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=4)
 182
 183 parser.add_argument("--escape_width", type=int, default=6)
 184
 185 parser.add_argument("--escape_T", type=int, default=20)
 186
 187 ######################################################################
 188
 189 args = parser.parse_args()
 190
 191 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 192
 193 if args.result_dir is None:
 194     args.result_dir = f"results_{args.task}"
 195
 196 ######################################################################
 197
 198 default_task_args = {
 199     "file": {
 200         "model": "37M",
 201         "batch_size": 25,
 202         "nb_train_samples": 250000,
 203         "nb_test_samples": 10000,
 204     },
 205     "addition": {
 206         "model": "352M",
 207         "batch_size": 25,
 208         "nb_train_samples": 250000,
 209         "nb_test_samples": 10000,
 210     },
 211     "byheart": {
 212         "model": "37M",
 213         "batch_size": 25,
 214         "nb_train_samples": 50000,
 215         "nb_test_samples": 10000,
 216     },
 217     "expr": {
 218         "model": "352M",
 219         "batch_size": 25,
 220         "nb_train_samples": 2500000,
 221         "nb_test_samples": 10000,
 222     },
 223     "grid": {
 224         "model": "37M",
 225         "batch_size": 25,
 226         "nb_train_samples": 250000,
 227         "nb_test_samples": 10000,
 228     },
 229     "qmlp": {
 230         "model": "37M",
 231         "batch_size": 10,
 232         "nb_train_samples": 100000,
 233         "nb_test_samples": 1000,
 234     },
 235     "guessop": {
 236         "model": "352M",
 237         "batch_size": 25,
 238         "nb_train_samples": 1000000,
 239         "nb_test_samples": 10000,
 240     },
 241     "learnop": {
 242         "model": "37M",
 243         "batch_size": 25,
 244         "nb_train_samples": 50000,
 245         "nb_test_samples": 10000,
 246     },
 247     "maze": {
 248         "model": "37M",
 249         "batch_size": 5,
 250         "nb_train_samples": 100000,
 251         "nb_test_samples": 10000,
 252     },
 253     "picoclvr": {
 254         "model": "37M",
 255         "batch_size": 25,
 256         "nb_train_samples": 250000,
 257         "nb_test_samples": 10000,
 258     },
 259     "rpl": {
 260         "model": "352M",
 261         "batch_size": 5,
 262         "nb_train_samples": 2500000,
 263         "nb_test_samples": 10000,
 264     },
 265     "snake": {
 266         "model": "37M",
 267         "batch_size": 25,
 268         "nb_train_samples": 250000,
 269         "nb_test_samples": 10000,
 270     },
 271     "stack": {
 272         "model": "37M",
 273         "batch_size": 25,
 274         "nb_train_samples": 100000,
 275         "nb_test_samples": 1000,
 276     },
 277     "twotargets": {
 278         "model": "37M",
 279         "batch_size": 25,
 280         "nb_train_samples": 50000,
 281         "nb_test_samples": 10000,
 282     },
 283     "memory": {
 284         "model": "37M",
 285         "batch_size": 100,
 286         "nb_train_samples": 25000,
 287         "nb_test_samples": 1000,
 288     },
 289     "mixing": {
 290         "model": "37M",
 291         "batch_size": 25,
 292         "nb_train_samples": 250000,
 293         "nb_test_samples": 10000,
 294     },
 295     "mnist": {
 296         "model": "37M",
 297         "batch_size": 10,
 298         "nb_train_samples": 60000,
 299         "nb_test_samples": 10000,
 300     },
 301     "escape": {
 302         "model": "37M",
 303         "batch_size": 25,
 304         "nb_train_samples": 25000,
 305         "nb_test_samples": 10000,
 306     },
 307 }
 308
 309 if args.task in default_task_args:
 310     for k, v in default_task_args[args.task].items():
 311         if getattr(args, k) is None:
 312             setattr(args, k, v)
 313
 314 ######################################################################
 315
