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
   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=25)
  50
  51 parser.add_argument("--batch_size", type=int, default=None)
  52
  53 parser.add_argument("--nb_train_samples", type=int, default=None)
  54
  55 parser.add_argument("--nb_test_samples", type=int, default=None)
  56
  57 parser.add_argument("--optim", type=str, default="adam")
  58
  59 parser.add_argument("--learning_rate", type=float, default=1e-4)
  60
  61 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  62
  63 ########################################
  64
  65 parser.add_argument("--model", type=str, default=None)
  66
  67 parser.add_argument("--dim_model", type=int, default=None)
  68
  69 parser.add_argument("--dim_keys", type=int, default=None)
  70
  71 parser.add_argument("--dim_hidden", type=int, default=None)
  72
  73 parser.add_argument("--nb_heads", type=int, default=None)
  74
  75 parser.add_argument("--nb_blocks", type=int, default=None)
  76
  77 parser.add_argument("--dropout", type=float, default=0.1)
  78
  79 ########################################
  80
  81 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  82
  83 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  84
  85 parser.add_argument("--overwrite_results", action="store_true", default=False)
  86
  87 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  88
  89 ##############################
  90 # filetask
  91
  92 parser.add_argument("--filetask_train_file", type=str, default=None)
  93
  94 parser.add_argument("--filetask_test_file", type=str, default=None)
  95
  96 ##############################
  97 # rpl options
  98
  99 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 100
 101 parser.add_argument("--rpl_max_input", type=int, default=9)
 102
 103 parser.add_argument("--rpl_prog_len", type=int, default=8)
 104
 105 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 106
 107 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 108
 109 ##############################
 110 # grid options
 111
 112 parser.add_argument("--grid_size", type=int, default=6)
 113
 114 parser.add_argument("--grid_fraction_play", type=float, default=0)
 115
 116 ##############################
 117 # picoclvr options
 118
 119 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 120
 121 parser.add_argument("--picoclvr_height", type=int, default=12)
 122
 123 parser.add_argument("--picoclvr_width", type=int, default=16)
 124
 125 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 126
 127 ##############################
 128 # Maze options
 129
 130 parser.add_argument("--maze_height", type=int, default=13)
 131
 132 parser.add_argument("--maze_width", type=int, default=21)
 133
 134 parser.add_argument("--maze_nb_walls", type=int, default=15)
 135
 136 ##############################
 137 # Snake options
 138
 139 parser.add_argument("--snake_height", type=int, default=9)
 140
 141 parser.add_argument("--snake_width", type=int, default=12)
 142
 143 parser.add_argument("--snake_nb_colors", type=int, default=5)
 144
 145 parser.add_argument("--snake_length", type=int, default=200)
 146
 147 ##############################
 148 # Stack options
 149
 150 parser.add_argument("--stack_nb_steps", type=int, default=100)
 151
 152 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 153
 154 parser.add_argument("--stack_nb_digits", type=int, default=3)
 155
 156 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 157
 158 ##############################
 159 # Expr options
 160
 161 parser.add_argument("--expr_nb_variables", type=int, default=5)
 162
 163 parser.add_argument("--expr_sequence_length", type=int, default=40)
 164
 165 parser.add_argument("--expr_operand_max", type=int, default=9)
 166
 167 parser.add_argument("--expr_result_max", type=int, default=99)
 168
 169 parser.add_argument("--expr_input_file", type=str, default=None)
 170
 171 ##############################
 172 # Mixing
 173
 174 parser.add_argument("--mixing_hard", action="store_true", default=False)
