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 # picoclvr options
 104
 105 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 106
 107 parser.add_argument("--picoclvr_height", type=int, default=12)
 108
 109 parser.add_argument("--picoclvr_width", type=int, default=16)
 110
 111 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 112
 113 ##############################
 114 # Maze options
 115
 116 parser.add_argument("--maze_height", type=int, default=13)
 117
 118 parser.add_argument("--maze_width", type=int, default=21)
 119
 120 parser.add_argument("--maze_nb_walls", type=int, default=15)
 121
 122 ##############################
 123 # Snake options
 124
 125 parser.add_argument("--snake_height", type=int, default=9)
 126
 127 parser.add_argument("--snake_width", type=int, default=12)
 128
 129 parser.add_argument("--snake_nb_colors", type=int, default=5)
 130
 131 parser.add_argument("--snake_length", type=int, default=200)
 132
 133 ##############################
 134 # Stack options
 135
 136 parser.add_argument("--stack_nb_steps", type=int, default=100)
 137
 138 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 139
 140 parser.add_argument("--stack_nb_digits", type=int, default=3)
 141
 142 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 143
 144 ##############################
 145 # Expr options
 146
 147 parser.add_argument("--expr_nb_variables", type=int, default=5)
 148
 149 parser.add_argument("--expr_sequence_length", type=int, default=40)
 150
 151 parser.add_argument("--expr_operand_max", type=int, default=9)
 152
 153 parser.add_argument("--expr_result_max", type=int, default=99)
 154
 155 parser.add_argument("--expr_input_file", type=str, default=None)
 156
 157 ##############################
 158 # World options
 159
 160 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 161
 162 ######################################################################
 163
 164 args = parser.parse_args()
 165
 166 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 167
 168 if args.result_dir is None:
 169     args.result_dir = f"results_{args.task}"
 170
 171 ######################################################################
 172
 173 default_task_args = {
 174     "byheart": {
 175         "model": "37M",
 176         "nb_epochs": 2,
 177         "batch_size": 25,
 178         "nb_train_samples": 50000,
 179         "nb_test_samples": 10000,
 180     },
 181     "learnop": {
 182         "model": "37M",
 183         "nb_epochs": 15,
 184         "batch_size": 25,
 185         "nb_train_samples": 50000,
 186         "nb_test_samples": 10000,
 187     },
 188     "guessop": {
 189         "model": "352M",
 190         "nb_epochs": 5,
 191         "batch_size": 25,
 192         "nb_train_samples": 1000000,
 193         "nb_test_samples": 10000,
 194     },
 195     "twotargets": {
 196         "model": "37M",
 197         "nb_epochs": 10,
 198         "batch_size": 25,
 199         "nb_train_samples": 50000,
 200         "nb_test_samples": 10000,
 201     },
 202     "addition": {
 203         "model": "352M",
 204         "nb_epochs": 50,
 205         "batch_size": 25,
 206         "nb_train_samples": 250000,
 207         "nb_test_samples": 10000,
 208     },
 209     "picoclvr": {
 210         "model": "37M",
 211         "nb_epochs": 25,
 212         "batch_size": 25,
 213         "nb_train_samples": 250000,
 214         "nb_test_samples": 10000,
 215     },
 216     "mnist": {
 217         "model": "37M",
 218         "nb_epochs": 25,
 219         "batch_size": 10,
 220         "nb_train_samples": 60000,
 221         "nb_test_samples": 10000,
 222     },
 223     "maze": {
 224         "model": "37M",
 225         "nb_epochs": 25,
 226         "batch_size": 5,
 227         "nb_train_samples": 100000,