 316 default_model_args = {
 317     "17K": {
 318         "dim_model": 32,
 319         "dim_keys": 32,
 320         "dim_hidden": 32,
 321         "nb_heads": 2,
 322         "nb_blocks": 2,
 323     },
 324     "4M": {
 325         "dim_model": 256,
 326         "dim_keys": 32,
 327         "dim_hidden": 1024,
 328         "nb_heads": 4,
 329         "nb_blocks": 6,
 330     },
 331     "37M": {
 332         "dim_model": 512,
 333         "dim_keys": 64,
 334         "dim_hidden": 2048,
 335         "nb_heads": 8,
 336         "nb_blocks": 12,
 337     },
 338     "122M": {
 339         "dim_model": 768,
 340         "dim_keys": 64,
 341         "dim_hidden": 2048,
 342         "nb_heads": 8,
 343         "nb_blocks": 24,
 344     },
 345     "352M": {
 346         "dim_model": 1024,
 347         "dim_keys": 64,
 348         "dim_hidden": 2048,
 349         "nb_heads": 8,
 350         "nb_blocks": 48,
 351     },
 352 }
 353
 354 if args.model in default_model_args:
 355     for k, v in default_model_args[args.model].items():
 356         if getattr(args, k) is None:
 357             setattr(args, k, v)
 358 else:
 359     raise ValueError(f"Unknown model {args.model}")
 360
 361 ######################################################################
 362
 363 try:
 364     os.mkdir(args.result_dir)
 365 except FileExistsError:
 366     if not args.overwrite_results:
 367         print(f"result directory {args.result_dir} already exists")
 368         exit(1)
 369
 370 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 371
 372 if args.seed >= 0:
 373     # torch.backends.cudnn.deterministic = True
 374     # torch.backends.cudnn.benchmark = False
 375     # torch.use_deterministic_algorithms(True)
 376     torch.manual_seed(args.seed)
 377     if torch.cuda.is_available():
 378         torch.cuda.manual_seed_all(args.seed)
 379
 380 ######################################################################
 381
 382
 383 def log_string(s):
 384     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 385
 386     if log_file is not None:
 387         log_file.write(t + s + "\n")
 388         log_file.flush()
 389
 390     print(t + s)
 391     sys.stdout.flush()
 392
 393
 394 log_string(f"argv {' '.join(sys.argv)}")
 395
 396 for n in vars(args):
 397     log_string(f"args.{n} {getattr(args, n)}")
 398
 399
 400 ######################################################################
 401
 402
 403 def picoclvr_pruner_horizontal_green(p):
 404     return not ("green" in p and ("left" in p or "right" in p))
 405
 406
 407 picoclvr_pruner_train = (
 408     picoclvr_pruner_horizontal_green
 409     if args.picocvlr_prune_properties in {"train+eval"}
 410     else None
 411 )
 412
 413 picoclvr_pruner_eval = (
 414     (lambda p: not picoclvr_pruner_horizontal_green(p))
 415     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 416     else None
 417 )
 418
 419 ######################################################################
 420
 421 if args.task == "file":
 422     assert (
 423         args.filetask_train_file is not None and args.filetask_test_file is not None
 424     ), "You have to specify the task train and test files"
 425     task = tasks.TaskFromFile(
 426         args.filetask_train_file,
 427         args.filetask_test_file,
 428         nb_train_samples=args.nb_train_samples,
 429         nb_test_samples=args.nb_test_samples,
 430         batch_size=args.batch_size,
 431         shuffle=True,
 432         device=device,
 433     )
 434     args.max_percents_of_test_in_train = 0
 435
 436 elif args.task == "byheart":
 437     task = tasks.SandBox(
 438         problem=problems.ProblemByHeart(),
 439         nb_train_samples=args.nb_train_samples,
 440         nb_test_samples=args.nb_test_samples,
 441         batch_size=args.batch_size,
 442         logger=log_string,
 443         device=device,
 444     )
 445     args.max_percents_of_test_in_train = -1
 446
