nb_train_samples=args.nb_train_samples,
 639         nb_test_samples=args.nb_test_samples,
 640         batch_size=args.physical_batch_size,
 641         result_dir=args.result_dir,
 642         logger=log_string,
 643         device=device,
 644     )
 645
 646 elif args.task == "greed":
 647     task = tasks.Greed(
 648         nb_train_samples=args.nb_train_samples,
 649         nb_test_samples=args.nb_test_samples,
 650         batch_size=args.physical_batch_size,
 651         height=args.greed_height,
 652         width=args.greed_width,
 653         T=args.greed_T,
 654         nb_walls=args.greed_nb_walls,
 655         nb_coins=args.greed_nb_coins,
 656         logger=log_string,
 657         device=device,
 658     )
 659
 660 else:
 661     raise ValueError(f"Unknown task {args.task}")
 662
 663 ######################################################################
 664
 665 log_string(f"device {device}")
 666
 667 vocabulary_size = task.vocabulary_size()
 668
 669 log_string(f"vocabulary_size {vocabulary_size}")
 670
 671 ##############################
 672
 673 models = []
 674
 675 for k in range(2):
 676     models.append(
 677         mygpt.MyGPT(
 678             vocabulary_size=vocabulary_size,
 679             dim_model=args.dim_model,
 680             dim_keys=args.dim_keys,
 681             dim_hidden=args.dim_hidden,
 682             nb_heads=args.nb_heads,
 683             nb_blocks=args.nb_blocks,
 684             causal=True,
 685             dropout=args.dropout,
 686         ).to(device)
 687     )
 688
 689
 690 nb_parameters = sum(p.numel() for p in models[0].parameters())
 691 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 692
 693 ######################################################################
 694
 695 # Compute the entropy of the training tokens
 696
 697 token_count = 0
 698 for input in task.batches(split="train", desc="train-entropy"):
 699     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 700 token_probas = token_count / token_count.sum()
 701 entropy = -torch.xlogy(token_probas, token_probas).sum()
 702 train_set_perplexity = math.exp(entropy)
 703
 704 ######################################################################
 705 # A bit of paranoia never hurts
 706
 707 if args.max_percents_of_test_in_train >= 0:
 708
 709     def subsets_as_tuples(batches, cs):
 710         s = set()
 711         for batch in batches:
 712             for x in batch:
 713                 s.add(tuple([v.item() for v in x]))
 714                 if len(s) == cs:
 715                     yield s
 716                     s = set()
 717         yield s
 718
 719     nb_test, nb_in_train = 0, 0
 720     for test_subset in subsets_as_tuples(
 721         task.batches(split="test", desc="test-check"), 25000
 722     ):
 723         in_train = set()
 724         for train_subset in subsets_as_tuples(
 725             task.batches(split="train", desc="train-check"), 25000
 726         ):
 727             in_train.update(test_subset.intersection(train_subset))
 728         nb_in_train += len(in_train)
 729         nb_test += len(test_subset)
 730
 731     log_string(
 732         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 733     )
 734
 735     assert (
 736         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 737     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 738
 739 ##############################
 740
 741 if args.learning_rate_schedule == "cos":
 742     learning_rate_schedule = {}
 743     for n_epoch in range(args.nb_epochs):
 744         u = n_epoch / args.nb_epochs * math.pi
 745         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 746 else:
 747     if args.learning_rate_schedule is not None:
 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     else:
 755         u = {}
 756
 757     learning_rate_schedule = {}
 758     learning_rate = args.learning_rate
 759     for n_epoch in range(args.nb_epochs):
 760         if n_epoch in u:
 761             learning_rate = u[n_epoch]
 762         learning_rate_schedule[n_epoch] = learning_rate
 763
 764 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 765
 766 time_pred_result = None
 767
 768 ######################################################################
 769
 770
 771 def one_epoch(model, task, learning_rate):
 772     log_string(f"learning_rate {learning_rate}")
 773
 774     if args.optim == "sgd":
 775         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 776     elif args.optim == "adam":
 777         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 778     elif args.optim == "adamw":
 779         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 780     else:
 781         raise ValueError(f"Unknown optimizer {args.optim}.")