 175
 176 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 177
 178 ##############################
 179 # greed options
 180
 181 parser.add_argument("--greed_height", type=int, default=5)
 182
 183 parser.add_argument("--greed_width", type=int, default=7)
 184
 185 parser.add_argument("--greed_T", type=int, default=25)
 186
 187 parser.add_argument("--greed_nb_walls", type=int, default=5)
 188
 189 parser.add_argument("--greed_nb_coins", type=int, default=2)
 190
 191 ######################################################################
 192
 193 args = parser.parse_args()
 194
 195 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 196
 197 if args.result_dir is None:
 198     args.result_dir = f"results_{args.task}"
 199
 200 ######################################################################
 201
 202 default_task_args = {
 203     "file": {
 204         "model": "37M",
 205         "batch_size": 25,
 206         "nb_train_samples": 250000,
 207         "nb_test_samples": 10000,
 208     },
 209     "addition": {
 210         "model": "352M",
 211         "batch_size": 25,
 212         "nb_train_samples": 250000,
 213         "nb_test_samples": 10000,
 214     },
 215     "byheart": {
 216         "model": "37M",
 217         "batch_size": 25,
 218         "nb_train_samples": 50000,
 219         "nb_test_samples": 10000,
 220     },
 221     "expr": {
 222         "model": "352M",
 223         "batch_size": 25,
 224         "nb_train_samples": 2500000,
 225         "nb_test_samples": 10000,
 226     },
 227     "grid": {
 228         "model": "37M",
 229         "batch_size": 25,
 230         "nb_train_samples": 250000,
 231         "nb_test_samples": 10000,
 232     },
 233     "qmlp": {
 234         "model": "37M",
 235         "batch_size": 10,
 236         "nb_train_samples": 100000,
 237         "nb_test_samples": 1000,
 238     },
 239     "guessop": {
 240         "model": "352M",
 241         "batch_size": 25,
 242         "nb_train_samples": 1000000,
 243         "nb_test_samples": 10000,
 244     },
 245     "learnop": {
 246         "model": "37M",
 247         "batch_size": 25,
 248         "nb_train_samples": 50000,
 249         "nb_test_samples": 10000,
 250     },
 251     "maze": {
 252         "model": "37M",
 253         "batch_size": 5,
 254         "nb_train_samples": 100000,
 255         "nb_test_samples": 10000,
 256     },
 257     "picoclvr": {
 258         "model": "37M",
 259         "batch_size": 25,
 260         "nb_train_samples": 250000,
 261         "nb_test_samples": 10000,
 262     },
 263     "rpl": {
 264         "model": "352M",
 265         "batch_size": 5,
 266         "nb_train_samples": 2500000,
 267         "nb_test_samples": 10000,
 268     },
 269     "snake": {
 270         "model": "37M",
 271         "batch_size": 25,
 272         "nb_train_samples": 250000,
 273         "nb_test_samples": 10000,
 274     },
 275     "stack": {
 276         "model": "37M",
 277         "batch_size": 25,
 278         "nb_train_samples": 100000,
 279         "nb_test_samples": 1000,
 280     },
 281     "twotargets": {
 282         "model": "37M",
 283         "batch_size": 25,
 284         "nb_train_samples": 50000,
 285         "nb_test_samples": 10000,
 286     },
 287     "memory": {
 288         "model": "37M",
 289         "batch_size": 100,
 290         "nb_train_samples": 25000,
 291         "nb_test_samples": 1000,
 292     },
 293     "mixing": {
 294         "model": "37M",
 295         "batch_size": 25,
 296         "nb_train_samples": 250000,
 297         "nb_test_samples": 10000,
 298     },
 299     "mnist": {
 300         "model": "37M",
 301         "batch_size": 10,
 302         "nb_train_samples": 60000,
 303         "nb_test_samples": 10000,
 304     },
 305     "greed": {
 306         "model": "37M",
 307         "batch_size": 25,
 308         "nb_train_samples": 25000,
 309         "nb_test_samples": 10000,
 310     },
 311 }
 312
 313 if args.task in default_task_args:
 314     for k, v in default_task_args[args.task].items():
