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("--proportion_memex", type=float, default=0)
  91
  92 parser.add_argument("--dim_model", type=int, default=None)
  93
  94 parser.add_argument("--dim_keys", type=int, default=None)
  95
  96 parser.add_argument("--dim_hidden", type=int, default=None)
  97
  98 parser.add_argument("--nb_heads", type=int, default=None)
  99
 100 parser.add_argument("--nb_lines", type=int, default=None)
 101
 102 parser.add_argument("--caterpillar_height", type=int, default=None)
 103
 104 parser.add_argument("--gate_dropout_proba", type=float, default=0.0)
 105
 106 parser.add_argument("--gate_dropout_sync", type=str2bool, default=False)
 107
 108 parser.add_argument("--gate_dropout_replace", type=str2bool, default=False)
 109
 110 parser.add_argument("--rho_inner_loss", type=float, default=0.0)
 111
 112 parser.add_argument("--nb_blocks", type=int, default=None)
 113
 114 parser.add_argument("--dropout", type=float, default=0.1)
 115
 116 ########################################
 117
 118 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 119
 120 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 121
 122 parser.add_argument("--continue_training", action="store_true", default=False)
 123
 124 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 125
 126 ##############################
 127 # rpl options
 128
 129 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 130
 131 parser.add_argument("--rpl_max_input", type=int, default=9)
 132
 133 parser.add_argument("--rpl_prog_len", type=int, default=8)
 134
 135 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 136
 137 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 138
 139 ##############################
 140 # grid options
 141
 142 parser.add_argument("--grid_size", type=int, default=6)
 143
 144 parser.add_argument("--grid_nb_colors", type=int, default=6)
 145
 146 parser.add_argument("--grid_nb_shapes", type=int, default=6)
 147
 148 ##############################
 149 # picoclvr options
 150
 151 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 152
 153 parser.add_argument("--picoclvr_height", type=int, default=12)
 154
 155 parser.add_argument("--picoclvr_width", type=int, default=16)
 156
 157 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 158
 159 ##############################
 160 # Maze options
 161
 162 parser.add_argument("--maze_height", type=int, default=13)
 163
 164 parser.add_argument("--maze_width", type=int, default=21)
 165
 166 parser.add_argument("--maze_nb_walls", type=int, default=15)
 167
 168 ##############################
 169 # Snake options
 170
 171 parser.add_argument("--snake_height", type=int, default=9)
 172
 173 parser.add_argument("--snake_width", type=int, default=12)
 174
 175 parser.add_argument("--snake_nb_colors", type=int, default=5)
 176
 177 parser.add_argument("--snake_length", type=int, default=200)
 178
 179 ##############################
 180 # Stack options
 181
 182 parser.add_argument("--stack_nb_steps", type=int, default=100)
 183
 184 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 185
 186 parser.add_argument("--stack_nb_digits", type=int, default=3)
 187
 188 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 189
 190 ##############################
 191 # Expr options
 192
 193 parser.add_argument("--expr_nb_variables", type=int, default=5)
 194
 195 parser.add_argument("--expr_sequence_length", type=int, default=40)
 196
 197 parser.add_argument("--expr_operand_max", type=int, default=9)
 198
 199 parser.add_argument("--expr_result_max", type=int, default=99)
 200
 201 parser.add_argument("--expr_input_file", type=str, default=None)
 202
 203 ##############################
 204 # Memory
 205
 206 parser.add_argument("--memory_len_total", type=int, default=32)
 207
 208 ##############################
 209 # Mixing
 210
 211 parser.add_argument("--mixing_hard", action="store_true", default=False)
 212
 213 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 214
 215 ######################################################################
 216
 217 # args = parser.parse_args()
 218
 219 args, sup_args = parser.parse_known_args()
 220
 221 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 222
 223 if args.result_dir is None:
 224     args.result_dir = f"results_{args.task}_{args.model}"
