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
  20 def str2bool(x):
  21     x = x.lower()
  22     if x in {"1", "true", "yes"}:
  23         return True
  24     elif x in {"0", "false", "no"}:
  25         return False
  26     else:
  27         raise ValueError
  28
  29
  30 parser = argparse.ArgumentParser(
  31     description="An implementation of GPT with cache.",
  32     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  33 )
  34
  35 parser.add_argument(
  36     "--task",
  37     type=str,
  38     default="twotargets",
  39     help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
  40 )
  41
  42 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  43
  44 parser.add_argument("--result_dir", type=str, default=None)
  45
  46 parser.add_argument("--seed", type=int, default=0)
  47
  48 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  49
  50 parser.add_argument("--force_cpu", type=str2bool, default=False)
  51
  52 ########################################
  53
  54 parser.add_argument("--nb_epochs", type=int, default=50)
  55
  56 parser.add_argument("--batch_size", type=int, default=None)
  57
  58 parser.add_argument("--nb_train_samples", type=int, default=None)
  59
  60 parser.add_argument("--nb_test_samples", type=int, default=None)
  61
  62 parser.add_argument("--optim", type=str, default="adam")
  63
  64 ########################################
  65
  66 parser.add_argument("--nb_warmup_iter", type=int, default=100)
  67
  68 parser.add_argument("--nb_decay_iter", type=int, default=5000)
  69
  70 parser.add_argument("--learning_rate", type=float, default=6e-4)
  71
  72 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
  73
  74 # legacy
  75
  76 parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
  77
  78 parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
  79
  80 parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
  81
  82 parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
  83
  84 ########################################
  85
  86 parser.add_argument("--model", type=str, default=None)
  87
  88 parser.add_argument("--attention", type=str, default=None)
  89
  90 parser.add_argument("--dim_model", type=int, default=None)
  91
  92 parser.add_argument("--dim_keys", type=int, default=None)
  93
  94 parser.add_argument("--dim_hidden", type=int, default=None)
  95
  96 parser.add_argument("--nb_heads", type=int, default=None)
  97
  98 parser.add_argument("--nb_lines", type=int, default=None)
  99
 100 parser.add_argument("--caterpillar_height", type=int, default=None)
 101
 102 parser.add_argument("--rho", type=float, default=0.0)
 103
 104 parser.add_argument("--nb_blocks", type=int, default=None)
 105
 106 parser.add_argument("--dropout", type=float, default=0.1)
 107
 108 ########################################
 109
 110 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 111
 112 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 113
 114 parser.add_argument("--continue_training", action="store_true", default=False)
 115
 116 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 117
 118 ##############################
 119 # rpl options
 120
 121 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 122
 123 parser.add_argument("--rpl_max_input", type=int, default=9)
 124
 125 parser.add_argument("--rpl_prog_len", type=int, default=8)
 126
 127 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 128
 129 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 130
 131 ##############################
 132 # grid options
 133
 134 parser.add_argument("--grid_size", type=int, default=6)
 135
 136 ##############################
 137 # picoclvr options
 138
 139 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 140
 141 parser.add_argument("--picoclvr_height", type=int, default=12)
 142
 143 parser.add_argument("--picoclvr_width", type=int, default=16)
 144
