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