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, datetime
   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="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, escape",
  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 # filetask
  91
  92 parser.add_argument("--filetask_train_file", type=str, default=None)
  93
  94 parser.add_argument("--filetask_test_file", type=str, default=None)
  95
  96 ##############################
  97 # rpl options
  98
  99 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 100
 101 parser.add_argument("--rpl_max_input", type=int, default=9)
 102
 103 parser.add_argument("--rpl_prog_len", type=int, default=8)
 104
 105 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 106
 107 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 108
 109 ##############################
 110 # grid options
 111
 112 parser.add_argument("--grid_size", type=int, default=6)
 113
 114 parser.add_argument("--grid_fraction_play", type=float, default=0)
 115
 116 ##############################
 117 # picoclvr options
 118
 119 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 120
 121 parser.add_argument("--picoclvr_height", type=int, default=12)
 122
 123 parser.add_argument("--picoclvr_width", type=int, default=16)
 124
 125 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 126
 127 ##############################
 128 # Maze options
 129
 130 parser.add_argument("--maze_height", type=int, default=13)
 131
 132 parser.add_argument("--maze_width", type=int, default=21)
 133
 134 parser.add_argument("--maze_nb_walls", type=int, default=15)
 135
 136 ##############################
 137 # Snake options
 138
 139 parser.add_argument("--snake_height", type=int, default=9)
 140
 141 parser.add_argument("--snake_width", type=int, default=12)
 142
 143 parser.add_argument("--snake_nb_colors", type=int, default=5)
 144
 145 parser.add_argument("--snake_length", type=int, default=200)
 146
 147 ##############################
 148 # Stack options
 149
 150 parser.add_argument("--stack_nb_steps", type=int, default=100)
 151
 152 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 153
 154 parser.add_argument("--stack_nb_digits", type=int, default=3)
 155
 156 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 157
 158 ##############################
 159 # Expr options
 160
 161 parser.add_argument("--expr_nb_variables", type=int, default=5)
 162
 163 parser.add_argument("--expr_sequence_length", type=int, default=40)
 164
 165 parser.add_argument("--expr_operand_max", type=int, default=9)
 166
 167 parser.add_argument("--expr_result_max", type=int, default=99)
 168
 169 parser.add_argument("--expr_input_file", type=str, default=None)
 170
 171 ##############################
 172 # Mixing
 173
 174 parser.add_argument("--mixing_hard", action="store_true", default=False)
 175
 176 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 177
 178 ##############################
 179 # escape options
 180
 181 parser.add_argument("--escape_height", type=int, default=5)
 182
 183 parser.add_argument("--escape_width", type=int, default=7)
 184
 185 parser.add_argument("--escape_T", type=int, default=25)
 186
 187 parser.add_argument("--escape_nb_walls", type=int, default=5)
 188
 189 ######################################################################
 190
 191 args = parser.parse_args()
 192
 193 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 194
 195 if args.result_dir is None:
 196     args.result_dir = f"results_{args.task}"
 197
 198 ######################################################################
 199
 200 default_task_args = {
 201     "file": {
 202         "model": "37M",
 203         "batch_size": 25,
 204         "nb_train_samples": 250000,
 205         "nb_test_samples": 10000,
 206     },
 207     "addition": {
 208         "model": "352M",
 209         "batch_size": 25,
 210         "nb_train_samples": 250000,
 211         "nb_test_samples": 10000,
 212     },
 213     "byheart": {
 214         "model": "37M",
 215         "batch_size": 25,
 216         "nb_train_samples": 50000,
 217         "nb_test_samples": 10000,
 218     },
 219     "expr": {
 220         "model": "352M",