 145 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 146
 147 ##############################
 148 # Maze options
 149
 150 parser.add_argument("--maze_height", type=int, default=13)
 151
 152 parser.add_argument("--maze_width", type=int, default=21)
 153
 154 parser.add_argument("--maze_nb_walls", type=int, default=15)
 155
 156 ##############################
 157 # Snake options
 158
 159 parser.add_argument("--snake_height", type=int, default=9)
 160
 161 parser.add_argument("--snake_width", type=int, default=12)
 162
 163 parser.add_argument("--snake_nb_colors", type=int, default=5)
 164
 165 parser.add_argument("--snake_length", type=int, default=200)
 166
 167 ##############################
 168 # Stack options
 169
 170 parser.add_argument("--stack_nb_steps", type=int, default=100)
 171
 172 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 173
 174 parser.add_argument("--stack_nb_digits", type=int, default=3)
 175
 176 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 177
 178 ##############################
 179 # Expr options
 180
 181 parser.add_argument("--expr_nb_variables", type=int, default=5)
 182
 183 parser.add_argument("--expr_sequence_length", type=int, default=40)
 184
 185 parser.add_argument("--expr_operand_max", type=int, default=9)
 186
 187 parser.add_argument("--expr_result_max", type=int, default=99)
 188
 189 parser.add_argument("--expr_input_file", type=str, default=None)
 190
 191 ##############################
 192 # Memory
 193
 194 parser.add_argument("--memory_len_total", type=int, default=32)
 195
 196 ##############################
 197 # Mixing
 198
 199 parser.add_argument("--mixing_hard", action="store_true", default=False)
 200
 201 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 202
 203 ######################################################################
 204
 205 args = parser.parse_args()
 206
 207 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 208
 209 if args.result_dir is None:
 210     args.result_dir = f"results_{args.task}_{args.model}"
 211
 212 ######################################################################
 213
 214 if not args.force_cpu and torch.cuda.is_available():
 215     device = torch.device("cuda")
 216     torch.backends.cuda.matmul.allow_tf32 = True
 217 else:
 218     device = torch.device("cpu")
 219
 220 ######################################################################
 221
 222 default_task_args = {
 223     "addition": {
 224         "model": "352M",
 225         "batch_size": 25,
 226         "nb_train_samples": 250000,
 227         "nb_test_samples": 10000,
 228     },
 229     "byheart": {
 230         "model": "37M",
 231         "batch_size": 25,
 232         "nb_train_samples": 50000,
 233         "nb_test_samples": 10000,
 234     },
 235     "expr": {
 236         "model": "352M",
 237         "batch_size": 25,
 238         "nb_train_samples": 2500000,
 239         "nb_test_samples": 10000,
 240     },
 241     "grid": {
 242         "model": "37M",
 243         "batch_size": 25,
 244         "nb_train_samples": 250000,
 245         "nb_test_samples": 10000,
 246     },
 247     "qmlp": {
 248         "model": "37M",
 249         "batch_size": 10,
 250         "nb_train_samples": 100000,
 251         "nb_test_samples": 1000,
 252     },
 253     "guessop": {
 254         "model": "352M",
 255         "batch_size": 25,
 256         "nb_train_samples": 1000000,
 257         "nb_test_samples": 10000,
 258     },
 259     "learnop": {
 260         "model": "37M",
 261         "batch_size": 25,
 262         "nb_train_samples": 50000,
 263         "nb_test_samples": 10000,
 264     },
 265     "maze": {
 266         "model": "37M",
 267         "batch_size": 5,
 268         "nb_train_samples": 100000,
 269         "nb_test_samples": 10000,
 270     },
 271     "picoclvr": {
 272         "model": "37M",
 273         "batch_size": 25,
