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