 225
 226 ######################################################################
 227
 228 if not args.force_cpu and torch.cuda.is_available():
 229     device = torch.device("cuda")
 230     torch.backends.cuda.matmul.allow_tf32 = True
 231 else:
 232     device = torch.device("cpu")
 233
 234 ######################################################################
 235
 236 default_task_args = {
 237     "addition": {
 238         "model": "352M",
 239         "batch_size": 25,
 240         "nb_train_samples": 250000,
 241         "nb_test_samples": 10000,
 242     },
 243     "byheart": {
 244         "model": "37M",
 245         "batch_size": 25,
 246         "nb_train_samples": 50000,
 247         "nb_test_samples": 10000,
 248     },
 249     "expr": {
 250         "model": "352M",
 251         "batch_size": 25,
 252         "nb_train_samples": 2500000,
 253         "nb_test_samples": 10000,
 254     },
 255     "grid": {
 256         "model": "37M",
 257         "batch_size": 25,
 258         "nb_train_samples": 250000,
 259         "nb_test_samples": 10000,
 260     },
 261     "qmlp": {
 262         "model": "37M",
 263         "batch_size": 10,
 264         "nb_train_samples": 100000,
 265         "nb_test_samples": 1000,
 266     },
 267     "guessop": {
 268         "model": "352M",
 269         "batch_size": 25,
 270         "nb_train_samples": 1000000,
 271         "nb_test_samples": 10000,
 272     },
 273     "learnop": {
 274         "model": "37M",
 275         "batch_size": 25,
 276         "nb_train_samples": 50000,
 277         "nb_test_samples": 10000,
 278     },
 279     "maze": {
 280         "model": "37M",
 281         "batch_size": 5,
 282         "nb_train_samples": 100000,
 283         "nb_test_samples": 10000,
 284     },
 285     "picoclvr": {
 286         "model": "37M",
 287         "batch_size": 25,
 288         "nb_train_samples": 250000,
 289         "nb_test_samples": 10000,
 290     },
 291     "rpl": {
 292         "model": "352M",
 293         "batch_size": 5,
 294         "nb_train_samples": 2500000,
 295         "nb_test_samples": 10000,
 296     },
 297     "snake": {
 298         "model": "37M",
 299         "batch_size": 25,
 300         "nb_train_samples": 250000,
 301         "nb_test_samples": 10000,
 302     },
 303     "stack": {
 304         "model": "37M",
 305         "batch_size": 25,
 306         "nb_train_samples": 100000,
 307         "nb_test_samples": 1000,
 308     },
 309     "twotargets": {
 310         "model": "37M",
 311         "batch_size": 25,
 312         "nb_train_samples": 50000,
 313         "nb_test_samples": 10000,
 314     },
 315     "memory": {
 316         "model": "37M",
 317         "batch_size": 25,
 318         "nb_train_samples": 25000,
 319         "nb_test_samples": 10000,
 320     },
 321     "mixing": {
 322         "model": "37M",
 323         "batch_size": 25,
 324         "nb_train_samples": 250000,
 325         "nb_test_samples": 10000,
 326     },
 327     "mnist": {
 328         "model": "37M",
 329         "batch_size": 10,
 330         "nb_train_samples": 60000,
 331         "nb_test_samples": 10000,
 332     },
 333 }
 334
 335 if args.task in default_task_args:
 336     for k, v in default_task_args[args.task].items():
 337         if getattr(args, k) is None:
 338             setattr(args, k, v)
 339
 340 ######################################################################
 341
 342 default_model_args = {
 343     "17K": {
 344         "attention": "mha",
 345         "dim_model": 32,
 346         "dim_keys": 32,
 347         "dim_hidden": 32,
 348         "nb_heads": 2,
 349         "nb_blocks": 2,
 350     },
 351     "17K-C": {
 352         "attention": "caterpillar",
 353         "dim_model": 32,
 354         "dim_keys": 32,
 355         "dim_hidden": 32,
 356         "nb_heads": 2,
 357         "nb_lines": 16,
 358         "caterpillar_height": 4,
 359         "nb_blocks": 2,
 360     },
 361     "4M": {
 362         "attention": "mha",
 363         "dim_model": 256,
 364         "dim_keys": 32,
 365         "dim_hidden": 1024,
 366         "nb_heads": 4,
 367         "nb_blocks": 6,
 368     },
 369     "4M-C": {
 370         "attention": "caterpillar",
 371         "dim_model": 256,
 372         "dim_keys": 32,
