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