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