 447 elif args.task == "learnop":
 448     task = tasks.SandBox(
 449         problem=problems.ProblemLearnOperator(),
 450         nb_train_samples=args.nb_train_samples,
 451         nb_test_samples=args.nb_test_samples,
 452         batch_size=args.batch_size,
 453         logger=log_string,
 454         device=device,
 455     )
 456
 457
 458 elif args.task == "guessop":
 459     task = tasks.SandBox(
 460         problem=problems.ProblemGuessOperator(),
 461         nb_train_samples=args.nb_train_samples,
 462         nb_test_samples=args.nb_test_samples,
 463         batch_size=args.batch_size,
 464         logger=log_string,
 465         device=device,
 466     )
 467
 468
 469 elif args.task == "twotargets":
 470     task = tasks.SandBox(
 471         problem=problems.ProblemTwoTargets(),
 472         nb_train_samples=args.nb_train_samples,
 473         nb_test_samples=args.nb_test_samples,
 474         batch_size=args.batch_size,
 475         logger=log_string,
 476         device=device,
 477     )
 478
 479 elif args.task == "memory":
 480     task = tasks.SandBox(
 481         problem=problems.ProblemMemory(),
 482         nb_train_samples=args.nb_train_samples,
 483         nb_test_samples=args.nb_test_samples,
 484         batch_size=args.batch_size,
 485         logger=log_string,
 486         device=device,
 487     )
 488
 489 elif args.task == "mixing":
 490     task = tasks.SandBox(
 491         problem=problems.ProblemMixing(
 492             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 493         ),
 494         nb_train_samples=args.nb_train_samples,
 495         nb_test_samples=args.nb_test_samples,
 496         batch_size=args.batch_size,
 497         logger=log_string,
 498         device=device,
 499     )
 500
 501 elif args.task == "addition":
 502     task = tasks.SandBox(
 503         problem=problems.ProblemAddition(),
 504         nb_train_samples=args.nb_train_samples,
 505         nb_test_samples=args.nb_test_samples,
 506         batch_size=args.batch_size,
 507         logger=log_string,
 508         device=device,
 509     )
 510
 511 elif args.task == "picoclvr":
 512     task = tasks.PicoCLVR(
 513         nb_train_samples=args.nb_train_samples,
 514         nb_test_samples=args.nb_test_samples,
 515         batch_size=args.batch_size,
 516         height=args.picoclvr_height,
 517         width=args.picoclvr_width,
 518         nb_colors=args.picoclvr_nb_colors,
 519         logger=log_string,
 520         device=device,
 521         pruner_train=picoclvr_pruner_train,
 522         pruner_eval=picoclvr_pruner_eval,
 523     )
 524
 525 elif args.task == "mnist":
 526     task = tasks.MNIST(
 527         nb_train_samples=args.nb_train_samples,
 528         nb_test_samples=args.nb_test_samples,
 529         batch_size=args.batch_size,
 530         device=device,
 531     )
 532
 533 elif args.task == "maze":
 534     task = tasks.Maze(
 535         nb_train_samples=args.nb_train_samples,
 536         nb_test_samples=args.nb_test_samples,
 537         batch_size=args.batch_size,
 538         height=args.maze_height,
 539         width=args.maze_width,
 540         nb_walls=args.maze_nb_walls,
 541         device=device,
 542     )
 543
 544 elif args.task == "snake":
 545     task = tasks.Snake(
 546         nb_train_samples=args.nb_train_samples,
 547         nb_test_samples=args.nb_test_samples,
 548         batch_size=args.batch_size,
 549         height=args.snake_height,
 550         width=args.snake_width,
 551         nb_colors=args.snake_nb_colors,
 552         length=args.snake_length,
 553         prompt_length=args.snake_length // 2,
 554         device=device,
 555     )
 556
 557 elif args.task == "stack":
 558     task = tasks.Stack(
 559         nb_train_samples=args.nb_train_samples,
 560         nb_test_samples=args.nb_test_samples,
 561         batch_size=args.batch_size,
 562         logger=log_string,