 373         "dim_hidden": 1024,
 374         "nb_heads": 4,
 375         "nb_lines": 32,
 376         "caterpillar_height": 4,
 377         "nb_blocks": 6,
 378     },
 379     "37M": {
 380         "attention": "mha",
 381         "dim_model": 512,
 382         "dim_keys": 64,
 383         "dim_hidden": 2048,
 384         "nb_heads": 8,
 385         "nb_blocks": 12,
 386     },
 387     "37M-C": {
 388         "attention": "caterpillar",
 389         "dim_model": 512,
 390         "dim_keys": 64,
 391         "dim_hidden": 2048,
 392         "nb_heads": 8,
 393         "nb_lines": 256,
 394         "caterpillar_height": 32,
 395         "nb_blocks": 12,
 396     },
 397     "122M": {
 398         "attention": "mha",
 399         "dim_model": 768,
 400         "dim_keys": 64,
 401         "dim_hidden": 2048,
 402         "nb_heads": 8,
 403         "nb_blocks": 24,
 404     },
 405     "122M-C": {
 406         "attention": "caterpillar",
 407         "dim_model": 768,
 408         "dim_keys": 64,
 409         "dim_hidden": 2048,
 410         "nb_heads": 8,
 411         "nb_lines": 128,
 412         "nb_blocks": 24,
 413     },
 414     "352M": {
 415         "attention": "mha",
 416         "dim_model": 1024,
 417         "dim_keys": 64,
 418         "dim_hidden": 2048,
 419         "nb_heads": 8,
 420         "nb_blocks": 48,
 421     },
 422     "352M-C": {
 423         "attention": "caterpillar",
 424         "dim_model": 1024,
 425         "dim_keys": 64,
 426         "dim_hidden": 2048,
 427         "nb_heads": 8,
 428         "nb_lines": 128,
 429         "nb_blocks": 48,
 430     },
 431 }
 432
 433 if args.model in default_model_args:
 434     for k, v in default_model_args[args.model].items():
 435         if getattr(args, k) is None:
 436             setattr(args, k, v)
 437 else:
 438     raise ValueError(f"Unknown model {args.model}")
 439
 440 ######################################################################
 441
 442 try:
 443     os.mkdir(args.result_dir)
 444 except FileExistsError:
 445     if not args.continue_training:
 446         print(f"result directory {args.result_dir} already exists")
 447         exit(1)
 448
 449 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
 450
 451 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 452
 453 if args.seed >= 0:
 454     # torch.backends.cudnn.deterministic = True
 455     # torch.backends.cudnn.benchmark = False
 456     # torch.use_deterministic_algorithms(True)
 457     torch.manual_seed(args.seed)
 458     if torch.cuda.is_available():
 459         torch.cuda.manual_seed_all(args.seed)
 460
 461 ######################################################################
 462
 463
 464 def log_string(s):
 465     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 466
 467     if log_file is not None:
 468         log_file.write(t + s + "\n")
 469         log_file.flush()
 470
 471     print(t + s)
 472     sys.stdout.flush()
 473
 474
 475 with os.popen("sha256sum *.py") as f:
 476     for l in f:
 477         log_string(f"sha256sum {l.strip()}")
 478
 479 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 480 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 481
 482 log_string(f"argv {' '.join(sys.argv)}")
 483
 484 for n in vars(args):
 485     log_string(f"args.{n} {getattr(args, n)}")
 486
 487 for k, v in sup_args.items():
 488     log_string(f'sup_args["{k}"] "{v}"')
 489
 490
 491 ######################################################################
 492
 493
 494 def get_lr(n_epoch, it):
 495     if args.legacy_lr_schedule:
 496         # my crude scheduling to compare to previous baseline, added
 497         # warmup though
 498
 499         if it < args.nb_warmup_iter:
 500             return args.legacy_large_lr * it / args.nb_warmup_iter
 501         elif n_epoch < args.legacy_nb_epoch_large_lr:
 502             return args.legacy_large_lr
 503         else:
 504             return args.legacy_small_lr
 505
 506     # from nanoGPT
 507
 508     # 1) linear warmup for warmup_iter steps
 509     if it < args.nb_warmup_iter:
 510         return args.learning_rate * it / args.nb_warmup_iter
 511     # 2) if it > nb_decay_iter, return min learning rate
 512     if it > args.nb_decay_iter:
 513         return args.min_learning_rate