 228         "nb_test_samples": 10000,
 229     },
 230     "snake": {
 231         "model": "37M",
 232         "nb_epochs": 5,
 233         "batch_size": 25,
 234         "nb_train_samples": 250000,
 235         "nb_test_samples": 10000,
 236     },
 237     "stack": {
 238         "model": "37M",
 239         "nb_epochs": 15,
 240         "batch_size": 25,
 241         "nb_train_samples": 100000,
 242         "nb_test_samples": 1000,
 243     },
 244     "expr": {
 245         "model": "352M",
 246         "nb_epochs": 25,
 247         "batch_size": 25,
 248         "nb_train_samples": 2500000,
 249         "nb_test_samples": 10000,
 250     },
 251     "rpl": {
 252         "model": "122M",
 253         "nb_epochs": 50,
 254         "batch_size": 5,
 255         "nb_train_samples": 1000000,
 256         "nb_test_samples": 10000,
 257     },
 258     "world": {
 259         "model": "37M",
 260         "nb_epochs": 10,
 261         "batch_size": 25,
 262         "nb_train_samples": 25000,
 263         "nb_test_samples": 1000,
 264     },
 265     "grid": {
 266         "model": "37M",
 267         "nb_epochs": 25,
 268         "batch_size": 25,
 269         "nb_train_samples": 250000,
 270         "nb_test_samples": 10000,
 271     },
 272 }
 273
 274 if args.task in default_task_args:
 275     for k, v in default_task_args[args.task].items():
 276         if getattr(args, k) is None:
 277             setattr(args, k, v)
 278
 279 ######################################################################
 280
 281 default_model_args = {
 282     "17K": {
 283         "dim_model": 32,
 284         "dim_keys": 32,
 285         "dim_hidden": 32,
 286         "nb_heads": 2,
 287         "nb_blocks": 2,
 288     },
 289     "37M": {
 290         "dim_model": 512,
 291         "dim_keys": 64,
 292         "dim_hidden": 2048,
 293         "nb_heads": 8,
 294         "nb_blocks": 12,
 295     },
 296     "122M": {
 297         "dim_model": 768,
 298         "dim_keys": 64,
 299         "dim_hidden": 2048,
 300         "nb_heads": 8,
 301         "nb_blocks": 24,
 302     },
 303     "352M": {
 304         "dim_model": 1024,
 305         "dim_keys": 64,
 306         "dim_hidden": 2048,
 307         "nb_heads": 8,
 308         "nb_blocks": 48,
 309     },
 310 }
 311
 312 if args.model in default_model_args:
 313     for k, v in default_model_args[args.model].items():
 314         if getattr(args, k) is None:
 315             setattr(args, k, v)
 316 else:
 317     raise ValueError(f"Unknown model {args.model}")
 318
 319 ######################################################################
 320
 321 try:
 322     os.mkdir(args.result_dir)
 323 except FileExistsError:
 324     if not args.overwrite_results:
 325         print(f"result directory {args.result_dir} already exists")
 326         exit(1)
 327
 328 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 329
 330 if args.seed >= 0:
 331     # torch.backends.cudnn.deterministic = True
 332     # torch.backends.cudnn.benchmark = False
 333     # torch.use_deterministic_algorithms(True)
 334     torch.manual_seed(args.seed)
 335     if torch.cuda.is_available():
 336         torch.cuda.manual_seed_all(args.seed)
 337
 338 ######################################################################
 339
 340
 341 def log_string(s):
 342     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 343
 344     if log_file is not None:
 345         log_file.write(t + s + "\n")
 346         log_file.flush()
 347
 348     print(t + s)
 349     sys.stdout.flush()
 350
 351
 352 for n in vars(args):
 353     log_string(f"args.{n} {getattr(args, n)}")
 354
 355
 356 ######################################################################
 357
 358
 359 def picoclvr_pruner_horizontal_green(p):