 782
 783     model.train()
 784
 785     nb_train_samples, acc_train_loss = 0, 0.0
 786
 787     for input in task.batches(split="train"):
 788         input = input.to(device)
 789
 790         if nb_train_samples % args.batch_size == 0:
 791             optimizer.zero_grad()
 792
 793         output = model(mygpt.BracketedSequence(input)).x
 794         loss = F.cross_entropy(output.transpose(1, 2), input)
 795         acc_train_loss += loss.item() * input.size(0)
 796
 797         nb_train_samples += input.size(0)
 798
 799         loss.backward()
 800
 801         if nb_train_samples % args.batch_size == 0:
 802             optimizer.step()
 803
 804     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 805
 806     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 807
 808
 809 ######################################################################
 810
 811
 812 def run_tests(model, task, deterministic_synthesis):
 813     with torch.autograd.no_grad():
 814         model.eval()
 815
 816         nb_test_samples, acc_test_loss = 0, 0.0
 817         nb_samples_accumulated = 0
 818
 819         for input in task.batches(split="test"):
 820             input = input.to(device)
 821
 822             bs = model(mygpt.BracketedSequence(input))
 823             output = bs.x
 824
 825             loss = F.cross_entropy(output.transpose(1, 2), input)
 826
 827             acc_test_loss += loss.item() * input.size(0)
 828
 829             nb_test_samples += input.size(0)
 830
 831         main_test_accuracy = task.produce_results(
 832             n_epoch=n_epoch,
 833             model=model,
 834             result_dir=args.result_dir,
 835             logger=log_string,
 836             deterministic_synthesis=deterministic_synthesis,
 837         )
 838
 839         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 840
 841         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 842
 843     return main_test_accuracy
 844
 845
 846 ######################################################################
 847
 848
 849 def create_quizzes(
 850     model,
 851     other_models,
 852     task,
 853     nb_for_train=1000,
 854     nb_for_test=100,
 855     nb_runs=10,
 856     nb_min_correct=9,
 857     nb_max_correct=9,
 858 ):
 859     kept = []
 860
 861     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 862         new_quizzes, nb_correct = task.create_new_quizzes(
 863             n_epoch=n_epoch,
 864             result_dir=args.result_dir,
 865             logger=log_string,
 866             nb=4 * (nb_for_train + nb_for_test),
 867             model=model,
 868             other_models=other_models,
 869             nb_runs=nb_runs,
 870         )
 871
 872         to_keep = new_quizzes[
 873             torch.logical_and(
 874                 nb_correct >= nb_min_correct, nb_correct <= nb_max_correct
 875             )
 876         ]
 877         log_string(f"keep {to_keep.size(0)} quizzes")
 878         kept.append(to_keep)
 879
 880     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 881
 882     task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
 883     task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
 884
 885     task.save_image(
 886         new_quizzes[:96],
 887         args.result_dir,
 888         f"world_new_{n_epoch:04d}.png",
 889         log_string,
 890     )
 891
 892
 893 ######################################################################
 894
 895 accuracy_to_make_quizzes = 0.975
 896
 897 for n_epoch in range(args.nb_epochs):
 898     learning_rate = learning_rate_schedule[n_epoch]
 899
 900     for m in models:
 901         one_epoch(m, task, learning_rate)
 902         test_accuracy = run_tests(m, task, deterministic_synthesis=False)
 903
 904         if test_accuracy >= accuracy_to_make_quizzes:
 905             other_models = models.copy()
 906             other_models.remove(m)
 907             create_quizzes(m, other_models, task)
 908
 909     # --------------------------------------------
 910
 911     time_current_result = datetime.datetime.now()
 912     if time_pred_result is not None:
 913         log_string(
 914             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 915         )
 916     time_pred_result = time_current_result
 917
 918 ######################################################################