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, warnings
   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
  28 def str2bool(x):
  29     x = x.lower()
  30     if x in {"1", "true", "yes"}:
  31         return True
  32     elif x in {"0", "false", "no"}:
  33         return False
  34     else:
  35         raise ValueError
  36
  37
  38 parser = argparse.ArgumentParser(
  39     description="An implementation of GPT with cache.",
  40     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  41 )
  42
  43 parser.add_argument(
  44     "--task",
  45     type=str,
  46     default="twotargets",
  47     help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
  48 )
  49
  50 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  51
  52 parser.add_argument("--result_dir", type=str, default=None)
  53
  54 parser.add_argument("--seed", type=int, default=0)
  55
  56 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  57
  58 ########################################
  59
  60 parser.add_argument("--nb_epochs", type=int, default=50)
  61
  62 parser.add_argument("--batch_size", type=int, default=None)
  63
  64 parser.add_argument("--nb_train_samples", type=int, default=None)
  65
  66 parser.add_argument("--nb_test_samples", type=int, default=None)
  67
  68 parser.add_argument("--optim", type=str, default="adam")
  69
  70 ########################################
  71
  72 parser.add_argument("--nb_warmup_iter", type=int, default=100)
  73
  74 parser.add_argument("--nb_decay_iter", type=int, default=5000)
  75
  76 parser.add_argument("--learning_rate", type=float, default=6e-4)
  77
  78 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
  79
  80 # legacy
  81
  82 parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
  83
  84 parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
  85
  86 parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
  87
  88 parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
  89
  90 ########################################
  91
  92 parser.add_argument("--model", type=str, default=None)
  93
  94 parser.add_argument("--attention", type=str, default=None)
  95
  96 parser.add_argument("--dim_model", type=int, default=None)
  97
  98 parser.add_argument("--dim_keys", type=int, default=None)
  99
 100 parser.add_argument("--dim_hidden", type=int, default=None)
 101
 102 parser.add_argument("--nb_heads", type=int, default=None)
 103
 104 parser.add_argument("--nb_lines", type=int, default=None)
 105
 106 parser.add_argument("--caterpillar_height", type=int, default=None)
 107
 108 parser.add_argument("--rho", type=float, default=0.0)
 109
 110 parser.add_argument("--nb_blocks", type=int, default=None)
 111
 112 parser.add_argument("--dropout", type=float, default=0.1)
 113
 114 ########################################
 115
 116 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 117
 118 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 119
 120 parser.add_argument("--continue_training", action="store_true", default=False)
 121
 122 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 123
 124 ##############################
 125 # rpl options
 126
 127 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 128
 129 parser.add_argument("--rpl_max_input", type=int, default=9)
 130
 131 parser.add_argument("--rpl_prog_len", type=int, default=8)
 132
 133 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 134
 135 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 136
 137 ##############################
 138 # grid options
 139
 140 parser.add_argument("--grid_size", type=int, default=6)
 141
 142 ##############################
 143 # picoclvr options
 144
 145 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 146
 147 parser.add_argument("--picoclvr_height", type=int, default=12)
 148
 149 parser.add_argument("--picoclvr_width", type=int, default=16)