 221         "batch_size": 25,
 222         "nb_train_samples": 2500000,
 223         "nb_test_samples": 10000,
 224     },
 225     "grid": {
 226         "model": "37M",
 227         "batch_size": 25,
 228         "nb_train_samples": 250000,
 229         "nb_test_samples": 10000,
 230     },
 231     "qmlp": {
 232         "model": "37M",
 233         "batch_size": 10,
 234         "nb_train_samples": 100000,
 235         "nb_test_samples": 1000,
 236     },
 237     "guessop": {
 238         "model": "352M",
 239         "batch_size": 25,
 240         "nb_train_samples": 1000000,
 241         "nb_test_samples": 10000,
 242     },
 243     "learnop": {
 244         "model": "37M",
 245         "batch_size": 25,
 246         "nb_train_samples": 50000,
 247         "nb_test_samples": 10000,
 248     },
 249     "maze": {
 250         "model": "37M",
 251         "batch_size": 5,
 252         "nb_train_samples": 100000,
 253         "nb_test_samples": 10000,
 254     },
 255     "picoclvr": {
 256         "model": "37M",
 257         "batch_size": 25,
 258         "nb_train_samples": 250000,
 259         "nb_test_samples": 10000,
 260     },
 261     "rpl": {
 262         "model": "352M",
 263         "batch_size": 5,
 264         "nb_train_samples": 2500000,
 265         "nb_test_samples": 10000,
 266     },
 267     "snake": {
 268         "model": "37M",
 269         "batch_size": 25,
 270         "nb_train_samples": 250000,
 271         "nb_test_samples": 10000,
 272     },
 273     "stack": {
 274         "model": "37M",
 275         "batch_size": 25,
 276         "nb_train_samples": 100000,
 277         "nb_test_samples": 1000,
 278     },
 279     "twotargets": {
 280         "model": "37M",
 281         "batch_size": 25,
 282         "nb_train_samples": 50000,
 283         "nb_test_samples": 10000,
 284     },
 285     "memory": {
 286         "model": "37M",
 287         "batch_size": 100,
 288         "nb_train_samples": 25000,
 289         "nb_test_samples": 1000,
 290     },
 291     "mixing": {
 292         "model": "37M",
 293         "batch_size": 25,
 294         "nb_train_samples": 250000,
 295         "nb_test_samples": 10000,
 296     },
 297     "mnist": {
 298         "model": "37M",
 299         "batch_size": 10,
 300         "nb_train_samples": 60000,
 301         "nb_test_samples": 10000,
 302     },
 303     "escape": {
 304         "model": "37M",
 305         "batch_size": 25,
 306         "nb_train_samples": 25000,
 307         "nb_test_samples": 10000,
 308     },
 309 }
 310
 311 if args.task in default_task_args:
 312     for k, v in default_task_args[args.task].items():
 313         if getattr(args, k) is None:
 314             setattr(args, k, v)
 315
 316 ######################################################################
 317
 318 default_model_args = {
 319     "17K": {
 320         "dim_model": 32,
 321         "dim_keys": 32,
 322         "dim_hidden": 32,
 323         "nb_heads": 2,
 324         "nb_blocks": 2,
 325     },
 326     "4M": {
 327         "dim_model": 256,
 328         "dim_keys": 32,
 329         "dim_hidden": 1024,
 330         "nb_heads": 4,
 331         "nb_blocks": 6,
 332     },
 333     "37M": {
 334         "dim_model": 512,
 335         "dim_keys": 64,
 336         "dim_hidden": 2048,
 337         "nb_heads": 8,
 338         "nb_blocks": 12,
 339     },
 340     "122M": {
 341         "dim_model": 768,
 342         "dim_keys": 64,
 343         "dim_hidden": 2048,
 344         "nb_heads": 8,
 345         "nb_blocks": 24,
 346     },
 347     "352M": {
 348         "dim_model": 1024,
 349         "dim_keys": 64,
 350         "dim_hidden": 2048,
 351         "nb_heads": 8,
 352         "nb_blocks": 48,
 353     },
 354 }
 355
 356 if args.model in default_model_args:
 357     for k, v in default_model_args[args.model].items():
 358         if getattr(args, k) is None:
 359             setattr(args, k, v)
 360 else:
 361     raise ValueError(f"Unknown model {args.model}")
 362
 363 ######################################################################
 364
 365 try:
 366     os.mkdir(args.result_dir)
 367 except FileExistsError:
 368     if not args.overwrite_results:
