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 # Misc
 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 for n in vars(args):
 352     log_string(f"args.{n} {getattr(args, n)}")
 353
 354
 355 ######################################################################
 356
 357
 358 def picoclvr_pruner_horizontal_green(p):
 359     return not ("green" in p and ("left" in p or "right" in p))
 360
 361
 362 picoclvr_pruner_train = (
 363     picoclvr_pruner_horizontal_green
 364     if args.picocvlr_prune_properties in {"train+eval"}
 365     else None
 366 )
 367
 368 picoclvr_pruner_eval = (
 369     (lambda p: not picoclvr_pruner_horizontal_green(p))
 370     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 371     else None
 372 )
 373
 374 ######################################################################
 375
 376 if args.task == "byheart":
 377     task = tasks.SandBox(
 378         problem=problems.ProblemByHeart(),
 379         nb_train_samples=args.nb_train_samples,
 380         nb_test_samples=args.nb_test_samples,
 381         batch_size=args.batch_size,
 382         logger=log_string,
 383         device=device,
 384     )
 385     args.max_percents_of_test_in_train = -1
 386
 387 elif args.task == "learnop":
 388     task = tasks.SandBox(
 389         problem=problems.ProblemLearnOperator(),
 390         nb_train_samples=args.nb_train_samples,
 391         nb_test_samples=args.nb_test_samples,
 392         batch_size=args.batch_size,
 393         logger=log_string,
 394         device=device,
 395     )
 396
 397
 398 elif args.task == "guessop":
 399     task = tasks.SandBox(
 400         problem=problems.ProblemGuessOperator(),
 401         nb_train_samples=args.nb_train_samples,
 402         nb_test_samples=args.nb_test_samples,
 403         batch_size=args.batch_size,
 404         logger=log_string,
 405         device=device,
 406     )
 407
 408
 409 elif args.task == "twotargets":
 410     task = tasks.SandBox(
 411         problem=problems.ProblemTwoTargets(),
 412         nb_train_samples=args.nb_train_samples,
 413         nb_test_samples=args.nb_test_samples,
 414         batch_size=args.batch_size,
 415         logger=log_string,
 416         device=device,
 417     )
 418
 419 elif args.task == "mixing":
 420     task = tasks.SandBox(
 421         problem=problems.ProblemMixing(
 422             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 423         ),
 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
 431 elif args.task == "addition":
 432     task = tasks.SandBox(
 433         problem=problems.ProblemAddition(),
 434         nb_train_samples=args.nb_train_samples,
 435         nb_test_samples=args.nb_test_samples,
 436         batch_size=args.batch_size,
 437         logger=log_string,
 438         device=device,
 439     )
 440
 441 elif args.task == "picoclvr":
 442     task = tasks.PicoCLVR(
 443         nb_train_samples=args.nb_train_samples,
 444         nb_test_samples=args.nb_test_samples,
 445         batch_size=args.batch_size,
 446         height=args.picoclvr_height,
 447         width=args.picoclvr_width,
 448         nb_colors=args.picoclvr_nb_colors,
 449         logger=log_string,
 450         device=device,
 451         pruner_train=picoclvr_pruner_train,
 452         pruner_eval=picoclvr_pruner_eval,
 453     )
 454
 455 elif args.task == "mnist":
 456     task = tasks.MNIST(
 457         nb_train_samples=args.nb_train_samples,
 458         nb_test_samples=args.nb_test_samples,
 459         batch_size=args.batch_size,
 460         device=device,
 461     )
 462
 463 elif args.task == "maze":
 464     task = tasks.Maze(
 465         nb_train_samples=args.nb_train_samples,
 466         nb_test_samples=args.nb_test_samples,
 467         batch_size=args.batch_size,
 468         height=args.maze_height,
 469         width=args.maze_width,
 470         nb_walls=args.maze_nb_walls,
 471         device=device,
 472     )
 473
 474 elif args.task == "snake":
 475     task = tasks.Snake(