 150
 151 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 152
 153 ##############################
 154 # Maze options
 155
 156 parser.add_argument("--maze_height", type=int, default=13)
 157
 158 parser.add_argument("--maze_width", type=int, default=21)
 159
 160 parser.add_argument("--maze_nb_walls", type=int, default=15)
 161
 162 ##############################
 163 # Snake options
 164
 165 parser.add_argument("--snake_height", type=int, default=9)
 166
 167 parser.add_argument("--snake_width", type=int, default=12)
 168
 169 parser.add_argument("--snake_nb_colors", type=int, default=5)
 170
 171 parser.add_argument("--snake_length", type=int, default=200)
 172
 173 ##############################
 174 # Stack options
 175
 176 parser.add_argument("--stack_nb_steps", type=int, default=100)
 177
 178 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 179
 180 parser.add_argument("--stack_nb_digits", type=int, default=3)
 181
 182 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 183
 184 ##############################
 185 # Expr options
 186
 187 parser.add_argument("--expr_nb_variables", type=int, default=5)
 188
 189 parser.add_argument("--expr_sequence_length", type=int, default=40)
 190
 191 parser.add_argument("--expr_operand_max", type=int, default=9)
 192
 193 parser.add_argument("--expr_result_max", type=int, default=99)
 194
 195 parser.add_argument("--expr_input_file", type=str, default=None)
 196
 197 ##############################
 198 # Memory
 199
 200 parser.add_argument("--memory_len_total", type=int, default=32)
 201
 202 ##############################
 203 # Mixing
 204
 205 parser.add_argument("--mixing_hard", action="store_true", default=False)
 206
 207 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 208
 209 ######################################################################
 210
 211 args = parser.parse_args()
 212
 213 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 214
 215 if args.result_dir is None:
 216     args.result_dir = f"results_{args.task}_{args.model}"
 217
 218 ######################################################################
 219
 220 default_task_args = {
 221     "addition": {
 222         "model": "352M",
 223         "batch_size": 25,
 224         "nb_train_samples": 250000,
 225         "nb_test_samples": 10000,
 226     },
 227     "byheart": {
 228         "model": "37M",
 229         "batch_size": 25,
 230         "nb_train_samples": 50000,
 231         "nb_test_samples": 10000,
 232     },
 233     "expr": {
 234         "model": "352M",
 235         "batch_size": 25,
 236         "nb_train_samples": 2500000,
 237         "nb_test_samples": 10000,
 238     },
 239     "grid": {
 240         "model": "37M",
 241         "batch_size": 25,
 242         "nb_train_samples": 250000,
 243         "nb_test_samples": 10000,
 244     },
 245     "qmlp": {
 246         "model": "37M",
 247         "batch_size": 10,
 248         "nb_train_samples": 100000,
 249         "nb_test_samples": 1000,
 250     },
 251     "guessop": {
 252         "model": "352M",
 253         "batch_size": 25,
 254         "nb_train_samples": 1000000,
 255         "nb_test_samples": 10000,
 256     },
 257     "learnop": {
 258         "model": "37M",
 259         "batch_size": 25,
 260         "nb_train_samples": 50000,
 261         "nb_test_samples": 10000,
 262     },
 263     "maze": {
 264         "model": "37M",
 265         "batch_size": 5,
 266         "nb_train_samples": 100000,
 267         "nb_test_samples": 10000,
 268     },
 269     "picoclvr": {
 270         "model": "37M",
 271         "batch_size": 25,
 272         "nb_train_samples": 250000,
 273         "nb_test_samples": 10000,
 274     },
 275     "rpl": {
 276         "model": "352M",
 277         "batch_size": 5,
 278         "nb_train_samples": 2500000,
