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 parser.add_argument("--grid_nb_colors", type=int, default=6)
 137
 138 parser.add_argument("--grid_nb_shapes", type=int, default=6)
 139
 140 ##############################
 141 # picoclvr options
 142
 143 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 144
 145 parser.add_argument("--picoclvr_height", type=int, default=12)
 146
 147 parser.add_argument("--picoclvr_width", type=int, default=16)
 148
 149 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 150
 151 ##############################
 152 # Maze options
 153
 154 parser.add_argument("--maze_height", type=int, default=13)
 155
 156 parser.add_argument("--maze_width", type=int, default=21)
 157
 158 parser.add_argument("--maze_nb_walls", type=int, default=15)
 159
 160 ##############################
 161 # Snake options
 162
 163 parser.add_argument("--snake_height", type=int, default=9)
 164
 165 parser.add_argument("--snake_width", type=int, default=12)
 166
 167 parser.add_argument("--snake_nb_colors", type=int, default=5)
 168
 169 parser.add_argument("--snake_length", type=int, default=200)
 170
 171 ##############################
 172 # Stack options
 173
 174 parser.add_argument("--stack_nb_steps", type=int, default=100)
 175
 176 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 177
 178 parser.add_argument("--stack_nb_digits", type=int, default=3)
 179
 180 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 181
 182 ##############################
 183 # Expr options
 184
 185 parser.add_argument("--expr_nb_variables", type=int, default=5)
 186
 187 parser.add_argument("--expr_sequence_length", type=int, default=40)
 188
 189 parser.add_argument("--expr_operand_max", type=int, default=9)
 190
 191 parser.add_argument("--expr_result_max", type=int, default=99)
 192
 193 parser.add_argument("--expr_input_file", type=str, default=None)
 194
 195 ##############################
 196 # Memory
 197
 198 parser.add_argument("--memory_len_total", type=int, default=32)
 199
 200 ##############################
 201 # Mixing
 202
 203 parser.add_argument("--mixing_hard", action="store_true", default=False)
 204
 205 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 206
 207 ######################################################################
 208
 209 # args = parser.parse_args()
 210
 211 args, sup_args = parser.parse_known_args()
 212
 213 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 214
 215 if args.result_dir is None:
 216     args.result_dir = f"results_{args.task}_{args.model}"
 217
 218 ######################################################################
 219
 220 if not args.force_cpu and torch.cuda.is_available():
 221     device = torch.device("cuda")
 222     torch.backends.cuda.matmul.allow_tf32 = True
 223 else:
 224     device = torch.device("cpu")
 225
 226 ######################################################################
 227
 228 default_task_args = {
 229     "addition": {
 230         "model": "352M",
 231         "batch_size": 25,
 232         "nb_train_samples": 250000,
 233         "nb_test_samples": 10000,
 234     },
 235     "byheart": {
 236         "model": "37M",
 237         "batch_size": 25,
 238         "nb_train_samples": 50000,
 239         "nb_test_samples": 10000,
 240     },
 241     "expr": {
 242         "model": "352M",
 243         "batch_size": 25,
 244         "nb_train_samples": 2500000,
 245         "nb_test_samples": 10000,
 246     },
 247     "grid": {
 248         "model": "37M",
 249         "batch_size": 25,
 250         "nb_train_samples": 250000,
 251         "nb_test_samples": 10000,
 252     },
 253     "qmlp": {
 254         "model": "37M",
 255         "batch_size": 10,
 256         "nb_train_samples": 100000,
 257         "nb_test_samples": 1000,
 258     },
 259     "guessop": {
 260         "model": "352M",
 261         "batch_size": 25,
 262         "nb_train_samples": 1000000,
 263         "nb_test_samples": 10000,
 264     },
 265     "learnop": {
 266         "model": "37M",
 267         "batch_size": 25,