 369         print(f"result directory {args.result_dir} already exists")
 370         exit(1)
 371
 372 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 373
 374 if args.seed >= 0:
 375     # torch.backends.cudnn.deterministic = True
 376     # torch.backends.cudnn.benchmark = False
 377     # torch.use_deterministic_algorithms(True)
 378     torch.manual_seed(args.seed)
 379     if torch.cuda.is_available():
 380         torch.cuda.manual_seed_all(args.seed)
 381
 382 ######################################################################
 383
 384
 385 def log_string(s):
 386     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 387
 388     if log_file is not None:
 389         log_file.write(t + s + "\n")
 390         log_file.flush()
 391
 392     print(t + s)
 393     sys.stdout.flush()
 394
 395
 396 log_string(f"argv {' '.join(sys.argv)}")
 397
 398 for n in vars(args):
 399     log_string(f"args.{n} {getattr(args, n)}")
 400
 401
 402 ######################################################################
 403
 404
 405 def picoclvr_pruner_horizontal_green(p):
 406     return not ("green" in p and ("left" in p or "right" in p))
 407
 408
 409 picoclvr_pruner_train = (
 410     picoclvr_pruner_horizontal_green
 411     if args.picocvlr_prune_properties in {"train+eval"}
 412     else None
 413 )
 414
 415 picoclvr_pruner_eval = (
 416     (lambda p: not picoclvr_pruner_horizontal_green(p))
 417     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 418     else None
 419 )
 420
 421 ######################################################################
 422
 423 if args.task == "file":
 424     assert (
 425         args.filetask_train_file is not None and args.filetask_test_file is not None
 426     ), "You have to specify the task train and test files"
 427     task = tasks.TaskFromFile(
 428         args.filetask_train_file,
 429         args.filetask_test_file,
 430         nb_train_samples=args.nb_train_samples,
 431         nb_test_samples=args.nb_test_samples,
 432         batch_size=args.batch_size,
 433         shuffle=True,
 434         device=device,
 435     )
 436     args.max_percents_of_test_in_train = 0
 437
 438 elif args.task == "byheart":
 439     task = tasks.SandBox(
 440         problem=problems.ProblemByHeart(),
 441         nb_train_samples=args.nb_train_samples,
 442         nb_test_samples=args.nb_test_samples,
 443         batch_size=args.batch_size,
 444         logger=log_string,
 445         device=device,
 446     )
 447     args.max_percents_of_test_in_train = -1
 448
 449 elif args.task == "learnop":
 450     task = tasks.SandBox(
 451         problem=problems.ProblemLearnOperator(),
 452         nb_train_samples=args.nb_train_samples,
 453         nb_test_samples=args.nb_test_samples,
 454         batch_size=args.batch_size,
 455         logger=log_string,
 456         device=device,
 457     )
 458
 459
 460 elif args.task == "guessop":
 461     task = tasks.SandBox(
 462         problem=problems.ProblemGuessOperator(),
 463         nb_train_samples=args.nb_train_samples,
 464         nb_test_samples=args.nb_test_samples,
 465         batch_size=args.batch_size,
 466         logger=log_string,
 467         device=device,
 468     )
 469
 470
 471 elif args.task == "twotargets":
 472     task = tasks.SandBox(
 473         problem=problems.ProblemTwoTargets(),
 474         nb_train_samples=args.nb_train_samples,
 475         nb_test_samples=args.nb_test_samples,
 476         batch_size=args.batch_size,
 477         logger=log_string,
 478         device=device,
 479     )
 480
 481 elif args.task == "memory":
 482     task = tasks.SandBox(
 483         problem=problems.ProblemMemory(),
 484         nb_train_samples=args.nb_train_samples,
 485         nb_test_samples=args.nb_test_samples,
 486         batch_size=args.batch_size,
 487         logger=log_string,
 488         device=device,
 489     )
 490
 491 elif args.task == "mixing":
 492     task = tasks.SandBox(
 493         problem=problems.ProblemMixing(
 494             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 495         ),
 496         nb_train_samples=args.nb_train_samples,