 563         nb_steps=args.stack_nb_steps,
 564         nb_stacks=args.stack_nb_stacks,
 565         nb_digits=args.stack_nb_digits,
 566         fraction_values_for_train=args.stack_fraction_values_for_train,
 567         device=device,
 568     )
 569
 570 elif args.task == "expr":
 571     task = tasks.Expr(
 572         nb_train_samples=args.nb_train_samples,
 573         nb_test_samples=args.nb_test_samples,
 574         nb_variables=args.expr_nb_variables,
 575         sequence_length=args.expr_sequence_length,
 576         operand_max=args.expr_operand_max,
 577         result_max=args.expr_result_max,
 578         batch_size=args.batch_size,
 579         device=device,
 580     )
 581
 582 elif args.task == "rpl":
 583     task = tasks.RPL(
 584         nb_train_samples=args.nb_train_samples,
 585         nb_test_samples=args.nb_test_samples,
 586         batch_size=args.batch_size,
 587         nb_starting_values=args.rpl_nb_starting_values,
 588         max_input=args.rpl_max_input,
 589         prog_len=args.rpl_prog_len,
 590         nb_runs=args.rpl_nb_runs,
 591         no_prog=args.rpl_no_prog,
 592         logger=log_string,
 593         device=device,
 594     )
 595
 596 elif args.task == "grid":
 597     task = tasks.Grid(
 598         nb_train_samples=args.nb_train_samples,
 599         nb_test_samples=args.nb_test_samples,
 600         batch_size=args.batch_size,
 601         size=args.grid_size,
 602         fraction_play=args.grid_fraction_play,
 603         logger=log_string,
 604         device=device,
 605     )
 606
 607 elif args.task == "qmlp":
 608     task = tasks.QMLP(
 609         nb_train_samples=args.nb_train_samples,
 610         nb_test_samples=args.nb_test_samples,
 611         batch_size=args.batch_size,
 612         result_dir=args.result_dir,
 613         logger=log_string,
 614         device=device,
 615     )
 616
 617 elif args.task == "escape":
 618     task = tasks.Escape(
 619         nb_train_samples=args.nb_train_samples,
 620         nb_test_samples=args.nb_test_samples,
 621         batch_size=args.batch_size,
 622         height=args.escape_height,
 623         width=args.escape_width,
 624         T=args.escape_T,
 625         logger=log_string,
 626         device=device,
 627     )
 628
 629 else:
 630     raise ValueError(f"Unknown task {args.task}")
 631
 632 ######################################################################
 633
 634 log_string(f"device {device}")
 635
 636 vocabulary_size = task.vocabulary_size()
 637
 638 log_string(f"vocabulary_size {vocabulary_size}")
 639
 640 ##############################
 641
 642 model = mygpt.MyGPT(
 643     vocabulary_size=vocabulary_size,
 644     dim_model=args.dim_model,
 645     dim_keys=args.dim_keys,
 646     dim_hidden=args.dim_hidden,
 647     nb_heads=args.nb_heads,
 648     nb_blocks=args.nb_blocks,
 649     causal=True,
 650     dropout=args.dropout,
 651 )
 652
 653 model.to(device)
 654
 655 nb_parameters = sum(p.numel() for p in model.parameters())
 656 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 657
 658 ######################################################################
 659
 660 nb_epochs_finished = 0
 661
 662 if args.no_checkpoint:
 663     log_string(f"not trying to load checkpoint.")
 664
 665 else:
 666     try:
 667         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 668         checkpoint = torch.load(checkpoint_name)
 669         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 670         model.load_state_dict(checkpoint["model_state"])
 671         torch.set_rng_state(checkpoint["rng_state"])
 672         if torch.cuda.is_available():
 673             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 674
 675         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 676
 677     except FileNotFoundError:
 678         log_string("starting from scratch.")
 679
 680     except:
 681         log_string("error when loading the checkpoint.")