 268         "nb_train_samples": 50000,
 269         "nb_test_samples": 10000,
 270     },
 271     "maze": {
 272         "model": "37M",
 273         "batch_size": 5,
 274         "nb_train_samples": 100000,
 275         "nb_test_samples": 10000,
 276     },
 277     "picoclvr": {
 278         "model": "37M",
 279         "batch_size": 25,
 280         "nb_train_samples": 250000,
 281         "nb_test_samples": 10000,
 282     },
 283     "rpl": {
 284         "model": "352M",
 285         "batch_size": 5,
 286         "nb_train_samples": 2500000,
 287         "nb_test_samples": 10000,
 288     },
 289     "snake": {
 290         "model": "37M",
 291         "batch_size": 25,
 292         "nb_train_samples": 250000,
 293         "nb_test_samples": 10000,
 294     },
 295     "stack": {
 296         "model": "37M",
 297         "batch_size": 25,
 298         "nb_train_samples": 100000,
 299         "nb_test_samples": 1000,
 300     },
 301     "twotargets": {
 302         "model": "37M",
 303         "batch_size": 25,
 304         "nb_train_samples": 50000,
 305         "nb_test_samples": 10000,
 306     },
 307     "memory": {
 308         "model": "37M",
 309         "batch_size": 25,
 310         "nb_train_samples": 25000,
 311         "nb_test_samples": 10000,
 312     },
 313     "mixing": {
 314         "model": "37M",
 315         "batch_size": 25,
 316         "nb_train_samples": 250000,
 317         "nb_test_samples": 10000,
 318     },
 319     "mnist": {
 320         "model": "37M",
 321         "batch_size": 10,
 322         "nb_train_samples": 60000,
 323         "nb_test_samples": 10000,
 324     },
 325 }
 326
 327 if args.task in default_task_args:
 328     for k, v in default_task_args[args.task].items():
 329         if getattr(args, k) is None:
 330             setattr(args, k, v)
 331
 332 ######################################################################
 333
 334 default_model_args = {
 335     "17K": {
 336         "attention": "mha",
 337         "dim_model": 32,
 338         "dim_keys": 32,
 339         "dim_hidden": 32,
 340         "nb_heads": 2,
 341         "nb_blocks": 2,
 342     },
 343     "17K-C": {
 344         "attention": "caterpillar",
 345         "dim_model": 32,
 346         "dim_keys": 32,
 347         "dim_hidden": 32,
 348         "nb_heads": 2,
 349         "nb_lines": 16,
 350         "caterpillar_height": 4,
 351         "nb_blocks": 2,
 352     },
 353     "4M": {
 354         "attention": "mha",
 355         "dim_model": 256,
 356         "dim_keys": 32,
 357         "dim_hidden": 1024,
 358         "nb_heads": 4,
 359         "nb_blocks": 6,
 360     },
 361     "4M-C": {
 362         "attention": "caterpillar",
 363         "dim_model": 256,
 364         "dim_keys": 32,
 365         "dim_hidden": 1024,
 366         "nb_heads": 4,
 367         "nb_lines": 32,
 368         "caterpillar_height": 4,
 369         "nb_blocks": 6,
 370     },
 371     "37M": {
 372         "attention": "mha",
 373         "dim_model": 512,
 374         "dim_keys": 64,
 375         "dim_hidden": 2048,
 376         "nb_heads": 8,
 377         "nb_blocks": 12,
 378     },
 379     "37M-C": {
 380         "attention": "caterpillar",
 381         "dim_model": 512,
 382         "dim_keys": 64,
 383         "dim_hidden": 2048,
 384         "nb_heads": 8,
 385         "nb_lines": 256,
 386         "caterpillar_height": 32,
 387         "nb_blocks": 12,
 388     },
 389     "122M": {
 390         "attention": "mha",
 391         "dim_model": 768,
 392         "dim_keys": 64,
 393         "dim_hidden": 2048,
 394         "nb_heads": 8,
 395         "nb_blocks": 24,
 396     },
 397     "122M-C": {
 398         "attention": "caterpillar",
 399         "dim_model": 768,
 400         "dim_keys": 64,
 401         "dim_hidden": 2048,
 402         "nb_heads": 8,
 403         "nb_lines": 128,
 404         "nb_blocks": 24,
 405     },
 406     "352M": {
 407         "attention": "mha",
 408         "dim_model": 1024,
 409         "dim_keys": 64,
 410         "dim_hidden": 2048,
 411         "nb_heads": 8,
 412         "nb_blocks": 48,