 514     # 3) in between, use cosine decay down to min learning rate
 515     decay_ratio = (it - args.nb_warmup_iter) / (
 516         args.nb_decay_iter - args.nb_warmup_iter
 517     )
 518     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 519     return args.min_learning_rate + coeff * (
 520         args.learning_rate - args.min_learning_rate
 521     )
 522
 523
 524 ######################################################################
 525
 526
 527 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 528
 529
 530 def picoclvr_pruner_horizontal_green(p):
 531     return not ("green" in p and ("left" in p or "right" in p))
 532
 533
 534 picoclvr_pruner_train = (
 535     picoclvr_pruner_horizontal_green
 536     if args.picocvlr_prune_properties in {"train+eval"}
 537     else None
 538 )
 539
 540 picoclvr_pruner_eval = (
 541     (lambda p: not picoclvr_pruner_horizontal_green(p))
 542     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 543     else None
 544 )
 545
 546 ######################################################################
 547
 548 device_data = device
 549
 550 if args.task == "byheart":
 551     task = tasks.SandBox(
 552         problem=problems.ProblemByHeart(),
 553         nb_train_samples=args.nb_train_samples,
 554         nb_test_samples=args.nb_test_samples,
 555         batch_size=args.batch_size,
 556         logger=log_string,
 557         device=device_data,
 558     )
 559     args.max_percents_of_test_in_train = -1
 560
 561 elif args.task == "learnop":
 562     task = tasks.SandBox(
 563         problem=problems.ProblemLearnOperator(),
 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
 572 elif args.task == "guessop":
 573     task = tasks.SandBox(
 574         problem=problems.ProblemGuessOperator(),
 575         nb_train_samples=args.nb_train_samples,
 576         nb_test_samples=args.nb_test_samples,
 577         batch_size=args.batch_size,
 578         logger=log_string,
 579         device=device_data,
 580     )
 581
 582
 583 elif args.task == "twotargets":
 584     task = tasks.SandBox(
 585         problem=problems.ProblemTwoTargets(),
 586         nb_train_samples=args.nb_train_samples,
 587         nb_test_samples=args.nb_test_samples,
 588         batch_size=args.batch_size,
 589         logger=log_string,
 590         device=device_data,
 591     )
 592
 593 elif args.task == "memory":
 594     task = tasks.SandBox(
 595         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 596         nb_train_samples=args.nb_train_samples,
 597         nb_test_samples=args.nb_test_samples,
 598         batch_size=args.batch_size,
 599         logger=log_string,
 600         device=device_data,
 601     )
 602
 603 elif args.task == "mixing":
 604     task = tasks.SandBox(
 605         problem=problems.ProblemMixing(
 606             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 607         ),
 608         nb_train_samples=args.nb_train_samples,
 609         nb_test_samples=args.nb_test_samples,
 610         batch_size=args.batch_size,
 611         logger=log_string,
 612         device=device_data,
 613     )
 614
 615 elif args.task == "addition":
 616     task = tasks.SandBox(
 617         problem=problems.ProblemAddition(),
 618         nb_train_samples=args.nb_train_samples,
 619         nb_test_samples=args.nb_test_samples,
 620         batch_size=args.batch_size,
 621         logger=log_string,
 622         device=device_data,
 623     )
 624
 625 elif args.task == "picoclvr":
 626     task = tasks.PicoCLVR(
 627         nb_train_samples=args.nb_train_samples,
 628         nb_test_samples=args.nb_test_samples,
 629         batch_size=args.batch_size,
 630         height=args.picoclvr_height,
 631         width=args.picoclvr_width,
 632         nb_colors=args.picoclvr_nb_colors,
 633         logger=log_string,
 634         device=device_data,
 635         pruner_train=picoclvr_pruner_train,
 636         pruner_eval=picoclvr_pruner_eval,
 637     )
 638
 639 elif args.task == "mnist":
 640     task = tasks.MNIST(
 641         nb_train_samples=args.nb_train_samples,
 642         nb_test_samples=args.nb_test_samples,
 643         batch_size=args.batch_size,
 644         device=device_data,
 645     )
 646
 647 elif args.task == "maze":
 648     task = tasks.Maze(
