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