 360     return not ("green" in p and ("left" in p or "right" in p))
 361
 362
 363 picoclvr_pruner_train = (
 364     picoclvr_pruner_horizontal_green
 365     if args.picocvlr_prune_properties in {"train+eval"}
 366     else None
 367 )
 368
 369 picoclvr_pruner_eval = (
 370     (lambda p: not picoclvr_pruner_horizontal_green(p))
 371     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 372     else None
 373 )
 374
 375 ######################################################################
 376
 377 if args.task == "byheart":
 378     task = tasks.SandBox(
 379         problem=problems.ProblemByHeart(),
 380         nb_train_samples=args.nb_train_samples,
 381         nb_test_samples=args.nb_test_samples,
 382         batch_size=args.batch_size,
 383         logger=log_string,
 384         device=device,
 385     )
 386     args.max_percents_of_test_in_train = -1
 387
 388 elif args.task == "learnop":
 389     task = tasks.SandBox(
 390         problem=problems.ProblemLearnOperator(),
 391         nb_train_samples=args.nb_train_samples,
 392         nb_test_samples=args.nb_test_samples,
 393         batch_size=args.batch_size,
 394         logger=log_string,
 395         device=device,
 396     )
 397
 398
 399 elif args.task == "guessop":
 400     task = tasks.SandBox(
 401         problem=problems.ProblemGuessOperator(),
 402         nb_train_samples=args.nb_train_samples,
 403         nb_test_samples=args.nb_test_samples,
 404         batch_size=args.batch_size,
 405         logger=log_string,
 406         device=device,
 407     )
 408
 409
 410 elif args.task == "twotargets":
 411     task = tasks.SandBox(
 412         problem=problems.ProblemTwoTargets(),
 413         nb_train_samples=args.nb_train_samples,
 414         nb_test_samples=args.nb_test_samples,
 415         batch_size=args.batch_size,
 416         logger=log_string,
 417         device=device,
 418     )
 419
 420 elif args.task == "addition":
 421     task = tasks.SandBox(
 422         problem=problems.ProblemAddition(),
 423         nb_train_samples=args.nb_train_samples,
 424         nb_test_samples=args.nb_test_samples,
 425         batch_size=args.batch_size,
 426         logger=log_string,
 427         device=device,
 428     )
 429
 430 elif args.task == "picoclvr":
 431     task = tasks.PicoCLVR(
 432         nb_train_samples=args.nb_train_samples,
 433         nb_test_samples=args.nb_test_samples,
 434         batch_size=args.batch_size,
 435         height=args.picoclvr_height,
 436         width=args.picoclvr_width,
 437         nb_colors=args.picoclvr_nb_colors,
 438         logger=log_string,
 439         device=device,
 440         pruner_train=picoclvr_pruner_train,
 441         pruner_eval=picoclvr_pruner_eval,
 442     )
 443
 444 elif args.task == "mnist":
 445     task = tasks.MNIST(
 446         nb_train_samples=args.nb_train_samples,
 447         nb_test_samples=args.nb_test_samples,
 448         batch_size=args.batch_size,
 449         device=device,
 450     )
 451
 452 elif args.task == "maze":
 453     task = tasks.Maze(
 454         nb_train_samples=args.nb_train_samples,
 455         nb_test_samples=args.nb_test_samples,
 456         batch_size=args.batch_size,
 457         height=args.maze_height,
 458         width=args.maze_width,
 459         nb_walls=args.maze_nb_walls,
 460         device=device,
 461     )
 462
 463 elif args.task == "snake":
 464     task = tasks.Snake(
 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.snake_height,
 469         width=args.snake_width,
 470         nb_colors=args.snake_nb_colors,
 471         length=args.snake_length,
 472         prompt_length=args.snake_length // 2,
 473         device=device,
 474     )
 475
 476 elif args.task == "stack":