 315         if getattr(args, k) is None:
 316             setattr(args, k, v)
 317
 318 ######################################################################
 319
 320 default_model_args = {
 321     "17K": {
 322         "dim_model": 32,
 323         "dim_keys": 32,
 324         "dim_hidden": 32,
 325         "nb_heads": 2,
 326         "nb_blocks": 2,
 327     },
 328     "4M": {
 329         "dim_model": 256,
 330         "dim_keys": 32,
 331         "dim_hidden": 1024,
 332         "nb_heads": 4,
 333         "nb_blocks": 6,
 334     },
 335     "37M": {
 336         "dim_model": 512,
 337         "dim_keys": 64,
 338         "dim_hidden": 2048,
 339         "nb_heads": 8,
 340         "nb_blocks": 12,
 341     },
 342     "122M": {
 343         "dim_model": 768,
 344         "dim_keys": 64,
 345         "dim_hidden": 2048,
 346         "nb_heads": 8,
 347         "nb_blocks": 24,
 348     },
 349     "352M": {
 350         "dim_model": 1024,
 351         "dim_keys": 64,
 352         "dim_hidden": 2048,
 353         "nb_heads": 8,
 354         "nb_blocks": 48,
 355     },
 356 }
 357
 358 if args.model in default_model_args:
 359     for k, v in default_model_args[args.model].items():
 360         if getattr(args, k) is None:
 361             setattr(args, k, v)
 362 else:
 363     raise ValueError(f"Unknown model {args.model}")
 364
 365 ######################################################################
 366
 367 try:
 368     os.mkdir(args.result_dir)
 369 except FileExistsError:
 370     if not args.overwrite_results:
 371         print(f"result directory {args.result_dir} already exists")
 372         exit(1)
 373
 374 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 375
 376 if args.seed >= 0:
 377     # torch.backends.cudnn.deterministic = True
 378     # torch.backends.cudnn.benchmark = False
 379     # torch.use_deterministic_algorithms(True)
 380     torch.manual_seed(args.seed)
 381     if torch.cuda.is_available():
 382         torch.cuda.manual_seed_all(args.seed)
 383
 384 ######################################################################
 385
 386
 387 def log_string(s):
 388     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 389
 390     if log_file is not None:
 391         log_file.write(t + s + "\n")
 392         log_file.flush()
 393
 394     print(t + s)
 395     sys.stdout.flush()
 396
 397
 398 log_string(f"argv {' '.join(sys.argv)}")
 399
 400 for n in vars(args):
 401     log_string(f"args.{n} {getattr(args, n)}")
 402
 403
 404 ######################################################################
 405
 406
 407 def picoclvr_pruner_horizontal_green(p):
 408     return not ("green" in p and ("left" in p or "right" in p))
 409
 410
 411 picoclvr_pruner_train = (
 412     picoclvr_pruner_horizontal_green
 413     if args.picocvlr_prune_properties in {"train+eval"}
 414     else None
 415 )
 416
 417 picoclvr_pruner_eval = (
 418     (lambda p: not picoclvr_pruner_horizontal_green(p))
 419     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 420     else None
 421 )
 422
 423 ######################################################################
 424
 425 if args.task == "file":
 426     assert (
 427         args.filetask_train_file is not None and args.filetask_test_file is not None
 428     ), "You have to specify the task train and test files"
 429     task = tasks.TaskFromFile(
 430         args.filetask_train_file,
 431         args.filetask_test_file,
 432         nb_train_samples=args.nb_train_samples,
 433         nb_test_samples=args.nb_test_samples,
 434         batch_size=args.batch_size,
 435         shuffle=True,
 436         device=device,
 437     )
 438     args.max_percents_of_test_in_train = 0
 439
 440 elif args.task == "byheart":
 441     task = tasks.SandBox(
 442         problem=problems.ProblemByHeart(),
 443         nb_train_samples=args.nb_train_samples,
 444         nb_test_samples=args.nb_test_samples,
 445         batch_size=args.batch_size,
 446         logger=log_string,