 497         nb_test_samples=args.nb_test_samples,
 498         batch_size=args.batch_size,
 499         logger=log_string,
 500         device=device,
 501     )
 502
 503 elif args.task == "addition":
 504     task = tasks.SandBox(
 505         problem=problems.ProblemAddition(),
 506         nb_train_samples=args.nb_train_samples,
 507         nb_test_samples=args.nb_test_samples,
 508         batch_size=args.batch_size,
 509         logger=log_string,
 510         device=device,
 511     )
 512
 513 elif args.task == "picoclvr":
 514     task = tasks.PicoCLVR(
 515         nb_train_samples=args.nb_train_samples,
 516         nb_test_samples=args.nb_test_samples,
 517         batch_size=args.batch_size,
 518         height=args.picoclvr_height,
 519         width=args.picoclvr_width,
 520         nb_colors=args.picoclvr_nb_colors,
 521         logger=log_string,
 522         device=device,
 523         pruner_train=picoclvr_pruner_train,
 524         pruner_eval=picoclvr_pruner_eval,
 525     )
 526
 527 elif args.task == "mnist":
 528     task = tasks.MNIST(
 529         nb_train_samples=args.nb_train_samples,
 530         nb_test_samples=args.nb_test_samples,
 531         batch_size=args.batch_size,
 532         device=device,
 533     )
 534
 535 elif args.task == "maze":
 536     task = tasks.Maze(
 537         nb_train_samples=args.nb_train_samples,
 538         nb_test_samples=args.nb_test_samples,
 539         batch_size=args.batch_size,
 540         height=args.maze_height,
 541         width=args.maze_width,
 542         nb_walls=args.maze_nb_walls,
 543         device=device,
 544     )
 545
 546 elif args.task == "snake":
 547     task = tasks.Snake(
 548         nb_train_samples=args.nb_train_samples,
 549         nb_test_samples=args.nb_test_samples,
 550         batch_size=args.batch_size,
 551         height=args.snake_height,
 552         width=args.snake_width,
 553         nb_colors=args.snake_nb_colors,
 554         length=args.snake_length,
 555         prompt_length=args.snake_length // 2,
 556         device=device,
 557     )
 558
 559 elif args.task == "stack":
 560     task = tasks.Stack(
 561         nb_train_samples=args.nb_train_samples,
 562         nb_test_samples=args.nb_test_samples,
 563         batch_size=args.batch_size,
 564         logger=log_string,
 565         nb_steps=args.stack_nb_steps,
 566         nb_stacks=args.stack_nb_stacks,
 567         nb_digits=args.stack_nb_digits,
 568         fraction_values_for_train=args.stack_fraction_values_for_train,
 569         device=device,
 570     )
 571
 572 elif args.task == "expr":
 573     task = tasks.Expr(
 574         nb_train_samples=args.nb_train_samples,
 575         nb_test_samples=args.nb_test_samples,
 576         nb_variables=args.expr_nb_variables,
 577         sequence_length=args.expr_sequence_length,
 578         operand_max=args.expr_operand_max,
 579         result_max=args.expr_result_max,
 580         batch_size=args.batch_size,
 581         device=device,
 582     )
 583
 584 elif args.task == "rpl":
 585     task = tasks.RPL(
 586         nb_train_samples=args.nb_train_samples,
 587         nb_test_samples=args.nb_test_samples,
 588         batch_size=args.batch_size,
 589         nb_starting_values=args.rpl_nb_starting_values,
 590         max_input=args.rpl_max_input,
 591         prog_len=args.rpl_prog_len,
 592         nb_runs=args.rpl_nb_runs,
 593         no_prog=args.rpl_no_prog,
 594         logger=log_string,
 595         device=device,
 596     )
 597
 598 elif args.task == "grid":
 599     task = tasks.Grid(
 600         nb_train_samples=args.nb_train_samples,
 601         nb_test_samples=args.nb_test_samples,
 602         batch_size=args.batch_size,
 603         size=args.grid_size,
 604         fraction_play=args.grid_fraction_play,
 605         logger=log_string,
 606         device=device,
 607     )
 608
 609 elif args.task == "qmlp":
 610     task = tasks.QMLP(
 611         nb_train_samples=args.nb_train_samples,
 612         nb_test_samples=args.nb_test_samples,
 613         batch_size=args.batch_size,
 614         result_dir=args.result_dir,
 615         logger=log_string,
 616         device=device,
 617     )