 649         nb_train_samples=args.nb_train_samples,
 650         nb_test_samples=args.nb_test_samples,
 651         batch_size=args.batch_size,
 652         height=args.maze_height,
 653         width=args.maze_width,
 654         nb_walls=args.maze_nb_walls,
 655         device=device_data,
 656     )
 657
 658 elif args.task == "snake":
 659     task = tasks.Snake(
 660         nb_train_samples=args.nb_train_samples,
 661         nb_test_samples=args.nb_test_samples,
 662         batch_size=args.batch_size,
 663         height=args.snake_height,
 664         width=args.snake_width,
 665         nb_colors=args.snake_nb_colors,
 666         length=args.snake_length,
 667         prompt_length=args.snake_length // 2,
 668         device=device_data,
 669     )
 670
 671 elif args.task == "stack":
 672     task = tasks.Stack(
 673         nb_train_samples=args.nb_train_samples,
 674         nb_test_samples=args.nb_test_samples,
 675         batch_size=args.batch_size,
 676         logger=log_string,
 677         nb_steps=args.stack_nb_steps,
 678         nb_stacks=args.stack_nb_stacks,
 679         nb_digits=args.stack_nb_digits,
 680         fraction_values_for_train=args.stack_fraction_values_for_train,
 681         device=device_data,
 682     )
 683
 684 elif args.task == "expr":
 685     task = tasks.Expr(
 686         nb_train_samples=args.nb_train_samples,
 687         nb_test_samples=args.nb_test_samples,
 688         nb_variables=args.expr_nb_variables,
 689         sequence_length=args.expr_sequence_length,
 690         operand_max=args.expr_operand_max,
 691         result_max=args.expr_result_max,
 692         batch_size=args.batch_size,
 693         device=device_data,
 694     )
 695
 696 elif args.task == "rpl":
 697     task = tasks.RPL(
 698         nb_train_samples=args.nb_train_samples,
 699         nb_test_samples=args.nb_test_samples,
 700         batch_size=args.batch_size,
 701         nb_starting_values=args.rpl_nb_starting_values,
 702         max_input=args.rpl_max_input,
 703         prog_len=args.rpl_prog_len,
 704         nb_runs=args.rpl_nb_runs,
 705         no_prog=args.rpl_no_prog,
 706         logger=log_string,
 707         device=device_data,
 708     )
 709
 710 elif args.task == "grid":
 711     task = tasks.Grid(
 712         nb_train_samples=args.nb_train_samples,
 713         nb_test_samples=args.nb_test_samples,
 714         batch_size=args.batch_size,
 715         size=args.grid_size,
 716         nb_shapes=args.grid_nb_shapes,
 717         nb_colors=args.grid_nb_colors,
 718         logger=log_string,
 719         device=device_data,
 720     )
 721
 722 elif args.task == "qmlp":
 723     task = tasks.QMLP(
 724         nb_train_samples=args.nb_train_samples,
 725         nb_test_samples=args.nb_test_samples,
 726         batch_size=args.batch_size,
 727         result_dir=args.result_dir,
 728         logger=log_string,
 729         device=device_data,
 730     )
 731
 732 else:
 733     raise ValueError(f"Unknown task {args.task}")
 734
 735 ######################################################################
 736
 737 log_string(f"device {device}")
 738
 739 vocabulary_size = task.vocabulary_size()
 740
 741 if args.proportion_memex > 0:
 742     vocabulary_size += 1
 743
 744 log_string(f"vocabulary_size {vocabulary_size}")
 745
 746 ##############################
 747
 748 model = mygpt.MyGPT(
 749     vocabulary_size=vocabulary_size,
 750     dim_model=args.dim_model,
 751     dim_keys=args.dim_keys,
 752     dim_hidden=args.dim_hidden,
 753     nb_heads=args.nb_heads,
 754     nb_lines=args.nb_lines,
 755     caterpillar_height=args.caterpillar_height,
 756     nb_blocks=args.nb_blocks,
 757     causal=True,
 758     dropout=args.dropout,
 759     attention_layer=args.attention,
 760     logger=log_string,
 761     args=args,
 762 )
 763
 764 model.to(device)
 765
 766 nb_parameters = sum(p.numel() for p in model.parameters())
 767 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 768
 769 ######################################################################
 770
 771 nb_epochs_finished = 0
 772
 773 if args.no_checkpoint:
 774     log_string(f"not trying to load checkpoint.")
 775
 776 else:
 777     try:
 778         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 779         checkpoint = torch.load(checkpoint_name)