 447         device=device,
 448     )
 449     args.max_percents_of_test_in_train = -1
 450
 451 elif args.task == "learnop":
 452     task = tasks.SandBox(
 453         problem=problems.ProblemLearnOperator(),
 454         nb_train_samples=args.nb_train_samples,
 455         nb_test_samples=args.nb_test_samples,
 456         batch_size=args.batch_size,
 457         logger=log_string,
 458         device=device,
 459     )
 460
 461
 462 elif args.task == "guessop":
 463     task = tasks.SandBox(
 464         problem=problems.ProblemGuessOperator(),
 465         nb_train_samples=args.nb_train_samples,
 466         nb_test_samples=args.nb_test_samples,
 467         batch_size=args.batch_size,
 468         logger=log_string,
 469         device=device,
 470     )
 471
 472
 473 elif args.task == "twotargets":
 474     task = tasks.SandBox(
 475         problem=problems.ProblemTwoTargets(),
 476         nb_train_samples=args.nb_train_samples,
 477         nb_test_samples=args.nb_test_samples,
 478         batch_size=args.batch_size,
 479         logger=log_string,
 480         device=device,
 481     )
 482
 483 elif args.task == "memory":
 484     task = tasks.SandBox(
 485         problem=problems.ProblemMemory(),
 486         nb_train_samples=args.nb_train_samples,
 487         nb_test_samples=args.nb_test_samples,
 488         batch_size=args.batch_size,
 489         logger=log_string,
 490         device=device,
 491     )
 492
 493 elif args.task == "mixing":
 494     task = tasks.SandBox(
 495         problem=problems.ProblemMixing(
 496             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 497         ),
 498         nb_train_samples=args.nb_train_samples,
 499         nb_test_samples=args.nb_test_samples,
 500         batch_size=args.batch_size,
 501         logger=log_string,
 502         device=device,
 503     )
 504
 505 elif args.task == "addition":
 506     task = tasks.SandBox(
 507         problem=problems.ProblemAddition(),
 508         nb_train_samples=args.nb_train_samples,
 509         nb_test_samples=args.nb_test_samples,
 510         batch_size=args.batch_size,
 511         logger=log_string,
 512         device=device,
 513     )
 514
 515 elif args.task == "picoclvr":
 516     task = tasks.PicoCLVR(
 517         nb_train_samples=args.nb_train_samples,
 518         nb_test_samples=args.nb_test_samples,
 519         batch_size=args.batch_size,
 520         height=args.picoclvr_height,
 521         width=args.picoclvr_width,
 522         nb_colors=args.picoclvr_nb_colors,
 523         logger=log_string,
 524         device=device,
 525         pruner_train=picoclvr_pruner_train,
 526         pruner_eval=picoclvr_pruner_eval,
 527     )
 528
 529 elif args.task == "mnist":
 530     task = tasks.MNIST(
 531         nb_train_samples=args.nb_train_samples,
 532         nb_test_samples=args.nb_test_samples,
 533         batch_size=args.batch_size,
 534         device=device,
 535     )
 536
 537 elif args.task == "maze":
 538     task = tasks.Maze(
 539         nb_train_samples=args.nb_train_samples,
 540         nb_test_samples=args.nb_test_samples,
 541         batch_size=args.batch_size,
 542         height=args.maze_height,
 543         width=args.maze_width,
 544         nb_walls=args.maze_nb_walls,
 545         device=device,
 546     )
 547
 548 elif args.task == "snake":
 549     task = tasks.Snake(
 550         nb_train_samples=args.nb_train_samples,
 551         nb_test_samples=args.nb_test_samples,
 552         batch_size=args.batch_size,
 553         height=args.snake_height,
 554         width=args.snake_width,
 555         nb_colors=args.snake_nb_colors,
 556         length=args.snake_length,
 557         prompt_length=args.snake_length // 2,
 558         device=device,
 559     )
 560
 561 elif args.task == "stack":
 562     task = tasks.Stack(
 563         nb_train_samples=args.nb_train_samples,
 564         nb_test_samples=args.nb_test_samples,
 565         batch_size=args.batch_size,
 566         logger=log_string,
