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         "attention": "mha",
 350         "dim_model": 512,
 351         "dim_keys": 64,
 352         "dim_hidden": 2048,
 353         "nb_heads": 8,
 354         "dim_rec_v": 64,
 355         "nb_blocks": 12,
 356     },
 357     "37M-C": {
 358         "attention": "caterpillar",
 359         "dim_model": 512,
 360         "dim_keys": 64,
 361         "dim_hidden": 2048,
 362         "nb_heads": 8,
 363         "nb_lines": 256,
 364         "caterpillar_height": 32,
 365         "dim_rec_v": 64,
 366         "nb_blocks": 12,
 367     },
 368     "122M": {
 369         "attention": "mha",
 370         "dim_model": 768,
 371         "dim_keys": 64,
 372         "dim_hidden": 2048,
 373         "nb_heads": 8,
 374         "dim_rec_v": 96,
 375         "nb_blocks": 24,
 376     },
 377     "122M-C": {
 378         "attention": "caterpillar",
 379         "dim_model": 768,
 380         "dim_keys": 64,
 381         "dim_hidden": 2048,
 382         "nb_heads": 8,
 383         "nb_lines": 128,
 384         "dim_rec_v": 96,
 385         "nb_blocks": 24,
 386     },
 387     "352M": {
 388         "attention": "mha",
 389         "dim_model": 1024,
 390         "dim_keys": 64,
 391         "dim_hidden": 2048,
 392         "nb_heads": 8,
 393         "dim_rec_v": 128,
 394         "nb_blocks": 48,
 395     },
 396     "352M-C": {
 397         "attention": "caterpillar",
 398         "dim_model": 1024,
 399         "dim_keys": 64,
 400         "dim_hidden": 2048,
 401         "nb_heads": 8,
 402         "nb_lines": 128,
 403         "dim_rec_v": 128,
 404         "nb_blocks": 48,
 405     },
 406 }
 407
 408 if args.model in default_model_args:
 409     for k, v in default_model_args[args.model].items():
 410         if getattr(args, k) is None:
 411             setattr(args, k, v)
 412 else:
 413     raise ValueError(f"Unknown model {args.model}")
 414
 415 ######################################################################
 416
 417 try:
 418     os.mkdir(args.result_dir)
 419 except FileExistsError:
 420     if not args.overwrite_results:
 421         print(f"result directory {args.result_dir} already exists")
 422         exit(1)
 423
 424 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 425
 426 if args.seed >= 0:
 427     # torch.backends.cudnn.deterministic = True
 428     # torch.backends.cudnn.benchmark = False
 429     # torch.use_deterministic_algorithms(True)
 430     torch.manual_seed(args.seed)
 431     if torch.cuda.is_available():
 432         torch.cuda.manual_seed_all(args.seed)
 433
 434 ######################################################################
 435
 436
 437 def log_string(s):
 438     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 439
 440     if log_file is not None:
 441         log_file.write(t + s + "\n")
 442         log_file.flush()
 443
 444     print(t + s)
 445     sys.stdout.flush()
 446
 447
 448 with os.popen("sha256sum *.py") as f:
 449     for l in f:
 450         log_string(f"sha256sum {l.strip()}")
 451
 452 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 453 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 454
 455 log_string(f"argv {' '.join(sys.argv)}")
 456
 457 for n in vars(args):
 458     log_string(f"args.{n} {getattr(args, n)}")
 459
 460
 461 ######################################################################
 462
 463 # from nanoGPT
 464
 465
 466 def get_lr(it):
 467     # 1) linear warmup for warmup_iter steps
 468     if it < args.nb_warmup_iter:
 469         return args.learning_rate * it / args.nb_warmup_iter
 470     # 2) if it > nb_decay_iter, return min learning rate
 471     if it > args.nb_decay_iter:
 472         return args.min_learning_rate
 473     # 3) in between, use cosine decay down to min learning rate
 474     decay_ratio = (it - args.nb_warmup_iter) / (
 475         args.nb_decay_iter - args.nb_warmup_iter
 476     )
 477     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 478     return args.min_learning_rate + coeff * (
