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, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid",
  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=None)
  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 # World options
 164
 165 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 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     "byheart": {
 180         "model": "37M",
 181         "nb_epochs": 2,
 182         "batch_size": 25,
 183         "nb_train_samples": 50000,
 184         "nb_test_samples": 10000,
 185     },
 186     "learnop": {
 187         "model": "37M",
 188         "nb_epochs": 15,
 189         "batch_size": 25,
 190         "nb_train_samples": 50000,
 191         "nb_test_samples": 10000,
 192     },
 193     "guessop": {
 194         "model": "352M",
 195         "nb_epochs": 5,
 196         "batch_size": 25,
 197         "nb_train_samples": 1000000,
 198         "nb_test_samples": 10000,
 199     },
 200     "twotargets": {
 201         "model": "37M",
 202         "nb_epochs": 10,
 203         "batch_size": 25,
 204         "nb_train_samples": 50000,
 205         "nb_test_samples": 10000,
 206     },
 207     "addition": {
 208         "model": "352M",
 209         "nb_epochs": 50,
 210         "batch_size": 25,
 211         "nb_train_samples": 250000,
 212         "nb_test_samples": 10000,
 213     },
 214     "picoclvr": {
 215         "model": "37M",
 216         "nb_epochs": 25,
 217         "batch_size": 25,
 218         "nb_train_samples": 250000,
 219         "nb_test_samples": 10000,
 220     },
 221     "mnist": {
 222         "model": "37M",
 223         "nb_epochs": 25,
 224         "batch_size": 10,
 225         "nb_train_samples": 60000,
 226         "nb_test_samples": 10000,
 227     },
 228     "maze": {
 229         "model": "37M",
 230         "nb_epochs": 25,
 231         "batch_size": 5,
 232         "nb_train_samples": 100000,
 233         "nb_test_samples": 10000,
 234     },
 235     "snake": {
 236         "model": "37M",
 237         "nb_epochs": 5,
 238         "batch_size": 25,
 239         "nb_train_samples": 250000,
 240         "nb_test_samples": 10000,
 241     },
 242     "stack": {
 243         "model": "37M",
 244         "nb_epochs": 15,
 245         "batch_size": 25,
 246         "nb_train_samples": 100000,
 247         "nb_test_samples": 1000,
 248     },
 249     "expr": {
 250         "model": "352M",
 251         "nb_epochs": 25,
 252         "batch_size": 25,
 253         "nb_train_samples": 2500000,
 254         "nb_test_samples": 10000,
 255     },
 256     "rpl": {
 257         "model": "122M",
 258         "nb_epochs": 50,
 259         "batch_size": 5,
 260         "nb_train_samples": 1000000,
 261         "nb_test_samples": 10000,
 262     },
 263     "world": {
 264         "model": "37M",
 265         "nb_epochs": 10,
 266         "batch_size": 25,
 267         "nb_train_samples": 25000,
 268         "nb_test_samples": 1000,
 269     },
 270     "grid": {
 271         "model": "37M",
 272         "nb_epochs": 25,
 273         "batch_size": 25,
 274         "nb_train_samples": 250000,
 275         "nb_test_samples": 10000,
 276     },
 277 }
 278
 279 if args.task in default_task_args:
 280     for k, v in default_task_args[args.task].items():
 281         if getattr(args, k) is None:
 282             setattr(args, k, v)
 283
 284 ######################################################################
 285
 286 default_model_args = {
 287     "17K": {
 288         "dim_model": 32,
 289         "dim_keys": 32,
 290         "dim_hidden": 32,
 291         "nb_heads": 2,
 292         "nb_blocks": 2,
 293     },
 294     "37M": {
 295         "dim_model": 512,
 296         "dim_keys": 64,
 297         "dim_hidden": 2048,
 298         "nb_heads": 8,
 299         "nb_blocks": 12,
 300     },
 301     "122M": {
 302         "dim_model": 768,
 303         "dim_keys": 64,
 304         "dim_hidden": 2048,
 305         "nb_heads": 8,
 306         "nb_blocks": 24,
 307     },
 308     "352M": {
 309         "dim_model": 1024,
 310         "dim_keys": 64,
 311         "dim_hidden": 2048,
 312         "nb_heads": 8,
 313         "nb_blocks": 48,
 314     },
 315 }
 316
 317 if args.model in default_model_args:
 318     for k, v in default_model_args[args.model].items():
 319         if getattr(args, k) is None:
 320             setattr(args, k, v)
 321 else:
 322     raise ValueError(f"Unknown model {args.model}")
 323
 324 ######################################################################
 325
 326 try:
 327     os.mkdir(args.result_dir)
 328 except FileExistsError:
 329     if not args.overwrite_results:
 330         print(f"result directory {args.result_dir} already exists")
 331         exit(1)
 332
 333 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 334
 335 if args.seed >= 0:
 336     # torch.backends.cudnn.deterministic = True
 337     # torch.backends.cudnn.benchmark = False
 338     # torch.use_deterministic_algorithms(True)
 339     torch.manual_seed(args.seed)
 340     if torch.cuda.is_available():
 341         torch.cuda.manual_seed_all(args.seed)
 342
 343 ######################################################################
 344
 345
 346 def log_string(s):
 347     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 348
 349     if log_file is not None:
 350         log_file.write(t + s + "\n")
 351         log_file.flush()
 352
 353     print(t + s)
 354     sys.stdout.flush()
 355
 356
 357 for n in vars(args):
 358     log_string(f"args.{n} {getattr(args, n)}")
 359
 360
 361 ######################################################################
 362
 363
 364 def picoclvr_pruner_horizontal_green(p):
 365     return not ("green" in p and ("left" in p or "right" in p))
 366
 367
 368 picoclvr_pruner_train = (
 369     picoclvr_pruner_horizontal_green
 370     if args.picocvlr_prune_properties in {"train+eval"}
 371     else None
 372 )
 373
 374 picoclvr_pruner_eval = (
 375     (lambda p: not picoclvr_pruner_horizontal_green(p))
 376     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 377     else None
 378 )
 379
 380 ######################################################################
 381
 382 if args.task == "byheart":
 383     task = tasks.SandBox(
 384         problem=problems.ProblemByHeart(),
 385         nb_train_samples=args.nb_train_samples,
 386         nb_test_samples=args.nb_test_samples,
 387         batch_size=args.batch_size,
 388         logger=log_string,
 389         device=device,
 390     )
 391     args.max_percents_of_test_in_train = -1
 392
 393 elif args.task == "learnop":
 394     task = tasks.SandBox(
 395         problem=problems.ProblemLearnOperator(),
 396         nb_train_samples=args.nb_train_samples,
 397         nb_test_samples=args.nb_test_samples,
 398         batch_size=args.batch_size,
 399         logger=log_string,
 400         device=device,
 401     )
 402
 403
 404 elif args.task == "guessop":
 405     task = tasks.SandBox(
 406         problem=problems.ProblemGuessOperator(),
 407         nb_train_samples=args.nb_train_samples,
 408         nb_test_samples=args.nb_test_samples,
 409         batch_size=args.batch_size,
 410         logger=log_string,
 411         device=device,
 412     )
 413
 414
 415 elif args.task == "twotargets":
 416     task = tasks.SandBox(
 417         problem=problems.ProblemTwoTargets(),
 418         nb_train_samples=args.nb_train_samples,
 419         nb_test_samples=args.nb_test_samples,
 420         batch_size=args.batch_size,
 421         logger=log_string,
 422         device=device,
 423     )
 424
 425 elif args.task == "addition":
 426     task = tasks.SandBox(
 427         problem=problems.ProblemAddition(),
 428         nb_train_samples=args.nb_train_samples,
 429         nb_test_samples=args.nb_test_samples,
 430         batch_size=args.batch_size,
 431         logger=log_string,
 432         device=device,
 433     )
 434
 435 elif args.task == "picoclvr":
 436     task = tasks.PicoCLVR(
 437         nb_train_samples=args.nb_train_samples,
 438         nb_test_samples=args.nb_test_samples,
 439         batch_size=args.batch_size,
 440         height=args.picoclvr_height,
 441         width=args.picoclvr_width,
 442         nb_colors=args.picoclvr_nb_colors,
 443         logger=log_string,
 444         device=device,
 445         pruner_train=picoclvr_pruner_train,
 446         pruner_eval=picoclvr_pruner_eval,
 447     )
 448
 449 elif args.task == "mnist":
 450     task = tasks.MNIST(
 451         nb_train_samples=args.nb_train_samples,
 452         nb_test_samples=args.nb_test_samples,
 453         batch_size=args.batch_size,
 454         device=device,
 455     )
 456
 457 elif args.task == "maze":
 458     task = tasks.Maze(
 459         nb_train_samples=args.nb_train_samples,
 460         nb_test_samples=args.nb_test_samples,