 567         nb_steps=args.stack_nb_steps,
 568         nb_stacks=args.stack_nb_stacks,
 569         nb_digits=args.stack_nb_digits,
 570         fraction_values_for_train=args.stack_fraction_values_for_train,
 571         device=device,
 572     )
 573
 574 elif args.task == "expr":
 575     task = tasks.Expr(
 576         nb_train_samples=args.nb_train_samples,
 577         nb_test_samples=args.nb_test_samples,
 578         nb_variables=args.expr_nb_variables,
 579         sequence_length=args.expr_sequence_length,
 580         operand_max=args.expr_operand_max,
 581         result_max=args.expr_result_max,
 582         batch_size=args.batch_size,
 583         device=device,
 584     )
 585
 586 elif args.task == "rpl":
 587     task = tasks.RPL(
 588         nb_train_samples=args.nb_train_samples,
 589         nb_test_samples=args.nb_test_samples,
 590         batch_size=args.batch_size,
 591         nb_starting_values=args.rpl_nb_starting_values,
 592         max_input=args.rpl_max_input,
 593         prog_len=args.rpl_prog_len,
 594         nb_runs=args.rpl_nb_runs,
 595         no_prog=args.rpl_no_prog,
 596         logger=log_string,
 597         device=device,
 598     )
 599
 600 elif args.task == "grid":
 601     task = tasks.Grid(
 602         nb_train_samples=args.nb_train_samples,
 603         nb_test_samples=args.nb_test_samples,
 604         batch_size=args.batch_size,
 605         size=args.grid_size,
 606         fraction_play=args.grid_fraction_play,
 607         logger=log_string,
 608         device=device,
 609     )
 610
 611 elif args.task == "qmlp":
 612     task = tasks.QMLP(
 613         nb_train_samples=args.nb_train_samples,
 614         nb_test_samples=args.nb_test_samples,
 615         batch_size=args.batch_size,
 616         result_dir=args.result_dir,
 617         logger=log_string,
 618         device=device,
 619     )
 620
 621 elif args.task == "greed":
 622     task = tasks.Greed(
 623         nb_train_samples=args.nb_train_samples,
 624         nb_test_samples=args.nb_test_samples,
 625         batch_size=args.batch_size,
 626         height=args.greed_height,
 627         width=args.greed_width,
 628         T=args.greed_T,
 629         nb_walls=args.greed_nb_walls,
 630         nb_coins=args.greed_nb_coins,
 631         logger=log_string,
 632         device=device,
 633     )
 634
 635 else:
 636     raise ValueError(f"Unknown task {args.task}")
 637
 638 ######################################################################
 639
 640 log_string(f"device {device}")
 641
 642 vocabulary_size = task.vocabulary_size()
 643
 644 log_string(f"vocabulary_size {vocabulary_size}")
 645
 646 ##############################
 647
 648 model = mygpt.MyGPT(
 649     vocabulary_size=vocabulary_size,
 650     dim_model=args.dim_model,
 651     dim_keys=args.dim_keys,
 652     dim_hidden=args.dim_hidden,
 653     nb_heads=args.nb_heads,
 654     nb_blocks=args.nb_blocks,
 655     causal=True,
 656     dropout=args.dropout,
 657 )
 658
 659 model.to(device)
 660
 661 nb_parameters = sum(p.numel() for p in model.parameters())
 662 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 663
 664 ######################################################################
 665
 666 nb_epochs_finished = 0
 667
 668 if args.no_checkpoint:
 669     log_string(f"not trying to load checkpoint.")
 670
 671 else:
 672     try:
 673         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 674         checkpoint = torch.load(checkpoint_name)
 675         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 676         model.load_state_dict(checkpoint["model_state"])
 677         torch.set_rng_state(checkpoint["rng_state"])
 678         if torch.cuda.is_available():
 679             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 680
 681         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 682
 683     except FileNotFoundError:
 684         log_string("starting from scratch.")
 685
 686     except:
 687         log_string("error when loading the checkpoint.")