 274         "nb_train_samples": 250000,
 275         "nb_test_samples": 10000,
 276     },
 277     "rpl": {
 278         "model": "352M",
 279         "batch_size": 5,
 280         "nb_train_samples": 2500000,
 281         "nb_test_samples": 10000,
 282     },
 283     "snake": {
 284         "model": "37M",
 285         "batch_size": 25,
 286         "nb_train_samples": 250000,
 287         "nb_test_samples": 10000,
 288     },
 289     "stack": {
 290         "model": "37M",
 291         "batch_size": 25,
 292         "nb_train_samples": 100000,
 293         "nb_test_samples": 1000,
 294     },
 295     "twotargets": {
 296         "model": "37M",
 297         "batch_size": 25,
 298         "nb_train_samples": 50000,
 299         "nb_test_samples": 10000,
 300     },
 301     "memory": {
 302         "model": "37M",
 303         "batch_size": 25,
 304         "nb_train_samples": 25000,
 305         "nb_test_samples": 10000,
 306     },
 307     "mixing": {
 308         "model": "37M",
 309         "batch_size": 25,
 310         "nb_train_samples": 250000,
 311         "nb_test_samples": 10000,
 312     },
 313     "mnist": {
 314         "model": "37M",
 315         "batch_size": 10,
 316         "nb_train_samples": 60000,
 317         "nb_test_samples": 10000,
 318     },
 319 }
 320
 321 if args.task in default_task_args:
 322     for k, v in default_task_args[args.task].items():
 323         if getattr(args, k) is None:
 324             setattr(args, k, v)
 325
 326 ######################################################################
 327
 328 default_model_args = {
 329     "17K": {
 330         "attention": "mha",
 331         "dim_model": 32,
 332         "dim_keys": 32,
 333         "dim_hidden": 32,
 334         "nb_heads": 2,
 335         "nb_blocks": 2,
 336     },
 337     "17K-C": {
 338         "attention": "caterpillar",
 339         "dim_model": 32,
 340         "dim_keys": 32,
 341         "dim_hidden": 32,
 342         "nb_heads": 2,
 343         "nb_lines": 16,
 344         "caterpillar_height": 4,
 345         "nb_blocks": 2,
 346     },
 347     "4M": {
 348         "attention": "mha",
 349         "dim_model": 256,
 350         "dim_keys": 32,
 351         "dim_hidden": 1024,
 352         "nb_heads": 4,
 353         "nb_blocks": 6,
 354     },
 355     "4M-C": {
 356         "attention": "caterpillar",
 357         "dim_model": 256,
 358         "dim_keys": 32,
 359         "dim_hidden": 1024,
 360         "nb_heads": 4,
 361         "nb_lines": 32,
 362         "caterpillar_height": 4,
 363         "nb_blocks": 6,
 364     },
 365     "37M": {
 366         "attention": "mha",
 367         "dim_model": 512,
 368         "dim_keys": 64,
 369         "dim_hidden": 2048,
 370         "nb_heads": 8,
 371         "nb_blocks": 12,
 372     },
 373     "37M-C": {
 374         "attention": "caterpillar",
 375         "dim_model": 512,
 376         "dim_keys": 64,
 377         "dim_hidden": 2048,
 378         "nb_heads": 8,
 379         "nb_lines": 256,
 380         "caterpillar_height": 32,
 381         "nb_blocks": 12,
 382     },
 383     "122M": {
 384         "attention": "mha",
 385         "dim_model": 768,
 386         "dim_keys": 64,
 387         "dim_hidden": 2048,
 388         "nb_heads": 8,
 389         "nb_blocks": 24,
 390     },
 391     "122M-C": {
 392         "attention": "caterpillar",
 393         "dim_model": 768,
 394         "dim_keys": 64,
 395         "dim_hidden": 2048,
 396         "nb_heads": 8,
 397         "nb_lines": 128,
 398         "nb_blocks": 24,
 399     },
 400     "352M": {
 401         "attention": "mha",
 402         "dim_model": 1024,
 403         "dim_keys": 64,
 404         "dim_hidden": 2048,
 405         "nb_heads": 8,
 406         "nb_blocks": 48,
 407     },
 408     "352M-C": {
 409         "attention": "caterpillar",
 410         "dim_model": 1024,
 411         "dim_keys": 64,
 412         "dim_hidden": 2048,