 461         batch_size=args.batch_size,
 462         height=args.maze_height,
 463         width=args.maze_width,
 464         nb_walls=args.maze_nb_walls,
 465         device=device,
 466     )
 467
 468 elif args.task == "snake":
 469     task = tasks.Snake(
 470         nb_train_samples=args.nb_train_samples,
 471         nb_test_samples=args.nb_test_samples,
 472         batch_size=args.batch_size,
 473         height=args.snake_height,
 474         width=args.snake_width,
 475         nb_colors=args.snake_nb_colors,
 476         length=args.snake_length,
 477         prompt_length=args.snake_length // 2,
 478         device=device,
 479     )
 480
 481 elif args.task == "stack":
 482     task = tasks.Stack(
 483         nb_train_samples=args.nb_train_samples,
 484         nb_test_samples=args.nb_test_samples,
 485         batch_size=args.batch_size,
 486         logger=log_string,
 487         nb_steps=args.stack_nb_steps,
 488         nb_stacks=args.stack_nb_stacks,
 489         nb_digits=args.stack_nb_digits,
 490         fraction_values_for_train=args.stack_fraction_values_for_train,
 491         device=device,
 492     )
 493
 494 elif args.task == "expr":
 495     task = tasks.Expr(
 496         nb_train_samples=args.nb_train_samples,
 497         nb_test_samples=args.nb_test_samples,
 498         nb_variables=args.expr_nb_variables,
 499         sequence_length=args.expr_sequence_length,
 500         operand_max=args.expr_operand_max,
 501         result_max=args.expr_result_max,
 502         batch_size=args.batch_size,
 503         device=device,
 504     )
 505
 506 elif args.task == "rpl":
 507     task = tasks.RPL(
 508         nb_train_samples=args.nb_train_samples,
 509         nb_test_samples=args.nb_test_samples,
 510         batch_size=args.batch_size,
 511         nb_starting_values=args.rpl_nb_starting_values,
 512         max_input=args.rpl_max_input,
 513         prog_len=args.rpl_prog_len,
 514         nb_runs=args.rpl_nb_runs,
 515         no_prog=args.rpl_no_prog,
 516         logger=log_string,
 517         device=device,
 518     )
 519
 520 elif args.task == "grid":
 521     task = tasks.Grid(
 522         nb_train_samples=args.nb_train_samples,
 523         nb_test_samples=args.nb_test_samples,
 524         batch_size=args.batch_size,
 525         size=args.grid_size,
 526         logger=log_string,
 527         device=device,
 528     )
 529
 530 elif args.task == "world":
 531     task = tasks.World(
 532         nb_train_samples=args.nb_train_samples,
 533         nb_test_samples=args.nb_test_samples,
 534         batch_size=args.batch_size,
 535         vqae_nb_epochs=args.world_vqae_nb_epochs,
 536         logger=log_string,
 537         device=device,
 538     )
 539
 540 else:
 541     raise ValueError(f"Unknown task {args.task}")
 542
 543 ######################################################################
 544
 545 log_string(f"device {device}")
 546
 547 vocabulary_size = task.vocabulary_size()
 548
 549 log_string(f"vocabulary_size {vocabulary_size}")
 550
 551 ##############################
 552
 553 model = mygpt.MyGPT(
 554     vocabulary_size=vocabulary_size,
 555     dim_model=args.dim_model,
 556     dim_keys=args.dim_keys,
 557     dim_hidden=args.dim_hidden,
 558     nb_heads=args.nb_heads,
 559     nb_blocks=args.nb_blocks,
 560     causal=True,
 561     dropout=args.dropout,
 562 )
 563
 564 model.to(device)
 565
 566 nb_parameters = sum(p.numel() for p in model.parameters())
 567 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 568
 569 ######################################################################
 570
 571 nb_epochs_finished = 0
 572
 573 if args.no_checkpoint:
 574     log_string(f"not trying to load checkpoint.")
 575
 576 else:
 577     try:
 578         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 579         checkpoint = torch.load(checkpoint_name)
 580         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 581         model.load_state_dict(checkpoint["model_state"])
 582         torch.set_rng_state(checkpoint["rng_state"])
 583         if torch.cuda.is_available():
 584             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 585
 586         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 587
 588     except FileNotFoundError:
 589         log_string("starting from scratch.")
 590
 591     except:
 592         log_string("error when loading the checkpoint.")