 688         exit(1)
 689
 690 ######################################################################
 691
 692 if args.task == "expr" and args.expr_input_file is not None:
 693     task.produce_results(
 694         n_epoch=nb_epochs_finished,
 695         model=model,
 696         result_dir=args.result_dir,
 697         logger=log_string,
 698         deterministic_synthesis=args.deterministic_synthesis,
 699         input_file=args.expr_input_file,
 700     )
 701
 702     exit(0)
 703
 704 ######################################################################
 705
 706 # Compute the entropy of the training tokens
 707
 708 token_count = 0
 709 for input in task.batches(split="train", desc="train-entropy"):
 710     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 711 token_probas = token_count / token_count.sum()
 712 entropy = -torch.xlogy(token_probas, token_probas).sum()
 713 train_set_perplexity = math.exp(entropy)
 714
 715 ######################################################################
 716 # A bit of paranoia never hurts
 717
 718 if args.max_percents_of_test_in_train >= 0:
 719
 720     def subsets_as_tuples(batches, cs):
 721         s = set()
 722         for batch in batches:
 723             for x in batch:
 724                 s.add(tuple([v.item() for v in x]))
 725                 if len(s) == cs:
 726                     yield s
 727                     s = set()
 728         yield s
 729
 730     nb_test, nb_in_train = 0, 0
 731     for test_subset in subsets_as_tuples(
 732         task.batches(split="test", desc="test-check"), 25000
 733     ):
 734         in_train = set()
 735         for train_subset in subsets_as_tuples(
 736             task.batches(split="train", desc="train-check"), 25000
 737         ):
 738             in_train.update(test_subset.intersection(train_subset))
 739         nb_in_train += len(in_train)
 740         nb_test += len(test_subset)
 741
 742     log_string(
 743         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 744     )
 745
 746     assert (
 747         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 748     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 749
 750 ##############################
 751
 752 if args.learning_rate_schedule == "cos":
 753     learning_rate_schedule = {}
 754     for n_epoch in range(args.nb_epochs):
 755         u = n_epoch / args.nb_epochs * math.pi
 756         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 757 else:
 758     u = {
 759         int(k): float(v)
 760         for k, v in [
 761             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 762         ]
 763     }
 764
 765     learning_rate_schedule = {}
 766     learning_rate = args.learning_rate
 767     for n_epoch in range(args.nb_epochs):
 768         if n_epoch in u:
 769             learning_rate = u[n_epoch]
 770         learning_rate_schedule[n_epoch] = learning_rate
 771
 772 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 773
 774 ##############################
 775
 776 nb_samples_seen = 0
 777
 778 if nb_epochs_finished >= args.nb_epochs:
 779     task.produce_results(
 780         n_epoch=nb_epochs_finished,
 781         model=model,
 782         result_dir=args.result_dir,
 783         logger=log_string,
 784         deterministic_synthesis=args.deterministic_synthesis,
 785     )
 786
 787 time_pred_result = None
 788
 789 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 790     learning_rate = learning_rate_schedule[n_epoch]
 791
 792     log_string(f"learning_rate {learning_rate}")
 793
 794     if args.optim == "sgd":
 795         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 796     elif args.optim == "adam":
 797         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 798     elif args.optim == "adamw":
 799         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 800     else:
 801         raise ValueError(f"Unknown optimizer {args.optim}.")
 802
 803     model.train()
 804
 805     nb_train_samples, acc_train_loss = 0, 0.0
 806
 807     for input in task.batches(split="train"):
 808         input = input.to(device)
 809         output = model(mygpt.BracketedSequence(input)).x
 810         loss = F.cross_entropy(output.transpose(1, 2), input)
 811         acc_train_loss += loss.item() * input.size(0)
 812         nb_train_samples += input.size(0)
 813         nb_samples_seen += input.size(0)
 814
 815         optimizer.zero_grad()
 816         loss.backward()
 817         optimizer.step()
 818
 819     with torch.autograd.no_grad():
 820         model.eval()
 821
 822         nb_test_samples, acc_test_loss = 0, 0.0
 823
 824         for input in task.batches(split="test"):
 825             input = input.to(device)
 826
 827             output = model(mygpt.BracketedSequence(input)).x
 828             loss = F.cross_entropy(output.transpose(1, 2), input)
 829             acc_test_loss += loss.item() * input.size(0)
 830             nb_test_samples += input.size(0)
 831
 832         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 833         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 834
 835         log_string(
 836             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 837         )
 838
 839         task.produce_results(
 840             n_epoch=n_epoch,
 841             model=model,
 842             result_dir=args.result_dir,
 843             logger=log_string,
 844             deterministic_synthesis=args.deterministic_synthesis,
 845         )
 846
 847         time_current_result = datetime.datetime.now()
 848         if time_pred_result is not None:
 849             log_string(
 850                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 851             )
 852         time_pred_result = time_current_result
 853
 854     checkpoint = {
 855         "nb_epochs_finished": n_epoch + 1,
 856         "model_state": model.state_dict(),
 857         "rng_state": torch.get_rng_state(),
 858     }
 859
 860     if torch.cuda.is_available():
 861         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 862
 863     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 864     torch.save(checkpoint, checkpoint_name)
 865     log_string(f"saved checkpoint {checkpoint_name}")
 866
 867 ######################################################################