 413     },
 414     "352M-C": {
 415         "attention": "caterpillar",
 416         "dim_model": 1024,
 417         "dim_keys": 64,
 418         "dim_hidden": 2048,
 419         "nb_heads": 8,
 420         "nb_lines": 128,
 421         "nb_blocks": 48,
 422     },
 423 }
 424
 425 if args.model in default_model_args:
 426     for k, v in default_model_args[args.model].items():
 427         if getattr(args, k) is None:
 428             setattr(args, k, v)
 429 else:
 430     raise ValueError(f"Unknown model {args.model}")
 431
 432 ######################################################################
 433
 434 try:
 435     os.mkdir(args.result_dir)
 436 except FileExistsError:
 437     if not args.continue_training:
 438         print(f"result directory {args.result_dir} already exists")
 439         exit(1)
 440
 441 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
 442
 443 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 444
 445 if args.seed >= 0:
 446     # torch.backends.cudnn.deterministic = True
 447     # torch.backends.cudnn.benchmark = False
 448     # torch.use_deterministic_algorithms(True)
 449     torch.manual_seed(args.seed)
 450     if torch.cuda.is_available():
 451         torch.cuda.manual_seed_all(args.seed)
 452
 453 ######################################################################
 454
 455
 456 def log_string(s):
 457     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 458
 459     if log_file is not None:
 460         log_file.write(t + s + "\n")
 461         log_file.flush()
 462
 463     print(t + s)
 464     sys.stdout.flush()
 465
 466
 467 with os.popen("sha256sum *.py") as f:
 468     for l in f:
 469         log_string(f"sha256sum {l.strip()}")
 470
 471 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 472 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 473
 474 log_string(f"argv {' '.join(sys.argv)}")
 475
 476 for n in vars(args):
 477     log_string(f"args.{n} {getattr(args, n)}")
 478
 479 for k, v in sup_args.items():
 480     log_string(f'sup_args["{k}"] "{v}"')
 481
 482
 483 ######################################################################
 484
 485
 486 def get_lr(n_epoch, it):
 487     if args.legacy_lr_schedule:
 488         # my crude scheduling to compare to previous baseline, added
 489         # warmup though
 490
 491         if it < args.nb_warmup_iter:
 492             return args.legacy_large_lr * it / args.nb_warmup_iter
 493         elif n_epoch < args.legacy_nb_epoch_large_lr:
 494             return args.legacy_large_lr
 495         else:
 496             return args.legacy_small_lr
 497
 498     # from nanoGPT
 499
 500     # 1) linear warmup for warmup_iter steps
 501     if it < args.nb_warmup_iter:
 502         return args.learning_rate * it / args.nb_warmup_iter
 503     # 2) if it > nb_decay_iter, return min learning rate
 504     if it > args.nb_decay_iter:
 505         return args.min_learning_rate
 506     # 3) in between, use cosine decay down to min learning rate
 507     decay_ratio = (it - args.nb_warmup_iter) / (
 508         args.nb_decay_iter - args.nb_warmup_iter
 509     )
 510     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 511     return args.min_learning_rate + coeff * (
 512         args.learning_rate - args.min_learning_rate
 513     )
 514
 515
 516 ######################################################################
 517
 518
 519 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 520
 521
 522 def picoclvr_pruner_horizontal_green(p):
 523     return not ("green" in p and ("left" in p or "right" in p))
 524
 525
 526 picoclvr_pruner_train = (
 527     picoclvr_pruner_horizontal_green
 528     if args.picocvlr_prune_properties in {"train+eval"}
 529     else None
 530 )
 531
 532 picoclvr_pruner_eval = (
 533     (lambda p: not picoclvr_pruner_horizontal_green(p))
 534     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 535     else None
 536 )
 537
 538 ######################################################################
 539
 540 device_data = device
 541
 542 if args.task == "byheart":