 479         args.learning_rate - args.min_learning_rate
 480     )
 481
 482
 483 ######################################################################
 484
 485
 486 def picoclvr_pruner_horizontal_green(p):
 487     return not ("green" in p and ("left" in p or "right" in p))
 488
 489
 490 picoclvr_pruner_train = (
 491     picoclvr_pruner_horizontal_green
 492     if args.picocvlr_prune_properties in {"train+eval"}
 493     else None
 494 )
 495
 496 picoclvr_pruner_eval = (
 497     (lambda p: not picoclvr_pruner_horizontal_green(p))
 498     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 499     else None
 500 )
 501
 502 ######################################################################
 503
 504 device_data = device
 505
 506 if args.task == "byheart":
 507     task = tasks.SandBox(
 508         problem=problems.ProblemByHeart(),
 509         nb_train_samples=args.nb_train_samples,
 510         nb_test_samples=args.nb_test_samples,
 511         batch_size=args.batch_size,
 512         logger=log_string,
 513         device=device_data,
 514     )
 515     args.max_percents_of_test_in_train = -1
 516
 517 elif args.task == "learnop":
 518     task = tasks.SandBox(
 519         problem=problems.ProblemLearnOperator(),
 520         nb_train_samples=args.nb_train_samples,
 521         nb_test_samples=args.nb_test_samples,
 522         batch_size=args.batch_size,
 523         logger=log_string,
 524         device=device_data,
 525     )
 526
 527
 528 elif args.task == "guessop":
 529     task = tasks.SandBox(
 530         problem=problems.ProblemGuessOperator(),
 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
 538
 539 elif args.task == "twotargets":
 540     task = tasks.SandBox(
 541         problem=problems.ProblemTwoTargets(),
 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 elif args.task == "memory":
 550     task = tasks.SandBox(
 551         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 552         nb_train_samples=args.nb_train_samples,
 553         nb_test_samples=args.nb_test_samples,
 554         batch_size=args.batch_size,
 555         logger=log_string,
 556         device=device_data,
 557     )
 558
 559 elif args.task == "mixing":
 560     task = tasks.SandBox(
 561         problem=problems.ProblemMixing(
 562             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 563         ),
 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 == "addition":
 572     task = tasks.SandBox(
 573         problem=problems.ProblemAddition(),
 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 == "picoclvr":
 582     task = tasks.PicoCLVR(
 583         nb_train_samples=args.nb_train_samples,
 584         nb_test_samples=args.nb_test_samples,
 585         batch_size=args.batch_size,
 586         height=args.picoclvr_height,
 587         width=args.picoclvr_width,
 588         nb_colors=args.picoclvr_nb_colors,
 589         logger=log_string,
 590         device=device_data,
 591         pruner_train=picoclvr_pruner_train,
 592         pruner_eval=picoclvr_pruner_eval,
 593     )
 594
 595 elif args.task == "mnist":
 596     task = tasks.MNIST(
 597         nb_train_samples=args.nb_train_samples,
 598         nb_test_samples=args.nb_test_samples,
 599         batch_size=args.batch_size,
 600         device=device_data,
 601     )
 602
 603 elif args.task == "maze":
 604     task = tasks.Maze(
 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.maze_height,
 609         width=args.maze_width,
 610         nb_walls=args.maze_nb_walls,
 611         device=device_data,
 612     )
 613
 614 elif args.task == "snake":
 615     task = tasks.Snake(
 616         nb_train_samples=args.nb_train_samples,
 617         nb_test_samples=args.nb_test_samples,
 618         batch_size=args.batch_size,
 619         height=args.snake_height,
 620         width=args.snake_width,
 621         nb_colors=args.snake_nb_colors,
 622         length=args.snake_length,
 623         prompt_length=args.snake_length // 2,