 780         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 781         model.load_state_dict(checkpoint["model_state"])
 782         torch.set_rng_state(checkpoint["rng_state"])
 783         if torch.cuda.is_available():
 784             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 785
 786         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 787
 788     except FileNotFoundError:
 789         log_string("starting from scratch.")
 790
 791     except:
 792         log_string("error when loading the checkpoint.")
 793         exit(1)
 794
 795 ######################################################################
 796
 797 if args.task == "expr" and args.expr_input_file is not None:
 798     task.produce_results(
 799         n_epoch=nb_epochs_finished,
 800         model=model,
 801         result_dir=args.result_dir,
 802         logger=log_string,
 803         deterministic_synthesis=args.deterministic_synthesis,
 804         input_file=args.expr_input_file,
 805     )
 806
 807     exit(0)
 808
 809 ######################################################################
 810
 811 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 812
 813 # Compute the entropy of the training tokens
 814
 815 token_count = 0
 816 for input in task.batches(split="train"):
 817     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 818 token_probas = token_count / token_count.sum()
 819 entropy = -torch.xlogy(token_probas, token_probas).sum()
 820 train_set_perplexity = math.exp(entropy)
 821
 822 ######################################################################
 823 # A bit of paranoia never hurts
 824
 825 if args.max_percents_of_test_in_train >= 0:
 826
 827     def subsets_as_tuples(batches, cs):
 828         s = set()
 829         for batch in batches:
 830             for x in batch:
 831                 s.add(tuple([v.item() for v in x]))
 832                 if len(s) == cs:
 833                     yield s
 834                     s = set()
 835         yield s
 836
 837     nb_test, nb_in_train = 0, 0
 838     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 839         in_train = set()
 840         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 841             in_train.update(test_subset.intersection(train_subset))
 842         nb_in_train += len(in_train)
 843         nb_test += len(test_subset)
 844
 845     log_string(
 846         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 847     )
 848
 849     assert (
 850         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 851     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 852
 853 ##############################
 854
 855 if "calibrate" in sup_args:
 856     for input in task.batches(split="train", desc="calibrate"):
 857         input = input.to(device)
 858         output = model(mygpt.BracketedSequence(input)).x
 859
 860     for n, m in model.named_modules():
 861         for a in dir(m):
 862             x = getattr(m, a)
 863             if isinstance(x, mygpt.Calibrator):
 864                 print(f"####### ${n} | ${a} ########################")
 865                 mean, std = x.moments()
 866                 print("mean\n", mean, "\n")
 867                 print("std\n", std, "\n")
 868                 print(f"############################################\n\n")
 869
 870     exit(0)
 871
 872 ##############################
 873
 874 nb_samples_seen = 0
 875
 876 if nb_epochs_finished >= nb_epochs:
 877     task.produce_results(
 878         n_epoch=nb_epochs_finished,
 879         model=model,
 880         result_dir=args.result_dir,
 881         logger=log_string,
 882         deterministic_synthesis=args.deterministic_synthesis,
 883     )
 884
 885 time_pred_result = datetime.datetime.now()
 886
 887 it = 0
 888
 889 n_batch = 0
 890
 891 for n_epoch in range(nb_epochs_finished, nb_epochs):
 892     if args.optim == "sgd":
 893         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 894     elif args.optim == "adam":
 895         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 896     elif args.optim == "adamw":
 897         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 898     else:
 899         raise ValueError(f"Unknown optimizer {args.optim}.")