 477     task = tasks.Stack(
 478         nb_train_samples=args.nb_train_samples,
 479         nb_test_samples=args.nb_test_samples,
 480         batch_size=args.batch_size,
 481         logger=log_string,
 482         nb_steps=args.stack_nb_steps,
 483         nb_stacks=args.stack_nb_stacks,
 484         nb_digits=args.stack_nb_digits,
 485         fraction_values_for_train=args.stack_fraction_values_for_train,
 486         device=device,
 487     )
 488
 489 elif args.task == "expr":
 490     task = tasks.Expr(
 491         nb_train_samples=args.nb_train_samples,
 492         nb_test_samples=args.nb_test_samples,
 493         nb_variables=args.expr_nb_variables,
 494         sequence_length=args.expr_sequence_length,
 495         operand_max=args.expr_operand_max,
 496         result_max=args.expr_result_max,
 497         batch_size=args.batch_size,
 498         device=device,
 499     )
 500
 501 elif args.task == "rpl":
 502     task = tasks.RPL(
 503         nb_train_samples=args.nb_train_samples,
 504         nb_test_samples=args.nb_test_samples,
 505         batch_size=args.batch_size,
 506         nb_starting_values=args.rpl_nb_starting_values,
 507         max_input=args.rpl_max_input,
 508         prog_len=args.rpl_prog_len,
 509         nb_runs=args.rpl_nb_runs,
 510         no_prog=args.rpl_no_prog,
 511         logger=log_string,
 512         device=device,
 513     )
 514
 515 elif args.task == "grid":
 516     task = tasks.Grid(
 517         nb_train_samples=args.nb_train_samples,
 518         nb_test_samples=args.nb_test_samples,
 519         batch_size=args.batch_size,
 520         height=args.picoclvr_height,
 521         width=args.picoclvr_width,
 522         logger=log_string,
 523         device=device,
 524     )
 525
 526 elif args.task == "world":
 527     task = tasks.World(
 528         nb_train_samples=args.nb_train_samples,
 529         nb_test_samples=args.nb_test_samples,
 530         batch_size=args.batch_size,
 531         vqae_nb_epochs=args.world_vqae_nb_epochs,
 532         logger=log_string,
 533         device=device,
 534     )
 535
 536 else:
 537     raise ValueError(f"Unknown task {args.task}")
 538
 539 ######################################################################
 540
 541 log_string(f"device {device}")
 542
 543 vocabulary_size = task.vocabulary_size()
 544
 545 log_string(f"vocabulary_size {vocabulary_size}")
 546
 547 ##############################
 548
 549 model = mygpt.MyGPT(
 550     vocabulary_size=vocabulary_size,
 551     dim_model=args.dim_model,
 552     dim_keys=args.dim_keys,
 553     dim_hidden=args.dim_hidden,
 554     nb_heads=args.nb_heads,
 555     nb_blocks=args.nb_blocks,
 556     causal=True,
 557     dropout=args.dropout,
 558 )
 559
 560 model.to(device)
 561
 562 nb_parameters = sum(p.numel() for p in model.parameters())
 563 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 564
 565 ######################################################################
 566
 567 nb_epochs_finished = 0
 568
 569 if args.no_checkpoint:
 570     log_string(f"not trying to load checkpoint.")
 571
 572 else:
 573     try:
 574         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 575         checkpoint = torch.load(checkpoint_name)
 576         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 577         model.load_state_dict(checkpoint["model_state"])
 578         torch.set_rng_state(checkpoint["rng_state"])
 579         if torch.cuda.is_available():
 580             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 581
 582         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 583
 584     except FileNotFoundError:
 585         log_string("starting from scratch.")
 586
 587     except:
 588         log_string("error when loading the checkpoint.")