 618
 619 elif args.task == "escape":
 620     task = tasks.Escape(
 621         nb_train_samples=args.nb_train_samples,
 622         nb_test_samples=args.nb_test_samples,
 623         batch_size=args.batch_size,
 624         height=args.escape_height,
 625         width=args.escape_width,
 626         T=args.escape_T,
 627         nb_walls=args.escape_nb_walls,
 628         logger=log_string,
 629         device=device,
 630     )
 631
 632 else:
 633     raise ValueError(f"Unknown task {args.task}")
 634
 635 ######################################################################
 636
 637 log_string(f"device {device}")
 638
 639 vocabulary_size = task.vocabulary_size()
 640
 641 log_string(f"vocabulary_size {vocabulary_size}")
 642
 643 ##############################
 644
 645 model = mygpt.MyGPT(
 646     vocabulary_size=vocabulary_size,
 647     dim_model=args.dim_model,
 648     dim_keys=args.dim_keys,
 649     dim_hidden=args.dim_hidden,
 650     nb_heads=args.nb_heads,
 651     nb_blocks=args.nb_blocks,
 652     causal=True,
 653     dropout=args.dropout,
 654 )
 655
 656 model.to(device)
 657
 658 nb_parameters = sum(p.numel() for p in model.parameters())
 659 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 660
 661 ######################################################################
 662
 663 nb_epochs_finished = 0
 664
 665 if args.no_checkpoint:
 666     log_string(f"not trying to load checkpoint.")
 667
 668 else:
 669     try:
 670         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 671         checkpoint = torch.load(checkpoint_name)
 672         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 673         model.load_state_dict(checkpoint["model_state"])
 674         torch.set_rng_state(checkpoint["rng_state"])
 675         if torch.cuda.is_available():
 676             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 677
 678         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 679
 680     except FileNotFoundError:
 681         log_string("starting from scratch.")
 682
 683     except:
 684         log_string("error when loading the checkpoint.")
 685         exit(1)
 686
 687 ######################################################################
 688
 689 if args.task == "expr" and args.expr_input_file is not None:
 690     task.produce_results(
 691         n_epoch=nb_epochs_finished,
 692         model=model,
 693         result_dir=args.result_dir,
 694         logger=log_string,
 695         deterministic_synthesis=args.deterministic_synthesis,
 696         input_file=args.expr_input_file,
 697     )
 698
 699     exit(0)
 700
 701 ######################################################################
 702
 703 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 704
 705 # Compute the entropy of the training tokens
 706
 707 token_count = 0
 708 for input in task.batches(split="train"):
 709     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 710 token_probas = token_count / token_count.sum()
 711 entropy = -torch.xlogy(token_probas, token_probas).sum()
 712 train_set_perplexity = math.exp(entropy)
 713
 714 ######################################################################
 715 # A bit of paranoia never hurts
 716
 717 if args.max_percents_of_test_in_train >= 0:
 718
 719     def subsets_as_tuples(batches, cs):
 720         s = set()
 721         for batch in batches:
 722             for x in batch:
 723                 s.add(tuple([v.item() for v in x]))
 724                 if len(s) == cs:
 725                     yield s
 726                     s = set()
 727         yield s
 728
 729     nb_test, nb_in_train = 0, 0
 730     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 731         in_train = set()
 732         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 733             in_train.update(test_subset.intersection(train_subset))
 734         nb_in_train += len(in_train)
 735         nb_test += len(test_subset)
 736
 737     log_string(
 738         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 739     )
 740
 741     assert (
 742         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 743     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 744