 900
 901     model.train()
 902
 903     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 904
 905     def add_memex(batches, proportion_memex):
 906         for input in batches:
 907             if torch.rand(1).item() < proportion_memex:
 908                 yield torch.cat(
 909                     [
 910                         input,
 911                         torch.full(
 912                             (input.size(0), 1), vocabulary_size - 1, device=input.device
 913                         ),
 914                         input,
 915                     ],
 916                     dim=1,
 917                 )
 918             yield input
 919
 920     train_batches = add_memex(task.batches(split="train"), args.proportion_memex)
 921
 922     for input in train_batches:
 923         model.reset_inner_loss()
 924         input = input.to(device)
 925
 926         output = model(mygpt.BracketedSequence(input)).x
 927         loss = F.cross_entropy(output.transpose(1, 2), input)
 928         inner_loss = model.get_inner_loss()
 929
 930         acc_train_loss += loss.item() * input.size(0)
 931         acc_train_inner_loss += inner_loss.item() * input.size(0)
 932
 933         nb_train_samples += input.size(0)
 934         nb_samples_seen += input.size(0)
 935
 936         total_loss = loss + (
 937             args.rho_inner_loss * inner_loss if args.rho_inner_loss > 0 else 0.0
 938         )
 939
 940         it += 1
 941         lr = get_lr(n_epoch, it)
 942         for param_group in optimizer.param_groups:
 943             param_group["lr"] = lr
 944
 945         # log_string(f"learning_rate {lr}")
 946
 947         optimizer.zero_grad()
 948         total_loss.backward()
 949         optimizer.step()
 950
 951         grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
 952
 953         loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
 954
 955         n_batch += 1
 956
 957     with torch.autograd.no_grad():
 958         model.eval()
 959
 960         nb_test_samples, acc_test_loss = 0, 0.0
 961
 962         for input in task.batches(split="test"):
 963             input = input.to(device)
 964
 965             output = model(mygpt.BracketedSequence(input)).x
 966             loss = F.cross_entropy(output.transpose(1, 2), input)
 967             acc_test_loss += loss.item() * input.size(0)
 968             nb_test_samples += input.size(0)
 969
 970         log_string(
 971             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}"
 972         )
 973
 974         task.produce_results(
 975             n_epoch=n_epoch,
 976             model=model,
 977             result_dir=args.result_dir,
 978             logger=log_string,
 979             deterministic_synthesis=args.deterministic_synthesis,
 980         )
 981
 982         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 983         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 984
 985         log_string(
 986             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 987         )
 988
 989         time_current_result = datetime.datetime.now()
 990         log_string(
 991             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 992         )
 993         time_pred_result = time_current_result
 994
 995     checkpoint = {
 996         "nb_epochs_finished": n_epoch + 1,
 997         "model_state": model.state_dict(),
 998         "rng_state": torch.get_rng_state(),
 999     }
1000
1001     if torch.cuda.is_available():
1002         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1003
1004     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1005     torch.save(checkpoint, checkpoint_name)
1006     log_string(f"saved checkpoint {checkpoint_name}")
1007
1008 ######################################################################