 593         exit(1)
 594
 595 ######################################################################
 596
 597 if args.task == "expr" and args.expr_input_file is not None:
 598     task.produce_results(
 599         n_epoch=nb_epochs_finished,
 600         model=model,
 601         result_dir=args.result_dir,
 602         logger=log_string,
 603         deterministic_synthesis=args.deterministic_synthesis,
 604         input_file=args.expr_input_file,
 605     )
 606
 607     exit(0)
 608
 609 ######################################################################
 610
 611 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 612
 613 # Compute the entropy of the training tokens
 614
 615 token_count = 0
 616 for input in task.batches(split="train"):
 617     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 618 token_probas = token_count / token_count.sum()
 619 entropy = -torch.xlogy(token_probas, token_probas).sum()
 620 train_set_perplexity = math.exp(entropy)
 621
 622 ######################################################################
 623 # A bit of paranoia never hurts
 624
 625 if args.max_percents_of_test_in_train >= 0:
 626
 627     def subsets_as_tuples(batches, cs):
 628         s = set()
 629         for batch in batches:
 630             for x in batch:
 631                 s.add(tuple([v.item() for v in x]))
 632                 if len(s) == cs:
 633                     yield s
 634                     s = set()
 635         yield s
 636
 637     nb_test, nb_in_train = 0, 0
 638     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 639         in_train = set()
 640         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 641             in_train.update(test_subset.intersection(train_subset))
 642         nb_in_train += len(in_train)
 643         nb_test += len(test_subset)
 644
 645     log_string(
 646         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 647     )
 648
 649     assert (
 650         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 651     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 652
 653 ##############################
 654
 655 if args.learning_rate_schedule == "cos":
 656     learning_rate_schedule = {}
 657     for n_epoch in range(args.nb_epochs):
 658         u = n_epoch / args.nb_epochs * math.pi
 659         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 660 else:
 661     u = {
 662         int(k): float(v)
 663         for k, v in [
 664             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 665         ]
 666     }
 667
 668     learning_rate_schedule = {}
 669     learning_rate = args.learning_rate
 670     for n_epoch in range(args.nb_epochs):
 671         if n_epoch in u:
 672             learning_rate = u[n_epoch]
 673         learning_rate_schedule[n_epoch] = learning_rate
 674
 675 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 676
 677 ##############################
 678
 679 nb_samples_seen = 0
 680
 681 if nb_epochs_finished >= nb_epochs:
 682     task.produce_results(
 683         n_epoch=nb_epochs_finished,
 684         model=model,
 685         result_dir=args.result_dir,
 686         logger=log_string,
 687         deterministic_synthesis=args.deterministic_synthesis,
 688     )
 689
 690 for n_epoch in range(nb_epochs_finished, nb_epochs):
 691     learning_rate = learning_rate_schedule[n_epoch]
 692
 693     log_string(f"learning_rate {learning_rate}")
 694
 695     if args.optim == "sgd":
 696         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 697     elif args.optim == "adam":
 698         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 699     elif args.optim == "adamw":
 700         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 701     else:
 702         raise ValueError(f"Unknown optimizer {args.optim}.")
 703
 704     model.train()
 705
 706     nb_train_samples, acc_train_loss = 0, 0.0
 707
 708     for input in task.batches(split="train"):
 709         input = input.to(device)
 710         output = model(mygpt.BracketedSequence(input)).x
 711         loss = F.cross_entropy(output.transpose(1, 2), input)
 712         acc_train_loss += loss.item() * input.size(0)
 713         nb_train_samples += input.size(0)
 714         nb_samples_seen += input.size(0)
 715
 716         optimizer.zero_grad()
 717         loss.backward()
 718         optimizer.step()
 719
 720     with torch.autograd.no_grad():
 721         model.eval()
 722
 723         nb_test_samples, acc_test_loss = 0, 0.0
 724
 725         for input in task.batches(split="test"):
 726             input = input.to(device)
 727
 728             output = model(mygpt.BracketedSequence(input)).x
 729             loss = F.cross_entropy(output.transpose(1, 2), input)
 730             acc_test_loss += loss.item() * input.size(0)
 731             nb_test_samples += input.size(0)
 732
 733         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 734         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 735
 736         log_string(
 737             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 738         )
 739
 740         task.produce_results(
 741             n_epoch=n_epoch,
 742             model=model,
 743             result_dir=args.result_dir,
 744             logger=log_string,
 745             deterministic_synthesis=args.deterministic_synthesis,
 746         )
 747
 748     checkpoint = {
 749         "nb_epochs_finished": n_epoch + 1,
 750         "model_state": model.state_dict(),
 751         "rng_state": torch.get_rng_state(),
 752     }
 753
 754     if torch.cuda.is_available():
 755         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 756
 757     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 758     torch.save(checkpoint, checkpoint_name)
 759     log_string(f"saved checkpoint {checkpoint_name}")
 760
 761 ######################################################################