 476         nb_train_samples=args.nb_train_samples,
 477         nb_test_samples=args.nb_test_samples,
 478         batch_size=args.batch_size,
 479         height=args.snake_height,
 480         width=args.snake_width,
 481         nb_colors=args.snake_nb_colors,
 482         length=args.snake_length,
 483         prompt_length=args.snake_length // 2,
 484         device=device,
 485     )
 486
 487 elif args.task == "stack":
 488     task = tasks.Stack(
 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         nb_steps=args.stack_nb_steps,
 494         nb_stacks=args.stack_nb_stacks,
 495         nb_digits=args.stack_nb_digits,
 496         fraction_values_for_train=args.stack_fraction_values_for_train,
 497         device=device,
 498     )
 499
 500 elif args.task == "expr":
 501     task = tasks.Expr(
 502         nb_train_samples=args.nb_train_samples,
 503         nb_test_samples=args.nb_test_samples,
 504         nb_variables=args.expr_nb_variables,
 505         sequence_length=args.expr_sequence_length,
 506         operand_max=args.expr_operand_max,
 507         result_max=args.expr_result_max,
 508         batch_size=args.batch_size,
 509         device=device,
 510     )
 511
 512 elif args.task == "rpl":
 513     task = tasks.RPL(
 514         nb_train_samples=args.nb_train_samples,
 515         nb_test_samples=args.nb_test_samples,
 516         batch_size=args.batch_size,
 517         nb_starting_values=args.rpl_nb_starting_values,
 518         max_input=args.rpl_max_input,
 519         prog_len=args.rpl_prog_len,
 520         nb_runs=args.rpl_nb_runs,
 521         no_prog=args.rpl_no_prog,
 522         logger=log_string,
 523         device=device,
 524     )
 525
 526 elif args.task == "grid":
 527     task = tasks.Grid(
 528         nb_train_samples=args.nb_train_samples,
 529         nb_test_samples=args.nb_test_samples,
 530         batch_size=args.batch_size,
 531         size=args.grid_size,
 532         logger=log_string,
 533         device=device,
 534     )
 535
 536 elif args.task == "qmlp":
 537     task = tasks.QMLP(
 538         nb_train_samples=args.nb_train_samples,
 539         nb_test_samples=args.nb_test_samples,
 540         batch_size=args.batch_size,
 541         result_dir=args.result_dir,
 542         logger=log_string,
 543         device=device,
 544     )
 545
 546 else:
 547     raise ValueError(f"Unknown task {args.task}")
 548
 549 ######################################################################
 550
 551 log_string(f"device {device}")
 552
 553 vocabulary_size = task.vocabulary_size()
 554
 555 log_string(f"vocabulary_size {vocabulary_size}")
 556
 557 ##############################
 558
 559 model = mygpt.MyGPT(
 560     vocabulary_size=vocabulary_size,
 561     dim_model=args.dim_model,
 562     dim_keys=args.dim_keys,
 563     dim_hidden=args.dim_hidden,
 564     nb_heads=args.nb_heads,
 565     nb_blocks=args.nb_blocks,
 566     causal=True,
 567     dropout=args.dropout,
 568 )
 569
 570 model.to(device)
 571
 572 nb_parameters = sum(p.numel() for p in model.parameters())
 573 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 574
 575 ######################################################################
 576
 577 nb_epochs_finished = 0
 578
 579 if args.no_checkpoint:
 580     log_string(f"not trying to load checkpoint.")
 581
 582 else:
 583     try:
 584         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 585         checkpoint = torch.load(checkpoint_name)
 586         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 587         model.load_state_dict(checkpoint["model_state"])
 588         torch.set_rng_state(checkpoint["rng_state"])
 589         if torch.cuda.is_available():
 590             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 591
 592         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 593
 594     except FileNotFoundError:
 595         log_string("starting from scratch.")
 596
 597     except:
 598         log_string("error when loading the checkpoint.")