 624         device=device_data,
 625     )
 626
 627 elif args.task == "stack":
 628     task = tasks.Stack(
 629         nb_train_samples=args.nb_train_samples,
 630         nb_test_samples=args.nb_test_samples,
 631         batch_size=args.batch_size,
 632         logger=log_string,
 633         nb_steps=args.stack_nb_steps,
 634         nb_stacks=args.stack_nb_stacks,
 635         nb_digits=args.stack_nb_digits,
 636         fraction_values_for_train=args.stack_fraction_values_for_train,
 637         device=device_data,
 638     )
 639
 640 elif args.task == "expr":
 641     task = tasks.Expr(
 642         nb_train_samples=args.nb_train_samples,
 643         nb_test_samples=args.nb_test_samples,
 644         nb_variables=args.expr_nb_variables,
 645         sequence_length=args.expr_sequence_length,
 646         operand_max=args.expr_operand_max,
 647         result_max=args.expr_result_max,
 648         batch_size=args.batch_size,
 649         device=device_data,
 650     )
 651
 652 elif args.task == "rpl":
 653     task = tasks.RPL(
 654         nb_train_samples=args.nb_train_samples,
 655         nb_test_samples=args.nb_test_samples,
 656         batch_size=args.batch_size,
 657         nb_starting_values=args.rpl_nb_starting_values,
 658         max_input=args.rpl_max_input,
 659         prog_len=args.rpl_prog_len,
 660         nb_runs=args.rpl_nb_runs,
 661         no_prog=args.rpl_no_prog,
 662         logger=log_string,
 663         device=device_data,
 664     )
 665
 666 elif args.task == "grid":
 667     task = tasks.Grid(
 668         nb_train_samples=args.nb_train_samples,
 669         nb_test_samples=args.nb_test_samples,
 670         batch_size=args.batch_size,
 671         size=args.grid_size,
 672         logger=log_string,
 673         device=device_data,
 674     )
 675
 676 elif args.task == "qmlp":
 677     task = tasks.QMLP(
 678         nb_train_samples=args.nb_train_samples,
 679         nb_test_samples=args.nb_test_samples,
 680         batch_size=args.batch_size,
 681         result_dir=args.result_dir,
 682         logger=log_string,
 683         device=device_data,
 684     )
 685
 686 else:
 687     raise ValueError(f"Unknown task {args.task}")
 688
 689 ######################################################################
 690
 691 log_string(f"device {device}")
 692
 693 vocabulary_size = task.vocabulary_size()
 694
 695 log_string(f"vocabulary_size {vocabulary_size}")
 696
 697 ##############################
 698
 699 model = mygpt.MyGPT(
 700     vocabulary_size=vocabulary_size,
 701     dim_model=args.dim_model,
 702     dim_keys=args.dim_keys,
 703     dim_hidden=args.dim_hidden,
 704     nb_heads=args.nb_heads,
 705     nb_lines=args.nb_lines,
 706     caterpillar_height=args.caterpillar_height,
 707     dim_rec_v=args.dim_rec_v,
 708     nb_blocks=args.nb_blocks,
 709     causal=True,
 710     dropout=args.dropout,
 711     attention_layer=args.attention,
 712 )
 713
 714 model.to(device)
 715
 716 nb_parameters = sum(p.numel() for p in model.parameters())
 717 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 718
 719 ######################################################################
 720
 721 nb_epochs_finished = 0
 722
 723 if args.no_checkpoint:
 724     log_string(f"not trying to load checkpoint.")
 725
 726 else:
 727     try:
 728         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 729         checkpoint = torch.load(checkpoint_name)
 730         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 731         model.load_state_dict(checkpoint["model_state"])
 732         torch.set_rng_state(checkpoint["rng_state"])
 733         if torch.cuda.is_available():
 734             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 735
 736         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 737
 738     except FileNotFoundError:
 739         log_string("starting from scratch.")
 740
 741     except:
 742         log_string("error when loading the checkpoint.")