 589         exit(1)
 590
 591 ######################################################################
 592
 593 if args.task == "expr" and args.expr_input_file is not None:
 594     task.produce_results(
 595         n_epoch=nb_epochs_finished,
 596         model=model,
 597         result_dir=args.result_dir,
 598         logger=log_string,
 599         deterministic_synthesis=args.deterministic_synthesis,
 600         input_file=args.expr_input_file,
 601     )
 602
 603     exit(0)
 604
 605 ######################################################################
 606
 607 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 608
 609 # Compute the entropy of the training tokens
 610
 611 token_count = 0
 612 for input in task.batches(split="train"):
 613     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 614 token_probas = token_count / token_count.sum()
 615 entropy = -torch.xlogy(token_probas, token_probas).sum()
 616 train_set_perplexity = math.exp(entropy)
 617
 618 ######################################################################
 619 # A bit of paranoia never hurts
 620
 621 if args.max_percents_of_test_in_train >= 0:
 622
 623     def subsets_as_tuples(batches, cs):
 624         s = set()
 625         for batch in batches:
 626             for x in batch:
 627                 s.add(tuple([v.item() for v in x]))
 628                 if len(s) == cs:
 629                     yield s
 630                     s = set()
 631         yield s
 632
 633     nb_test, nb_in_train = 0, 0
 634     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 635         in_train = set()
 636         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 637             in_train.update(test_subset.intersection(train_subset))
 638         nb_in_train += len(in_train)
 639         nb_test += len(test_subset)
 640
 641     log_string(
 642         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 643     )
 644
 645     assert (
 646         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 647     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 648
 649 ##############################
 650
 651 if args.learning_rate_schedule == "cos":
 652     learning_rate_schedule = {}
 653     for n_epoch in range(args.nb_epochs):
 654         u = n_epoch / args.nb_epochs * math.pi
 655         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 656 else:
 657     u = {
 658         int(k): float(v)
 659         for k, v in [
 660             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 661         ]
 662     }
 663
 664     learning_rate_schedule = {}
 665     learning_rate = args.learning_rate
 666     for n_epoch in range(args.nb_epochs):
 667         if n_epoch in u:
 668             learning_rate = u[n_epoch]
 669         learning_rate_schedule[n_epoch] = learning_rate
 670
 671 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 672
 673 ##############################
 674
 675 nb_samples_seen = 0
 676
 677 if nb_epochs_finished >= nb_epochs:
 678     task.produce_results(
 679         n_epoch=nb_epochs_finished,
 680         model=model,
 681         result_dir=args.result_dir,
 682         logger=log_string,
 683         deterministic_synthesis=args.deterministic_synthesis,
 684     )
 685
 686 for n_epoch in range(nb_epochs_finished, nb_epochs):
 687     learning_rate = learning_rate_schedule[n_epoch]
 688
 689     log_string(f"learning_rate {learning_rate}")
 690
 691     if args.optim == "sgd":
 692         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 693     elif args.optim == "adam":
 694         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 695     elif args.optim == "adamw":
 696         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 697     else:
 698         raise ValueError(f"Unknown optimizer {args.optim}.")
 699
 700     model.train()
 701
 702     nb_train_samples, acc_train_loss = 0, 0.0
 703
 704     for input in task.batches(split="train"):
 705         input = input.to(device)
 706         output = model(mygpt.BracketedSequence(input)).x
 707         loss = F.cross_entropy(output.transpose(1, 2), input)
 708         acc_train_loss += loss.item() * input.size(0)
 709         nb_train_samples += input.size(0)
 710         nb_samples_seen += input.size(0)
 711
 712         optimizer.zero_grad()
 713         loss.backward()
 714         optimizer.step()
 715
 716     with torch.autograd.no_grad():
 717         model.eval()
 718
 719         nb_test_samples, acc_test_loss = 0, 0.0
 720
 721         for input in task.batches(split="test"):
 722             input = input.to(device)
 723
 724             output = model(mygpt.BracketedSequence(input)).x
 725             loss = F.cross_entropy(output.transpose(1, 2), input)
 726             acc_test_loss += loss.item() * input.size(0)
 727             nb_test_samples += input.size(0)
 728
 729         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 730         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 731
 732         log_string(
 733             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 734         )
 735
 736         task.produce_results(
 737             n_epoch=n_epoch,
 738             model=model,
 739             result_dir=args.result_dir,
 740             logger=log_string,
 741             deterministic_synthesis=args.deterministic_synthesis,
 742         )
 743
 744     checkpoint = {
 745         "nb_epochs_finished": n_epoch + 1,
 746         "model_state": model.state_dict(),
 747         "rng_state": torch.get_rng_state(),
 748     }
 749
 750     if torch.cuda.is_available():
 751         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 752
 753     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 754     torch.save(checkpoint, checkpoint_name)
 755     log_string(f"saved checkpoint {checkpoint_name}")
 756
 757 ######################################################################