 543     task = tasks.SandBox(
 544         problem=problems.ProblemByHeart(),
 545         nb_train_samples=args.nb_train_samples,
 546         nb_test_samples=args.nb_test_samples,
 547         batch_size=args.batch_size,
 548         logger=log_string,
 549         device=device_data,
 550     )
 551     args.max_percents_of_test_in_train = -1
 552
 553 elif args.task == "learnop":
 554     task = tasks.SandBox(
 555         problem=problems.ProblemLearnOperator(),
 556         nb_train_samples=args.nb_train_samples,
 557         nb_test_samples=args.nb_test_samples,
 558         batch_size=args.batch_size,
 559         logger=log_string,
 560         device=device_data,
 561     )
 562
 563
 564 elif args.task == "guessop":
 565     task = tasks.SandBox(
 566         problem=problems.ProblemGuessOperator(),
 567         nb_train_samples=args.nb_train_samples,
 568         nb_test_samples=args.nb_test_samples,
 569         batch_size=args.batch_size,
 570         logger=log_string,
 571         device=device_data,
 572     )
 573
 574
 575 elif args.task == "twotargets":
 576     task = tasks.SandBox(
 577         problem=problems.ProblemTwoTargets(),
 578         nb_train_samples=args.nb_train_samples,
 579         nb_test_samples=args.nb_test_samples,
 580         batch_size=args.batch_size,
 581         logger=log_string,
 582         device=device_data,
 583     )
 584
 585 elif args.task == "memory":
 586     task = tasks.SandBox(
 587         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 588         nb_train_samples=args.nb_train_samples,
 589         nb_test_samples=args.nb_test_samples,
 590         batch_size=args.batch_size,
 591         logger=log_string,
 592         device=device_data,
 593     )
 594
 595 elif args.task == "mixing":
 596     task = tasks.SandBox(
 597         problem=problems.ProblemMixing(
 598             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 599         ),
 600         nb_train_samples=args.nb_train_samples,
 601         nb_test_samples=args.nb_test_samples,
 602         batch_size=args.batch_size,
 603         logger=log_string,
 604         device=device_data,
 605     )
 606
 607 elif args.task == "addition":
 608     task = tasks.SandBox(
 609         problem=problems.ProblemAddition(),
 610         nb_train_samples=args.nb_train_samples,
 611         nb_test_samples=args.nb_test_samples,
 612         batch_size=args.batch_size,
 613         logger=log_string,
 614         device=device_data,
 615     )
 616
 617 elif args.task == "picoclvr":
 618     task = tasks.PicoCLVR(
 619         nb_train_samples=args.nb_train_samples,
 620         nb_test_samples=args.nb_test_samples,
 621         batch_size=args.batch_size,
 622         height=args.picoclvr_height,
 623         width=args.picoclvr_width,
 624         nb_colors=args.picoclvr_nb_colors,
 625         logger=log_string,
 626         device=device_data,
 627         pruner_train=picoclvr_pruner_train,
 628         pruner_eval=picoclvr_pruner_eval,
 629     )
 630
 631 elif args.task == "mnist":
 632     task = tasks.MNIST(
 633         nb_train_samples=args.nb_train_samples,
 634         nb_test_samples=args.nb_test_samples,
 635         batch_size=args.batch_size,
 636         device=device_data,
 637     )
 638
 639 elif args.task == "maze":
 640     task = tasks.Maze(
 641         nb_train_samples=args.nb_train_samples,
 642         nb_test_samples=args.nb_test_samples,
 643         batch_size=args.batch_size,
 644         height=args.maze_height,
 645         width=args.maze_width,
 646         nb_walls=args.maze_nb_walls,
 647         device=device_data,
 648     )
 649
 650 elif args.task == "snake":
 651     task = tasks.Snake(
 652         nb_train_samples=args.nb_train_samples,
 653         nb_test_samples=args.nb_test_samples,
 654         batch_size=args.batch_size,
 655         height=args.snake_height,
 656         width=args.snake_width,
 657         nb_colors=args.snake_nb_colors,
 658         length=args.snake_length,
 659         prompt_length=args.snake_length // 2,