 599         exit(1)
 600
 601 ######################################################################
 602
 603 if args.task == "expr" and args.expr_input_file is not None:
 604     task.produce_results(
 605         n_epoch=nb_epochs_finished,
 606         model=model,
 607         result_dir=args.result_dir,
 608         logger=log_string,
 609         deterministic_synthesis=args.deterministic_synthesis,
 610         input_file=args.expr_input_file,
 611     )
 612
 613     exit(0)
 614
 615 ######################################################################
 616
 617 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 618
 619 # Compute the entropy of the training tokens
 620
 621 token_count = 0
 622 for input in task.batches(split="train"):
 623     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 624 token_probas = token_count / token_count.sum()
 625 entropy = -torch.xlogy(token_probas, token_probas).sum()
 626 train_set_perplexity = math.exp(entropy)
 627
 628 ######################################################################
 629 # A bit of paranoia never hurts
 630
 631 if args.max_percents_of_test_in_train >= 0:
 632
 633     def subsets_as_tuples(batches, cs):
 634         s = set()
 635         for batch in batches:
 636             for x in batch:
 637                 s.add(tuple([v.item() for v in x]))
 638                 if len(s) == cs:
 639                     yield s
 640                     s = set()
 641         yield s
 642
 643     nb_test, nb_in_train = 0, 0
 644     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 645         in_train = set()
 646         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 647             in_train.update(test_subset.intersection(train_subset))
 648         nb_in_train += len(in_train)
 649         nb_test += len(test_subset)
 650
 651     log_string(
 652         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 653     )
 654
 655     assert (
 656         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 657     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 658
 659 ##############################
 660
 661 if args.learning_rate_schedule == "cos":
 662     learning_rate_schedule = {}
 663     for n_epoch in range(args.nb_epochs):
 664         u = n_epoch / args.nb_epochs * math.pi
 665         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 666 else:
 667     u = {
 668         int(k): float(v)
 669         for k, v in [
 670             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 671         ]
 672     }
 673
 674     learning_rate_schedule = {}
 675     learning_rate = args.learning_rate
 676     for n_epoch in range(args.nb_epochs):
 677         if n_epoch in u:
 678             learning_rate = u[n_epoch]
 679         learning_rate_schedule[n_epoch] = learning_rate
 680
 681 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 682
 683 ##############################
 684
 685 nb_samples_seen = 0
 686
 687 if nb_epochs_finished >= nb_epochs:
 688     task.produce_results(
 689         n_epoch=nb_epochs_finished,
 690         model=model,
 691         result_dir=args.result_dir,
 692         logger=log_string,
 693         deterministic_synthesis=args.deterministic_synthesis,
 694     )
 695
 696 for n_epoch in range(nb_epochs_finished, nb_epochs):
 697     learning_rate = learning_rate_schedule[n_epoch]
 698
 699     log_string(f"learning_rate {learning_rate}")
 700
 701     if args.optim == "sgd":
 702         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 703     elif args.optim == "adam":
 704         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 705     elif args.optim == "adamw":
 706         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 707     else:
 708         raise ValueError(f"Unknown optimizer {args.optim}.")
 709
 710     model.train()
 711
 712     nb_train_samples, acc_train_loss = 0, 0.0
 713
 714     for input in task.batches(split="train"):
 715         input = input.to(device)
 716         output = model(mygpt.BracketedSequence(input)).x
 717         loss = F.cross_entropy(output.transpose(1, 2), input)
 718         acc_train_loss += loss.item() * input.size(0)
 719         nb_train_samples += input.size(0)
 720         nb_samples_seen += input.size(0)
 721
 722         optimizer.zero_grad()
 723         loss.backward()
 724         optimizer.step()
 725
 726     with torch.autograd.no_grad():
 727         model.eval()
 728
 729         nb_test_samples, acc_test_loss = 0, 0.0
 730
 731         for input in task.batches(split="test"):
 732             input = input.to(device)
 733
 734             output = model(mygpt.BracketedSequence(input)).x
 735             loss = F.cross_entropy(output.transpose(1, 2), input)
 736             acc_test_loss += loss.item() * input.size(0)
 737             nb_test_samples += input.size(0)
 738
 739         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 740         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 741
 742         log_string(
 743             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 744         )
 745
 746         task.produce_results(
 747             n_epoch=n_epoch,
 748             model=model,
 749             result_dir=args.result_dir,
 750             logger=log_string,
 751             deterministic_synthesis=args.deterministic_synthesis,
 752         )
 753
 754     checkpoint = {
 755         "nb_epochs_finished": n_epoch + 1,
 756         "model_state": model.state_dict(),
 757         "rng_state": torch.get_rng_state(),
 758     }
 759
 760     if torch.cuda.is_available():
 761         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 762
 763     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 764     torch.save(checkpoint, checkpoint_name)
 765     log_string(f"saved checkpoint {checkpoint_name}")
 766
 767 ######################################################################