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