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