 413         "nb_heads": 8,
 414         "nb_lines": 128,
 415         "nb_blocks": 48,
 416     },
 417 }
 418
 419 if args.model in default_model_args:
 420     for k, v in default_model_args[args.model].items():
 421         if getattr(args, k) is None:
 422             setattr(args, k, v)
 423 else:
 424     raise ValueError(f"Unknown model {args.model}")
 425
 426 ######################################################################
 427
 428 try:
 429     os.mkdir(args.result_dir)
 430 except FileExistsError:
 431     if not args.continue_training:
 432         print(f"result directory {args.result_dir} already exists")
 433         exit(1)
 434
 435 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 436
 437 if args.seed >= 0:
 438     # torch.backends.cudnn.deterministic = True
 439     # torch.backends.cudnn.benchmark = False
 440     # torch.use_deterministic_algorithms(True)
 441     torch.manual_seed(args.seed)
 442     if torch.cuda.is_available():
 443         torch.cuda.manual_seed_all(args.seed)
 444
 445 ######################################################################
 446
 447
 448 def log_string(s):
 449     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 450
 451     if log_file is not None:
 452         log_file.write(t + s + "\n")
 453         log_file.flush()
 454
 455     print(t + s)
 456     sys.stdout.flush()
 457
 458
 459 with os.popen("sha256sum *.py") as f:
 460     for l in f:
 461         log_string(f"sha256sum {l.strip()}")
 462
 463 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 464 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 465
 466 log_string(f"argv {' '.join(sys.argv)}")
 467
 468 for n in vars(args):
 469     log_string(f"args.{n} {getattr(args, n)}")
 470
 471
 472 ######################################################################
 473
 474
 475 def get_lr(n_epoch, it):
 476     if args.legacy_lr_schedule:
 477         # my crude scheduling to compare to previous baseline, added
 478         # warmup though
 479
 480         if it < args.nb_warmup_iter:
 481             return args.legacy_large_lr * it / args.nb_warmup_iter
 482         elif n_epoch < args.legacy_nb_epoch_large_lr:
 483             return args.legacy_large_lr
 484         else:
 485             return args.legacy_small_lr
 486
 487     # from nanoGPT
 488
 489     # 1) linear warmup for warmup_iter steps
 490     if it < args.nb_warmup_iter:
 491         return args.learning_rate * it / args.nb_warmup_iter
 492     # 2) if it > nb_decay_iter, return min learning rate
 493     if it > args.nb_decay_iter:
 494         return args.min_learning_rate
 495     # 3) in between, use cosine decay down to min learning rate
 496     decay_ratio = (it - args.nb_warmup_iter) / (
 497         args.nb_decay_iter - args.nb_warmup_iter
 498     )
 499     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 500     return args.min_learning_rate + coeff * (
 501         args.learning_rate - args.min_learning_rate
 502     )
 503
 504
 505 ######################################################################
 506
 507
 508 def picoclvr_pruner_horizontal_green(p):
 509     return not ("green" in p and ("left" in p or "right" in p))
 510
 511
 512 picoclvr_pruner_train = (
 513     picoclvr_pruner_horizontal_green
 514     if args.picocvlr_prune_properties in {"train+eval"}
 515     else None
 516 )
 517
 518 picoclvr_pruner_eval = (
 519     (lambda p: not picoclvr_pruner_horizontal_green(p))
 520     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 521     else None
 522 )
 523
 524 ######################################################################
 525
 526 device_data = device
 527
 528 if args.task == "byheart":
 529     task = tasks.SandBox(
 530         problem=problems.ProblemByHeart(),
 531         nb_train_samples=args.nb_train_samples,
 532         nb_test_samples=args.nb_test_samples,