 660         device=device_data,
 661     )
 662
 663 elif args.task == "stack":
 664     task = tasks.Stack(
 665         nb_train_samples=args.nb_train_samples,
 666         nb_test_samples=args.nb_test_samples,
 667         batch_size=args.batch_size,
 668         logger=log_string,
 669         nb_steps=args.stack_nb_steps,
 670         nb_stacks=args.stack_nb_stacks,
 671         nb_digits=args.stack_nb_digits,
 672         fraction_values_for_train=args.stack_fraction_values_for_train,
 673         device=device_data,
 674     )
 675
 676 elif args.task == "expr":
 677     task = tasks.Expr(
 678         nb_train_samples=args.nb_train_samples,
 679         nb_test_samples=args.nb_test_samples,
 680         nb_variables=args.expr_nb_variables,
 681         sequence_length=args.expr_sequence_length,
 682         operand_max=args.expr_operand_max,
 683         result_max=args.expr_result_max,
 684         batch_size=args.batch_size,
 685         device=device_data,
 686     )
 687
 688 elif args.task == "rpl":
 689     task = tasks.RPL(
 690         nb_train_samples=args.nb_train_samples,
 691         nb_test_samples=args.nb_test_samples,
 692         batch_size=args.batch_size,
 693         nb_starting_values=args.rpl_nb_starting_values,
 694         max_input=args.rpl_max_input,
 695         prog_len=args.rpl_prog_len,
 696         nb_runs=args.rpl_nb_runs,
 697         no_prog=args.rpl_no_prog,
 698         logger=log_string,
 699         device=device_data,
 700     )
 701
 702 elif args.task == "grid":
 703     task = tasks.Grid(
 704         nb_train_samples=args.nb_train_samples,
 705         nb_test_samples=args.nb_test_samples,
 706         batch_size=args.batch_size,
 707         size=args.grid_size,
 708         nb_shapes=args.grid_nb_shapes,
 709         nb_colors=args.grid_nb_colors,
 710         logger=log_string,
 711         device=device_data,
 712     )
 713
 714 elif args.task == "qmlp":
 715     task = tasks.QMLP(
 716         nb_train_samples=args.nb_train_samples,
 717         nb_test_samples=args.nb_test_samples,
 718         batch_size=args.batch_size,
 719         result_dir=args.result_dir,
 720         logger=log_string,
 721         device=device_data,
 722     )
 723
 724 else:
 725     raise ValueError(f"Unknown task {args.task}")
 726
 727 ######################################################################
 728
 729 log_string(f"device {device}")
 730
 731 vocabulary_size = task.vocabulary_size()
 732
 733 log_string(f"vocabulary_size {vocabulary_size}")
 734
 735 ##############################
 736
 737 model = mygpt.MyGPT(
 738     vocabulary_size=vocabulary_size,
 739     dim_model=args.dim_model,
 740     dim_keys=args.dim_keys,
 741     dim_hidden=args.dim_hidden,
 742     nb_heads=args.nb_heads,
 743     nb_lines=args.nb_lines,
 744     caterpillar_height=args.caterpillar_height,
 745     nb_blocks=args.nb_blocks,
 746     causal=True,
 747     dropout=args.dropout,
 748     attention_layer=args.attention,
 749     logger=log_string,
 750     **sup_args,
 751 )
 752
 753 model.to(device)
 754
 755 nb_parameters = sum(p.numel() for p in model.parameters())
 756 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 757
 758 ######################################################################
 759
 760 nb_epochs_finished = 0
 761
 762 if args.no_checkpoint:
 763     log_string(f"not trying to load checkpoint.")
 764
 765 else:
 766     try:
 767         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 768         checkpoint = torch.load(checkpoint_name)
 769         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 770         model.load_state_dict(checkpoint["model_state"])
 771         torch.set_rng_state(checkpoint["rng_state"])
 772         if torch.cuda.is_available():
 773             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 774
 775         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 776
 777     except FileNotFoundError:
 778         log_string("starting from scratch.")
 779
 780     except:
 781         log_string("error when loading the checkpoint.")