 745 ##############################
 746
 747 if args.learning_rate_schedule == "cos":
 748     learning_rate_schedule = {}
 749     for n_epoch in range(args.nb_epochs):
 750         u = n_epoch / args.nb_epochs * math.pi
 751         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 752 else:
 753     u = {
 754         int(k): float(v)
 755         for k, v in [
 756             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 757         ]
 758     }
 759
 760     learning_rate_schedule = {}
 761     learning_rate = args.learning_rate
 762     for n_epoch in range(args.nb_epochs):
 763         if n_epoch in u:
 764             learning_rate = u[n_epoch]
 765         learning_rate_schedule[n_epoch] = learning_rate
 766
 767 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 768
 769 ##############################
 770
 771 nb_samples_seen = 0
 772
 773 if nb_epochs_finished >= nb_epochs:
 774     task.produce_results(
 775         n_epoch=nb_epochs_finished,
 776         model=model,
 777         result_dir=args.result_dir,
 778         logger=log_string,
 779         deterministic_synthesis=args.deterministic_synthesis,
 780     )
 781
 782 time_pred_result = None
 783
 784 for n_epoch in range(nb_epochs_finished, nb_epochs):
 785     learning_rate = learning_rate_schedule[n_epoch]
 786
 787     log_string(f"learning_rate {learning_rate}")
 788
 789     if args.optim == "sgd":
 790         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 791     elif args.optim == "adam":
 792         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 793     elif args.optim == "adamw":
 794         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 795     else:
 796         raise ValueError(f"Unknown optimizer {args.optim}.")
 797
 798     model.train()
 799
 800     nb_train_samples, acc_train_loss = 0, 0.0
 801
 802     for input in task.batches(split="train"):
 803         input = input.to(device)
 804         output = model(mygpt.BracketedSequence(input)).x
 805         loss = F.cross_entropy(output.transpose(1, 2), input)
 806         acc_train_loss += loss.item() * input.size(0)
 807         nb_train_samples += input.size(0)
 808         nb_samples_seen += input.size(0)
 809
 810         optimizer.zero_grad()
 811         loss.backward()
 812         optimizer.step()
 813
 814     with torch.autograd.no_grad():
 815         model.eval()
 816
 817         nb_test_samples, acc_test_loss = 0, 0.0
 818
 819         for input in task.batches(split="test"):
 820             input = input.to(device)
 821
 822             output = model(mygpt.BracketedSequence(input)).x
 823             loss = F.cross_entropy(output.transpose(1, 2), input)
 824             acc_test_loss += loss.item() * input.size(0)
 825             nb_test_samples += input.size(0)
 826
 827         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 828         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 829
 830         log_string(
 831             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 832         )
 833
 834         task.produce_results(
 835             n_epoch=n_epoch,
 836             model=model,
 837             result_dir=args.result_dir,
 838             logger=log_string,
 839             deterministic_synthesis=args.deterministic_synthesis,
 840         )
 841
 842         time_current_result = datetime.datetime.now()
 843         if time_pred_result is not None:
 844             log_string(
 845                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 846             )
 847         time_pred_result = time_current_result
 848
 849     checkpoint = {
 850         "nb_epochs_finished": n_epoch + 1,
 851         "model_state": model.state_dict(),
 852         "rng_state": torch.get_rng_state(),
 853     }
 854
 855     if torch.cuda.is_available():
 856         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 857
 858     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 859     torch.save(checkpoint, checkpoint_name)
 860     log_string(f"saved checkpoint {checkpoint_name}")
 861
 862 ######################################################################