 533         batch_size=args.batch_size,
 534         logger=log_string,
 535         device=device_data,
 536     )
 537     args.max_percents_of_test_in_train = -1
 538
 539 elif args.task == "learnop":
 540     task = tasks.SandBox(
 541         problem=problems.ProblemLearnOperator(),
 542         nb_train_samples=args.nb_train_samples,
 543         nb_test_samples=args.nb_test_samples,
 544         batch_size=args.batch_size,
 545         logger=log_string,
 546         device=device_data,
 547     )
 548
 549
 550 elif args.task == "guessop":
 551     task = tasks.SandBox(
 552         problem=problems.ProblemGuessOperator(),
 553         nb_train_samples=args.nb_train_samples,
 554         nb_test_samples=args.nb_test_samples,
 555         batch_size=args.batch_size,
 556         logger=log_string,
 557         device=device_data,
 558     )
 559
 560
 561 elif args.task == "twotargets":
 562     task = tasks.SandBox(
 563         problem=problems.ProblemTwoTargets(),
 564         nb_train_samples=args.nb_train_samples,
 565         nb_test_samples=args.nb_test_samples,
 566         batch_size=args.batch_size,
 567         logger=log_string,
 568         device=device_data,
 569     )
 570
 571 elif args.task == "memory":
 572     task = tasks.SandBox(
 573         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 574         nb_train_samples=args.nb_train_samples,
 575         nb_test_samples=args.nb_test_samples,
 576         batch_size=args.batch_size,
 577         logger=log_string,
 578         device=device_data,
 579     )
 580
 581 elif args.task == "mixing":
 582     task = tasks.SandBox(
 583         problem=problems.ProblemMixing(
 584             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 585         ),
 586         nb_train_samples=args.nb_train_samples,
 587         nb_test_samples=args.nb_test_samples,
 588         batch_size=args.batch_size,
 589         logger=log_string,
 590         device=device_data,
 591     )
 592
 593 elif args.task == "addition":
 594     task = tasks.SandBox(
 595         problem=problems.ProblemAddition(),
 596         nb_train_samples=args.nb_train_samples,
 597         nb_test_samples=args.nb_test_samples,
 598         batch_size=args.batch_size,
 599         logger=log_string,
 600         device=device_data,
 601     )
 602
 603 elif args.task == "picoclvr":
 604     task = tasks.PicoCLVR(
 605         nb_train_samples=args.nb_train_samples,
 606         nb_test_samples=args.nb_test_samples,
 607         batch_size=args.batch_size,
 608         height=args.picoclvr_height,
 609         width=args.picoclvr_width,
 610         nb_colors=args.picoclvr_nb_colors,
 611         logger=log_string,
 612         device=device_data,
 613         pruner_train=picoclvr_pruner_train,
 614         pruner_eval=picoclvr_pruner_eval,
 615     )
 616
 617 elif args.task == "mnist":
 618     task = tasks.MNIST(
 619         nb_train_samples=args.nb_train_samples,
 620         nb_test_samples=args.nb_test_samples,
 621         batch_size=args.batch_size,
 622         device=device_data,
 623     )
 624
 625 elif args.task == "maze":
 626     task = tasks.Maze(
 627         nb_train_samples=args.nb_train_samples,
 628         nb_test_samples=args.nb_test_samples,
 629         batch_size=args.batch_size,
 630         height=args.maze_height,
 631         width=args.maze_width,
 632         nb_walls=args.maze_nb_walls,
 633         device=device_data,
 634     )
 635
 636 elif args.task == "snake":
 637     task = tasks.Snake(
 638         nb_train_samples=args.nb_train_samples,
 639         nb_test_samples=args.nb_test_samples,
 640         batch_size=args.batch_size,
 641         height=args.snake_height,
 642         width=args.snake_width,
 643         nb_colors=args.snake_nb_colors,
 644         length=args.snake_length,
 645         prompt_length=args.snake_length // 2,
 646         device=device_data,
 647     )
 648