 279         "nb_test_samples": 10000,
 280     },
 281     "snake": {
 282         "model": "37M",
 283         "batch_size": 25,
 284         "nb_train_samples": 250000,
 285         "nb_test_samples": 10000,
 286     },
 287     "stack": {
 288         "model": "37M",
 289         "batch_size": 25,
 290         "nb_train_samples": 100000,
 291         "nb_test_samples": 1000,
 292     },
 293     "twotargets": {
 294         "model": "37M",
 295         "batch_size": 25,
 296         "nb_train_samples": 50000,
 297         "nb_test_samples": 10000,
 298     },
 299     "memory": {
 300         "model": "37M",
 301         "batch_size": 25,
 302         "nb_train_samples": 25000,
 303         "nb_test_samples": 10000,
 304     },
 305     "mixing": {
 306         "model": "37M",
 307         "batch_size": 25,
 308         "nb_train_samples": 250000,
 309         "nb_test_samples": 10000,
 310     },
 311     "mnist": {
 312         "model": "37M",
 313         "batch_size": 10,
 314         "nb_train_samples": 60000,
 315         "nb_test_samples": 10000,
 316     },
 317 }
 318
 319 if args.task in default_task_args:
 320     for k, v in default_task_args[args.task].items():
 321         if getattr(args, k) is None:
 322             setattr(args, k, v)
 323
 324 ######################################################################
 325
 326 default_model_args = {
 327     "17K": {
 328         "attention": "mha",
 329         "dim_model": 32,
 330         "dim_keys": 32,
 331         "dim_hidden": 32,
 332         "nb_heads": 2,
 333         "nb_blocks": 2,
 334     },
 335     "17K-C": {
 336         "attention": "caterpillar",
 337         "dim_model": 32,
 338         "dim_keys": 32,
 339         "dim_hidden": 32,
 340         "nb_heads": 2,
 341         "nb_lines": 16,
 342         "caterpillar_height": 4,
 343         "nb_blocks": 2,
 344     },
 345     "4M": {
 346         "attention": "mha",
 347         "dim_model": 256,
 348         "dim_keys": 32,
 349         "dim_hidden": 1024,
 350         "nb_heads": 4,
 351         "nb_blocks": 6,
 352     },
 353     "4M-C": {
 354         "attention": "caterpillar",
 355         "dim_model": 256,
 356         "dim_keys": 32,
 357         "dim_hidden": 1024,
 358         "nb_heads": 4,
 359         "nb_lines": 32,
 360         "caterpillar_height": 4,
 361         "nb_blocks": 6,
 362     },
 363     "37M": {
 364         "attention": "mha",
 365         "dim_model": 512,
 366         "dim_keys": 64,
 367         "dim_hidden": 2048,
 368         "nb_heads": 8,
 369         "nb_blocks": 12,
 370     },
 371     "37M-C": {
 372         "attention": "caterpillar",
 373         "dim_model": 512,
 374         "dim_keys": 64,
 375         "dim_hidden": 2048,
 376         "nb_heads": 8,
 377         "nb_lines": 256,
 378         "caterpillar_height": 32,
 379         "nb_blocks": 12,
 380     },
 381     "122M": {
 382         "attention": "mha",
 383         "dim_model": 768,
 384         "dim_keys": 64,
 385         "dim_hidden": 2048,
 386         "nb_heads": 8,
 387         "nb_blocks": 24,
 388     },
 389     "122M-C": {
 390         "attention": "caterpillar",
 391         "dim_model": 768,
 392         "dim_keys": 64,
 393         "dim_hidden": 2048,
 394         "nb_heads": 8,
 395         "nb_lines": 128,
 396         "nb_blocks": 24,
 397     },
 398     "352M": {
 399         "attention": "mha",
 400         "dim_model": 1024,
 401         "dim_keys": 64,
 402         "dim_hidden": 2048,
 403         "nb_heads": 8,
 404         "nb_blocks": 48,
 405     },
 406     "352M-C": {
 407         "attention": "caterpillar",
 408         "dim_model": 1024,
 409         "dim_keys": 64,
 410         "dim_hidden": 2048,
 411         "nb_heads": 8,
 412         "nb_lines": 128,
 413         "nb_blocks": 48,
 414     },
 415 }
 416
 417 if args.model in default_model_args:
 418     for k, v in default_model_args[args.model].items():
 419         if getattr(args, k) is None:
 420             setattr(args, k, v)
 421 else:
 422     raise ValueError(f"Unknown model {args.model}")
 423
 424 ######################################################################
 425
 426 try:
 427     os.mkdir(args.result_dir)
 428 except FileExistsError:
 429     if not args.continue_training:
 430         print(f"result directory {args.result_dir} already exists")
 431         exit(1)
 432
 433 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 434
 435 if args.seed >= 0:
 436     # torch.backends.cudnn.deterministic = True
 437     # torch.backends.cudnn.benchmark = False
 438     # torch.use_deterministic_algorithms(True)
 439     torch.manual_seed(args.seed)
 440     if torch.cuda.is_available():
 441         torch.cuda.manual_seed_all(args.seed)
 442
 443 ######################################################################
 444
 445
 446 def log_string(s):
 447     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 448
 449     if log_file is not None:
 450         log_file.write(t + s + "\n")
 451         log_file.flush()
 452
 453     print(t + s)
 454     sys.stdout.flush()
 455
 456
 457 with os.popen("sha256sum *.py") as f:
 458     for l in f:
 459         log_string(f"sha256sum {l.strip()}")
 460
 461 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 462 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 463
 464 log_string(f"argv {' '.join(sys.argv)}")
 465
 466 for n in vars(args):
 467     log_string(f"args.{n} {getattr(args, n)}")
 468
 469
 470 ######################################################################
 471
 472
 473 def get_lr(n_epoch, it):
 474     if args.legacy_lr_schedule:
 475         # my crude scheduling to compare to previous baseline, added
 476         # warmup though
 477
 478         if it < args.nb_warmup_iter:
 479             return args.legacy_large_lr * it / args.nb_warmup_iter
 480         elif n_epoch < args.legacy_nb_epoch_large_lr:
 481             return args.legacy_large_lr
 482         else:
 483             return args.legacy_small_lr
 484
 485     # from nanoGPT
 486
 487     # 1) linear warmup for warmup_iter steps
 488     if it < args.nb_warmup_iter:
 489         return args.learning_rate * it / args.nb_warmup_iter
 490     # 2) if it > nb_decay_iter, return min learning rate
 491     if it > args.nb_decay_iter:
 492         return args.min_learning_rate
 493     # 3) in between, use cosine decay down to min learning rate
 494     decay_ratio = (it - args.nb_warmup_iter) / (
 495         args.nb_decay_iter - args.nb_warmup_iter
 496     )
 497     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 498     return args.min_learning_rate + coeff * (
 499         args.learning_rate - args.min_learning_rate
 500     )
 501
 502
 503 ######################################################################
 504
 505
 506 def picoclvr_pruner_horizontal_green(p):
 507     return not ("green" in p and ("left" in p or "right" in p))
 508
 509
 510 picoclvr_pruner_train = (
 511     picoclvr_pruner_horizontal_green
 512     if args.picocvlr_prune_properties in {"train+eval"}
 513     else None
 514 )
 515
 516 picoclvr_pruner_eval = (
 517     (lambda p: not picoclvr_pruner_horizontal_green(p))
 518     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 519     else None
 520 )
 521
 522 ######################################################################
 523
 524 device_data = device
 525
 526 if args.task == "byheart":
 527     task = tasks.SandBox(
 528         problem=problems.ProblemByHeart(),
 529         nb_train_samples=args.nb_train_samples,
 530         nb_test_samples=args.nb_test_samples,
 531         batch_size=args.batch_size,
 532         logger=log_string,
 533         device=device_data,
 534     )