 782         exit(1)
 783
 784 ######################################################################
 785
 786 if args.task == "expr" and args.expr_input_file is not None:
 787     task.produce_results(
 788         n_epoch=nb_epochs_finished,
 789         model=model,
 790         result_dir=args.result_dir,
 791         logger=log_string,
 792         deterministic_synthesis=args.deterministic_synthesis,
 793         input_file=args.expr_input_file,
 794     )
 795
 796     exit(0)
 797
 798 ######################################################################
 799
 800 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 801
 802 # Compute the entropy of the training tokens
 803
 804 token_count = 0
 805 for input in task.batches(split="train"):
 806     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 807 token_probas = token_count / token_count.sum()
 808 entropy = -torch.xlogy(token_probas, token_probas).sum()
 809 train_set_perplexity = math.exp(entropy)
 810
 811 ######################################################################
 812 # A bit of paranoia never hurts
 813
 814 if args.max_percents_of_test_in_train >= 0:
 815
 816     def subsets_as_tuples(batches, cs):
 817         s = set()
 818         for batch in batches:
 819             for x in batch:
 820                 s.add(tuple([v.item() for v in x]))
 821                 if len(s) == cs:
 822                     yield s
 823                     s = set()
 824         yield s
 825
 826     nb_test, nb_in_train = 0, 0
 827     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 828         in_train = set()
 829         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 830             in_train.update(test_subset.intersection(train_subset))
 831         nb_in_train += len(in_train)
 832         nb_test += len(test_subset)
 833
 834     log_string(
 835         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 836     )
 837
 838     assert (
 839         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 840     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 841
 842 ##############################
 843
 844 if "calibrate" in sup_args:
 845     for input in task.batches(split="train", desc="calibrate"):
 846         input = input.to(device)
 847         output = model(mygpt.BracketedSequence(input)).x
 848
 849     for n, m in model.named_modules():
 850         for a in dir(m):
 851             x = getattr(m, a)
 852             if isinstance(x, mygpt.Calibrator):
 853                 print(f"####### ${n} | ${a} ########################")
 854                 mean, std = x.moments()
 855                 print("mean\n", mean, "\n")
 856                 print("std\n", std, "\n")
 857                 print(f"############################################\n\n")
 858
 859     exit(0)
 860
 861 ##############################
 862
 863 nb_samples_seen = 0
 864
 865 if nb_epochs_finished >= nb_epochs:
 866     task.produce_results(
 867         n_epoch=nb_epochs_finished,
 868         model=model,
 869         result_dir=args.result_dir,
 870         logger=log_string,
 871         deterministic_synthesis=args.deterministic_synthesis,
 872     )
 873
 874 time_pred_result = datetime.datetime.now()
 875
 876 it = 0
 877
 878 n_batch = 0
 879
 880 for n_epoch in range(nb_epochs_finished, nb_epochs):
 881     if args.optim == "sgd":
 882         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 883     elif args.optim == "adam":
 884         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 885     elif args.optim == "adamw":
 886         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 887     else:
 888         raise ValueError(f"Unknown optimizer {args.optim}.")
 889
 890     model.train()
 891
 892     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 893
 894     for input in task.batches(split="train"):
 895         model.reset_inner_loss()
 896         input = input.to(device)
 897
 898         output = model(mygpt.BracketedSequence(input)).x
 899         loss = F.cross_entropy(output.transpose(1, 2), input)
 900         inner_loss = model.get_inner_loss()
 901
 902         acc_train_loss += loss.item() * input.size(0)
 903         acc_train_inner_loss += inner_loss.item() * input.size(0)
 904
 905         nb_train_samples += input.size(0)
 906         nb_samples_seen += input.size(0)
 907
 908         total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
 909
 910         it += 1
 911         lr = get_lr(n_epoch, it)
 912         for param_group in optimizer.param_groups:
 913             param_group["lr"] = lr
 914
 915         # log_string(f"learning_rate {lr}")
 916
 917         optimizer.zero_grad()
 918         total_loss.backward()
 919         optimizer.step()
 920
 921         grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
 922
 923         loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
 924
 925         n_batch += 1
 926
 927     with torch.autograd.no_grad():
 928         model.eval()
 929
 930         nb_test_samples, acc_test_loss = 0, 0.0
 931
 932         for input in task.batches(split="test"):
 933             input = input.to(device)
 934
 935             output = model(mygpt.BracketedSequence(input)).x
 936             loss = F.cross_entropy(output.transpose(1, 2), input)
 937             acc_test_loss += loss.item() * input.size(0)
 938             nb_test_samples += input.size(0)
 939
 940         log_string(
 941             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}"
 942         )
 943
 944         task.produce_results(
 945             n_epoch=n_epoch,
 946             model=model,
 947             result_dir=args.result_dir,
 948             logger=log_string,
 949             deterministic_synthesis=args.deterministic_synthesis,
 950         )
 951
 952         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 953         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 954
 955         log_string(
 956             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 957         )
 958
 959         time_current_result = datetime.datetime.now()
 960         log_string(
 961             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 962         )
 963         time_pred_result = time_current_result
 964
 965     checkpoint = {
 966         "nb_epochs_finished": n_epoch + 1,
 967         "model_state": model.state_dict(),
 968         "rng_state": torch.get_rng_state(),
 969     }
 970
 971     if torch.cuda.is_available():
 972         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 973
 974     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 975     torch.save(checkpoint, checkpoint_name)
 976     log_string(f"saved checkpoint {checkpoint_name}")
 977
 978 ######################################################################