 682         exit(1)
 683
 684 ######################################################################
 685
 686 if args.task == "expr" and args.expr_input_file is not None:
 687     task.produce_results(
 688         n_epoch=nb_epochs_finished,
 689         model=model,
 690         result_dir=args.result_dir,
 691         logger=log_string,
 692         deterministic_synthesis=args.deterministic_synthesis,
 693         input_file=args.expr_input_file,
 694     )
 695
 696     exit(0)
 697
 698 ######################################################################
 699
 700 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 701
 702 # Compute the entropy of the training tokens
 703
 704 token_count = 0
 705 for input in task.batches(split="train"):
 706     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 707 token_probas = token_count / token_count.sum()
 708 entropy = -torch.xlogy(token_probas, token_probas).sum()
 709 train_set_perplexity = math.exp(entropy)
 710
 711 ######################################################################
 712 # A bit of paranoia never hurts
 713
 714 if args.max_percents_of_test_in_train >= 0:
 715
 716     def subsets_as_tuples(batches, cs):
 717         s = set()
 718         for batch in batches:
 719             for x in batch:
 720                 s.add(tuple([v.item() for v in x]))
 721                 if len(s) == cs:
 722                     yield s
 723                     s = set()
 724         yield s
 725
 726     nb_test, nb_in_train = 0, 0
 727     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 728         in_train = set()
 729         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 730             in_train.update(test_subset.intersection(train_subset))
 731         nb_in_train += len(in_train)
 732         nb_test += len(test_subset)
 733
 734     log_string(
 735         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 736     )
 737
 738     assert (
 739         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 740     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 741
 742 ##############################
 743
 744 if args.learning_rate_schedule == "cos":
 745     learning_rate_schedule = {}
 746     for n_epoch in range(args.nb_epochs):
 747         u = n_epoch / args.nb_epochs * math.pi
 748         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 749 else:
 750     u = {
 751         int(k): float(v)
 752         for k, v in [
 753             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 754         ]
 755     }
 756
 757     learning_rate_schedule = {}
 758     learning_rate = args.learning_rate
 759     for n_epoch in range(args.nb_epochs):
 760         if n_epoch in u:
 761             learning_rate = u[n_epoch]
 762         learning_rate_schedule[n_epoch] = learning_rate
 763
 764 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 765
 766 ##############################
 767
 768 nb_samples_seen = 0
 769
 770 if nb_epochs_finished >= nb_epochs:
 771     task.produce_results(
 772         n_epoch=nb_epochs_finished,
 773         model=model,
 774         result_dir=args.result_dir,
 775         logger=log_string,
 776         deterministic_synthesis=args.deterministic_synthesis,
 777     )
 778
 779 time_pred_result = None
 780
 781 for n_epoch in range(nb_epochs_finished, nb_epochs):
 782     learning_rate = learning_rate_schedule[n_epoch]
 783
 784     log_string(f"learning_rate {learning_rate}")
 785
 786     if args.optim == "sgd":
 787         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 788     elif args.optim == "adam":
 789         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 790     elif args.optim == "adamw":
 791         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 792     else:
 793         raise ValueError(f"Unknown optimizer {args.optim}.")
 794
 795     model.train()
 796
 797     nb_train_samples, acc_train_loss = 0, 0.0
 798
 799     for input in task.batches(split="train"):
 800         input = input.to(device)
 801         output = model(mygpt.BracketedSequence(input)).x
 802         loss = F.cross_entropy(output.transpose(1, 2), input)
 803         acc_train_loss += loss.item() * input.size(0)
 804         nb_train_samples += input.size(0)
 805         nb_samples_seen += input.size(0)
 806
 807         optimizer.zero_grad()
 808         loss.backward()
 809         optimizer.step()
 810
 811     with torch.autograd.no_grad():
 812         model.eval()
 813
 814         nb_test_samples, acc_test_loss = 0, 0.0
 815
 816         for input in task.batches(split="test"):
 817             input = input.to(device)
 818
 819             output = model(mygpt.BracketedSequence(input)).x
 820             loss = F.cross_entropy(output.transpose(1, 2), input)
 821             acc_test_loss += loss.item() * input.size(0)
 822             nb_test_samples += input.size(0)
 823
 824         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 825         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 826
 827         log_string(
 828             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 829         )
 830
 831         task.produce_results(
 832             n_epoch=n_epoch,
 833             model=model,
 834             result_dir=args.result_dir,
 835             logger=log_string,
 836             deterministic_synthesis=args.deterministic_synthesis,
 837         )
 838
 839         time_current_result = datetime.datetime.now()
 840         if time_pred_result is not None:
 841             log_string(
 842                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 843             )
 844         time_pred_result = time_current_result
 845
 846     checkpoint = {
 847         "nb_epochs_finished": n_epoch + 1,
 848         "model_state": model.state_dict(),
 849         "rng_state": torch.get_rng_state(),
 850     }
 851
 852     if torch.cuda.is_available():
 853         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 854
 855     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 856     torch.save(checkpoint, checkpoint_name)
 857     log_string(f"saved checkpoint {checkpoint_name}")
 858
 859 ######################################################################