 649 elif args.task == "stack":
 650     task = tasks.Stack(
 651         nb_train_samples=args.nb_train_samples,
 652         nb_test_samples=args.nb_test_samples,
 653         batch_size=args.batch_size,
 654         logger=log_string,
 655         nb_steps=args.stack_nb_steps,
 656         nb_stacks=args.stack_nb_stacks,
 657         nb_digits=args.stack_nb_digits,
 658         fraction_values_for_train=args.stack_fraction_values_for_train,
 659         device=device_data,
 660     )
 661
 662 elif args.task == "expr":
 663     task = tasks.Expr(
 664         nb_train_samples=args.nb_train_samples,
 665         nb_test_samples=args.nb_test_samples,
 666         nb_variables=args.expr_nb_variables,
 667         sequence_length=args.expr_sequence_length,
 668         operand_max=args.expr_operand_max,
 669         result_max=args.expr_result_max,
 670         batch_size=args.batch_size,
 671         device=device_data,
 672     )
 673
 674 elif args.task == "rpl":
 675     task = tasks.RPL(
 676         nb_train_samples=args.nb_train_samples,
 677         nb_test_samples=args.nb_test_samples,
 678         batch_size=args.batch_size,
 679         nb_starting_values=args.rpl_nb_starting_values,
 680         max_input=args.rpl_max_input,
 681         prog_len=args.rpl_prog_len,
 682         nb_runs=args.rpl_nb_runs,
 683         no_prog=args.rpl_no_prog,
 684         logger=log_string,
 685         device=device_data,
 686     )
 687
 688 elif args.task == "grid":
 689     task = tasks.Grid(
 690         nb_train_samples=args.nb_train_samples,
 691         nb_test_samples=args.nb_test_samples,
 692         batch_size=args.batch_size,
 693         size=args.grid_size,
 694         logger=log_string,
 695         device=device_data,
 696     )
 697
 698 elif args.task == "qmlp":
 699     task = tasks.QMLP(
 700         nb_train_samples=args.nb_train_samples,
 701         nb_test_samples=args.nb_test_samples,
 702         batch_size=args.batch_size,
 703         result_dir=args.result_dir,
 704         logger=log_string,
 705         device=device_data,
 706     )
 707
 708 else:
 709     raise ValueError(f"Unknown task {args.task}")
 710
 711 ######################################################################
 712
 713 log_string(f"device {device}")
 714
 715 vocabulary_size = task.vocabulary_size()
 716
 717 log_string(f"vocabulary_size {vocabulary_size}")
 718
 719 ##############################
 720
 721 model = mygpt.MyGPT(
 722     vocabulary_size=vocabulary_size,
 723     dim_model=args.dim_model,
 724     dim_keys=args.dim_keys,
 725     dim_hidden=args.dim_hidden,
 726     nb_heads=args.nb_heads,
 727     nb_lines=args.nb_lines,
 728     caterpillar_height=args.caterpillar_height,
 729     nb_blocks=args.nb_blocks,
 730     causal=True,
 731     dropout=args.dropout,
 732     attention_layer=args.attention,
 733 )
 734
 735 model.to(device)
 736
 737 nb_parameters = sum(p.numel() for p in model.parameters())
 738 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 739
 740 ######################################################################
 741
 742 nb_epochs_finished = 0
 743
 744 if args.no_checkpoint:
 745     log_string(f"not trying to load checkpoint.")
 746
 747 else:
 748     try:
 749         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 750         checkpoint = torch.load(checkpoint_name)
 751         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 752         model.load_state_dict(checkpoint["model_state"])
 753         torch.set_rng_state(checkpoint["rng_state"])
 754         if torch.cuda.is_available():
 755             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 756
 757         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 758
 759     except FileNotFoundError:
 760         log_string("starting from scratch.")
 761
 762     except:
 763         log_string("error when loading the checkpoint.")