 535     args.max_percents_of_test_in_train = -1
 536
 537 elif args.task == "learnop":
 538     task = tasks.SandBox(
 539         problem=problems.ProblemLearnOperator(),
 540         nb_train_samples=args.nb_train_samples,
 541         nb_test_samples=args.nb_test_samples,
 542         batch_size=args.batch_size,
 543         logger=log_string,
 544         device=device_data,
 545     )
 546
 547
 548 elif args.task == "guessop":
 549     task = tasks.SandBox(
 550         problem=problems.ProblemGuessOperator(),
 551         nb_train_samples=args.nb_train_samples,
 552         nb_test_samples=args.nb_test_samples,
 553         batch_size=args.batch_size,
 554         logger=log_string,
 555         device=device_data,
 556     )
 557
 558
 559 elif args.task == "twotargets":
 560     task = tasks.SandBox(
 561         problem=problems.ProblemTwoTargets(),
 562         nb_train_samples=args.nb_train_samples,
 563         nb_test_samples=args.nb_test_samples,
 564         batch_size=args.batch_size,
 565         logger=log_string,
 566         device=device_data,
 567     )
 568
 569 elif args.task == "memory":
 570     task = tasks.SandBox(
 571         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 572         nb_train_samples=args.nb_train_samples,
 573         nb_test_samples=args.nb_test_samples,
 574         batch_size=args.batch_size,
 575         logger=log_string,
 576         device=device_data,
 577     )
 578
 579 elif args.task == "mixing":
 580     task = tasks.SandBox(
 581         problem=problems.ProblemMixing(
 582             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 583         ),
 584         nb_train_samples=args.nb_train_samples,
 585         nb_test_samples=args.nb_test_samples,
 586         batch_size=args.batch_size,
 587         logger=log_string,
 588         device=device_data,
 589     )
 590
 591 elif args.task == "addition":
 592     task = tasks.SandBox(
 593         problem=problems.ProblemAddition(),
 594         nb_train_samples=args.nb_train_samples,
 595         nb_test_samples=args.nb_test_samples,
 596         batch_size=args.batch_size,
 597         logger=log_string,
 598         device=device_data,
 599     )
 600
 601 elif args.task == "picoclvr":
 602     task = tasks.PicoCLVR(
 603         nb_train_samples=args.nb_train_samples,
 604         nb_test_samples=args.nb_test_samples,
 605         batch_size=args.batch_size,
 606         height=args.picoclvr_height,
 607         width=args.picoclvr_width,
 608         nb_colors=args.picoclvr_nb_colors,
 609         logger=log_string,
 610         device=device_data,
 611         pruner_train=picoclvr_pruner_train,
 612         pruner_eval=picoclvr_pruner_eval,
 613     )
 614
 615 elif args.task == "mnist":
 616     task = tasks.MNIST(
 617         nb_train_samples=args.nb_train_samples,
 618         nb_test_samples=args.nb_test_samples,
 619         batch_size=args.batch_size,
 620         device=device_data,
 621     )
 622
 623 elif args.task == "maze":
 624     task = tasks.Maze(
 625         nb_train_samples=args.nb_train_samples,
 626         nb_test_samples=args.nb_test_samples,
 627         batch_size=args.batch_size,
 628         height=args.maze_height,
 629         width=args.maze_width,
 630         nb_walls=args.maze_nb_walls,
 631         device=device_data,
 632     )
 633
 634 elif args.task == "snake":
 635     task = tasks.Snake(
 636         nb_train_samples=args.nb_train_samples,
 637         nb_test_samples=args.nb_test_samples,
 638         batch_size=args.batch_size,
 639         height=args.snake_height,
 640         width=args.snake_width,
 641         nb_colors=args.snake_nb_colors,
 642         length=args.snake_length,
 643         prompt_length=args.snake_length // 2,
 644         device=device_data,
 645     )
 646
 647 elif args.task == "stack":
 648     task = tasks.Stack(