 743         exit(1)
 744
 745 ######################################################################
 746
 747 if args.task == "expr" and args.expr_input_file is not None:
 748     task.produce_results(
 749         n_epoch=nb_epochs_finished,
 750         model=model,
 751         result_dir=args.result_dir,
 752         logger=log_string,
 753         deterministic_synthesis=args.deterministic_synthesis,
 754         input_file=args.expr_input_file,
 755     )
 756
 757     exit(0)
 758
 759 ######################################################################
 760
 761 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 762
 763 # Compute the entropy of the training tokens
 764
 765 token_count = 0
 766 for input in task.batches(split="train"):
 767     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 768 token_probas = token_count / token_count.sum()
 769 entropy = -torch.xlogy(token_probas, token_probas).sum()
 770 train_set_perplexity = math.exp(entropy)
 771
 772 ######################################################################
 773 # A bit of paranoia never hurts
 774
 775 if args.max_percents_of_test_in_train >= 0:
 776
 777     def subsets_as_tuples(batches, cs):
 778         s = set()
 779         for batch in batches:
 780             for x in batch:
 781                 s.add(tuple([v.item() for v in x]))
 782                 if len(s) == cs:
 783                     yield s
 784                     s = set()
 785         yield s
 786
 787     nb_test, nb_in_train = 0, 0
 788     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 789         in_train = set()
 790         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 791             in_train.update(test_subset.intersection(train_subset))
 792         nb_in_train += len(in_train)
 793         nb_test += len(test_subset)
 794
 795     log_string(
 796         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 797     )
 798
 799     assert (
 800         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 801     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 802
 803 ##############################
 804
 805 nb_samples_seen = 0
 806
 807 if nb_epochs_finished >= nb_epochs:
 808     task.produce_results(
 809         n_epoch=nb_epochs_finished,
 810         model=model,
 811         result_dir=args.result_dir,
 812         logger=log_string,
 813         deterministic_synthesis=args.deterministic_synthesis,
 814     )
 815
 816 time_pred_result = None
 817
 818 it = 0
 819
 820 for n_epoch in range(nb_epochs_finished, nb_epochs):
 821     if args.optim == "sgd":
 822         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 823     elif args.optim == "adam":
 824         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 825     elif args.optim == "adamw":
 826         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 827     else:
 828         raise ValueError(f"Unknown optimizer {args.optim}.")
 829
 830     model.train()
 831
 832     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 833
 834     for input in task.batches(split="train"):
 835         model.reset_inner_loss()
 836         input = input.to(device)
 837
 838         output = model(mygpt.BracketedSequence(input)).x
 839         loss = F.cross_entropy(output.transpose(1, 2), input)
 840         inner_loss = model.get_inner_loss()
 841
 842         acc_train_loss += loss.item() * input.size(0)
 843         acc_train_inner_loss += inner_loss.item() * input.size(0)
 844
 845         nb_train_samples += input.size(0)
 846         nb_samples_seen += input.size(0)
 847
 848         total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
 849
 850         it += 1
 851         lr = get_lr(it)
 852         for param_group in optimizer.param_groups:
 853             param_group["lr"] = lr
 854
 855         # log_string(f"learning_rate {lr}")
 856
 857         optimizer.zero_grad()
 858         total_loss.backward()
 859         optimizer.step()
 860
 861     with torch.autograd.no_grad():
 862         model.eval()
 863
 864         nb_test_samples, acc_test_loss = 0, 0.0
 865
 866         for input in task.batches(split="test"):
 867             input = input.to(device)
 868
 869             output = model(mygpt.BracketedSequence(input)).x
 870             loss = F.cross_entropy(output.transpose(1, 2), input)
 871             acc_test_loss += loss.item() * input.size(0)
 872             nb_test_samples += input.size(0)
 873
 874         log_string(
 875             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}"
 876         )
 877
 878         task.produce_results(
 879             n_epoch=n_epoch,
 880             model=model,
 881             result_dir=args.result_dir,
 882             logger=log_string,
 883             deterministic_synthesis=args.deterministic_synthesis,
 884         )
 885
 886         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 887         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 888
 889         log_string(
 890             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 891         )
 892
 893         time_current_result = datetime.datetime.now()
 894         if time_pred_result is not None:
 895             log_string(
 896                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 897             )
 898         time_pred_result = time_current_result
 899
 900     checkpoint = {
 901         "nb_epochs_finished": n_epoch + 1,
 902         "model_state": model.state_dict(),
 903         "rng_state": torch.get_rng_state(),
 904     }
 905
 906     if torch.cuda.is_available():
 907         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 908
 909     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 910     torch.save(checkpoint, checkpoint_name)
 911     log_string(f"saved checkpoint {checkpoint_name}")
 912
 913 ######################################################################