 764         exit(1)
 765
 766 ######################################################################
 767
 768 if args.task == "expr" and args.expr_input_file is not None:
 769     task.produce_results(
 770         n_epoch=nb_epochs_finished,
 771         model=model,
 772         result_dir=args.result_dir,
 773         logger=log_string,
 774         deterministic_synthesis=args.deterministic_synthesis,
 775         input_file=args.expr_input_file,
 776     )
 777
 778     exit(0)
 779
 780 ######################################################################
 781
 782 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 783
 784 # Compute the entropy of the training tokens
 785
 786 token_count = 0
 787 for input in task.batches(split="train"):
 788     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 789 token_probas = token_count / token_count.sum()
 790 entropy = -torch.xlogy(token_probas, token_probas).sum()
 791 train_set_perplexity = math.exp(entropy)
 792
 793 ######################################################################
 794 # A bit of paranoia never hurts
 795
 796 if args.max_percents_of_test_in_train >= 0:
 797
 798     def subsets_as_tuples(batches, cs):
 799         s = set()
 800         for batch in batches:
 801             for x in batch:
 802                 s.add(tuple([v.item() for v in x]))
 803                 if len(s) == cs:
 804                     yield s
 805                     s = set()
 806         yield s
 807
 808     nb_test, nb_in_train = 0, 0
 809     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 810         in_train = set()
 811         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 812             in_train.update(test_subset.intersection(train_subset))
 813         nb_in_train += len(in_train)
 814         nb_test += len(test_subset)
 815
 816     log_string(
 817         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 818     )
 819
 820     assert (
 821         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 822     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 823
 824 ##############################
 825
 826 nb_samples_seen = 0
 827
 828 if nb_epochs_finished >= nb_epochs:
 829     task.produce_results(
 830         n_epoch=nb_epochs_finished,
 831         model=model,
 832         result_dir=args.result_dir,
 833         logger=log_string,
 834         deterministic_synthesis=args.deterministic_synthesis,
 835     )
 836
 837 time_pred_result = datetime.datetime.now()
 838
 839 it = 0
 840
 841 for n_epoch in range(nb_epochs_finished, nb_epochs):
 842     if args.optim == "sgd":
 843         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 844     elif args.optim == "adam":
 845         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 846     elif args.optim == "adamw":
 847         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 848     else:
 849         raise ValueError(f"Unknown optimizer {args.optim}.")
 850
 851     model.train()
 852
 853     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 854
 855     for input in task.batches(split="train"):
 856         model.reset_inner_loss()
 857         input = input.to(device)
 858
 859         output = model(mygpt.BracketedSequence(input)).x
 860         loss = F.cross_entropy(output.transpose(1, 2), input)
 861         inner_loss = model.get_inner_loss()
 862
 863         acc_train_loss += loss.item() * input.size(0)
 864         acc_train_inner_loss += inner_loss.item() * input.size(0)
 865
 866         nb_train_samples += input.size(0)
 867         nb_samples_seen += input.size(0)
 868
 869         total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
 870
 871         it += 1
 872         lr = get_lr(n_epoch, it)
 873         for param_group in optimizer.param_groups:
 874             param_group["lr"] = lr
 875
 876         # log_string(f"learning_rate {lr}")
 877
 878         optimizer.zero_grad()
 879         total_loss.backward()
 880         optimizer.step()
 881
 882     with torch.autograd.no_grad():
 883         model.eval()
 884
 885         nb_test_samples, acc_test_loss = 0, 0.0
 886
 887         for input in task.batches(split="test"):
 888             input = input.to(device)
 889
 890             output = model(mygpt.BracketedSequence(input)).x
 891             loss = F.cross_entropy(output.transpose(1, 2), input)
 892             acc_test_loss += loss.item() * input.size(0)
 893             nb_test_samples += input.size(0)
 894
 895         log_string(
 896             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}"
 897         )
 898
 899         task.produce_results(
 900             n_epoch=n_epoch,
 901             model=model,
 902             result_dir=args.result_dir,
 903             logger=log_string,
 904             deterministic_synthesis=args.deterministic_synthesis,
 905         )
 906
 907         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 908         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 909
 910         log_string(
 911             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 912         )
 913
 914         time_current_result = datetime.datetime.now()
 915         log_string(
 916             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 917         )
 918         time_pred_result = time_current_result
 919
 920     checkpoint = {
 921         "nb_epochs_finished": n_epoch + 1,
 922         "model_state": model.state_dict(),
 923         "rng_state": torch.get_rng_state(),
 924     }
 925
 926     if torch.cuda.is_available():
 927         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 928
 929     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 930     torch.save(checkpoint, checkpoint_name)
 931     log_string(f"saved checkpoint {checkpoint_name}")
 932
 933 ######################################################################