 649         nb_train_samples=args.nb_train_samples,
 650         nb_test_samples=args.nb_test_samples,
 651         batch_size=args.batch_size,
 652         logger=log_string,
 653         nb_steps=args.stack_nb_steps,
 654         nb_stacks=args.stack_nb_stacks,
 655         nb_digits=args.stack_nb_digits,
 656         fraction_values_for_train=args.stack_fraction_values_for_train,
 657         device=device_data,
 658     )
 659
 660 elif args.task == "expr":
 661     task = tasks.Expr(
 662         nb_train_samples=args.nb_train_samples,
 663         nb_test_samples=args.nb_test_samples,
 664         nb_variables=args.expr_nb_variables,
 665         sequence_length=args.expr_sequence_length,
 666         operand_max=args.expr_operand_max,
 667         result_max=args.expr_result_max,
 668         batch_size=args.batch_size,
 669         device=device_data,
 670     )
 671
 672 elif args.task == "rpl":
 673     task = tasks.RPL(
 674         nb_train_samples=args.nb_train_samples,
 675         nb_test_samples=args.nb_test_samples,
 676         batch_size=args.batch_size,
 677         nb_starting_values=args.rpl_nb_starting_values,
 678         max_input=args.rpl_max_input,
 679         prog_len=args.rpl_prog_len,
 680         nb_runs=args.rpl_nb_runs,
 681         no_prog=args.rpl_no_prog,
 682         logger=log_string,
 683         device=device_data,
 684     )
 685
 686 elif args.task == "grid":
 687     task = tasks.Grid(
 688         nb_train_samples=args.nb_train_samples,
 689         nb_test_samples=args.nb_test_samples,
 690         batch_size=args.batch_size,
 691         size=args.grid_size,
 692         logger=log_string,
 693         device=device_data,
 694     )
 695
 696 elif args.task == "qmlp":
 697     task = tasks.QMLP(
 698         nb_train_samples=args.nb_train_samples,
 699         nb_test_samples=args.nb_test_samples,
 700         batch_size=args.batch_size,
 701         result_dir=args.result_dir,
 702         logger=log_string,
 703         device=device_data,
 704     )
 705
 706 else:
 707     raise ValueError(f"Unknown task {args.task}")
 708
 709 ######################################################################
 710
 711 log_string(f"device {device}")
 712
 713 vocabulary_size = task.vocabulary_size()
 714
 715 log_string(f"vocabulary_size {vocabulary_size}")
 716
 717 ##############################
 718
 719 model = mygpt.MyGPT(
 720     vocabulary_size=vocabulary_size,
 721     dim_model=args.dim_model,
 722     dim_keys=args.dim_keys,
 723     dim_hidden=args.dim_hidden,
 724     nb_heads=args.nb_heads,
 725     nb_lines=args.nb_lines,
 726     caterpillar_height=args.caterpillar_height,
 727     nb_blocks=args.nb_blocks,
 728     causal=True,
 729     dropout=args.dropout,
 730     attention_layer=args.attention,
 731 )
 732
 733 model.to(device)
 734
 735 nb_parameters = sum(p.numel() for p in model.parameters())
 736 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 737
 738 ######################################################################
 739
 740 nb_epochs_finished = 0
 741
 742 if args.no_checkpoint:
 743     log_string(f"not trying to load checkpoint.")
 744
 745 else:
 746     try:
 747         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 748         checkpoint = torch.load(checkpoint_name)
 749         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 750         model.load_state_dict(checkpoint["model_state"])
 751         torch.set_rng_state(checkpoint["rng_state"])
 752         if torch.cuda.is_available():
 753             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 754
 755         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 756
 757     except FileNotFoundError:
 758         log_string("starting from scratch.")
 759
 760     except:
 761         log_string("error when loading the checkpoint.")
 762         exit(1)
 763
 764 ######################################################################
 765
 766 if args.task == "expr" and args.expr_input_file is not None:
 767     task.produce_results(
 768         n_epoch=nb_epochs_finished,
 769         model=model,
 770         result_dir=args.result_dir,
 771         logger=log_string,
 772         deterministic_synthesis=args.deterministic_synthesis,
 773         input_file=args.expr_input_file,
 774     )
 775
 776     exit(0)
 777
 778 ######################################################################
 779
 780 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 781
 782 # Compute the entropy of the training tokens
 783
 784 token_count = 0
 785 for input in task.batches(split="train"):
 786     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 787 token_probas = token_count / token_count.sum()
 788 entropy = -torch.xlogy(token_probas, token_probas).sum()
 789 train_set_perplexity = math.exp(entropy)
 790
 791 ######################################################################
 792 # A bit of paranoia never hurts
 793
 794 if args.max_percents_of_test_in_train >= 0:
 795
 796     def subsets_as_tuples(batches, cs):
 797         s = set()
 798         for batch in batches:
 799             for x in batch:
 800                 s.add(tuple([v.item() for v in x]))
 801                 if len(s) == cs:
 802                     yield s
 803                     s = set()
 804         yield s
 805
 806     nb_test, nb_in_train = 0, 0
 807     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 808         in_train = set()
 809         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 810             in_train.update(test_subset.intersection(train_subset))
 811         nb_in_train += len(in_train)
 812         nb_test += len(test_subset)
 813
 814     log_string(
 815         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 816     )
 817
 818     assert (
 819         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 820     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 821
 822 ##############################
 823
 824 nb_samples_seen = 0
 825
 826 if nb_epochs_finished >= nb_epochs:
 827     task.produce_results(
 828         n_epoch=nb_epochs_finished,
 829         model=model,
 830         result_dir=args.result_dir,
 831         logger=log_string,
 832         deterministic_synthesis=args.deterministic_synthesis,
 833     )
 834
 835 time_pred_result = None
 836
 837 it = 0
 838
 839 for n_epoch in range(nb_epochs_finished, nb_epochs):
 840     if args.optim == "sgd":
 841         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 842     elif args.optim == "adam":
 843         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 844     elif args.optim == "adamw":
 845         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 846     else:
 847         raise ValueError(f"Unknown optimizer {args.optim}.")
 848
 849     model.train()
 850
 851     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 852
 853     for input in task.batches(split="train"):
 854         model.reset_inner_loss()
 855         input = input.to(device)
 856
 857         output = model(mygpt.BracketedSequence(input)).x
 858         loss = F.cross_entropy(output.transpose(1, 2), input)
 859         inner_loss = model.get_inner_loss()
 860
 861         acc_train_loss += loss.item() * input.size(0)
 862         acc_train_inner_loss += inner_loss.item() * input.size(0)
 863
 864         nb_train_samples += input.size(0)
 865         nb_samples_seen += input.size(0)
 866
 867         total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
 868
 869         it += 1
 870         lr = get_lr(n_epoch, it)
 871         for param_group in optimizer.param_groups:
 872             param_group["lr"] = lr
 873
 874         # log_string(f"learning_rate {lr}")
 875
 876         optimizer.zero_grad()
 877         total_loss.backward()
 878         optimizer.step()
 879
 880     with torch.autograd.no_grad():
 881         model.eval()
 882
 883         nb_test_samples, acc_test_loss = 0, 0.0
 884
 885         for input in task.batches(split="test"):
 886             input = input.to(device)
 887
 888             output = model(mygpt.BracketedSequence(input)).x
 889             loss = F.cross_entropy(output.transpose(1, 2), input)
 890             acc_test_loss += loss.item() * input.size(0)
 891             nb_test_samples += input.size(0)
 892
 893         log_string(
 894             f"loss {n_epoch} train_loss {acc_train_loss/nb_train_samples} train_inner_loss {acc_train_inner_loss/nb_train_samples} test_prediction {acc_test_loss/nb_test_samples}"
 895         )
 896
 897         task.produce_results(
 898             n_epoch=n_epoch,
 899             model=model,
 900             result_dir=args.result_dir,
 901             logger=log_string,
 902             deterministic_synthesis=args.deterministic_synthesis,
 903         )
 904
 905         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 906         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 907
 908         log_string(
 909             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 910         )
 911
 912         time_current_result = datetime.datetime.now()
 913         if time_pred_result is not None:
 914             log_string(
 915                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 916             )
 917         time_pred_result = time_current_result
 918
 919     checkpoint = {
 920         "nb_epochs_finished": n_epoch + 1,
 921         "model_state": model.state_dict(),
 922         "rng_state": torch.get_rng_state(),
 923     }
 924
 925     if torch.cuda.is_available():
 926         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 927
 928     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 929     torch.save(checkpoint, checkpoint_name)
 930     log_string(f"saved checkpoint {checkpoint_name}")
 931
 932 ######################################################################