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("--memex_proba", type=float, default=0)
  91
  92 parser.add_argument("--memex_nb_epochs", type=float, default=1)
  93
  94 parser.add_argument("--dim_model", type=int, default=None)
  95
  96 parser.add_argument("--dim_keys", type=int, default=None)
  97
  98 parser.add_argument("--dim_hidden", type=int, default=None)
  99
 100 parser.add_argument("--nb_heads", type=int, default=None)
 101
 102 parser.add_argument("--nb_lines", type=int, default=None)
 103
 104 parser.add_argument("--caterpillar_height", type=int, default=None)
 105
 106 parser.add_argument("--gate_dropout_proba", type=float, default=0.0)
 107
 108 parser.add_argument("--gate_dropout_sync", type=str2bool, default=False)
 109
 110 parser.add_argument("--gate_dropout_replace", type=str2bool, default=False)
 111
 112 parser.add_argument("--rho_inner_loss", type=float, default=0.0)
 113
 114 parser.add_argument("--nb_blocks", type=int, default=None)
 115
 116 parser.add_argument("--dropout", type=float, default=0.1)
 117
 118 ########################################
 119
 120 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 121
 122 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 123
 124 parser.add_argument("--continue_training", action="store_true", default=False)
 125
 126 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 127
 128 ##############################
 129 # rpl options
 130
 131 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 132
 133 parser.add_argument("--rpl_max_input", type=int, default=9)
 134
 135 parser.add_argument("--rpl_prog_len", type=int, default=8)
 136
 137 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 138
 139 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 140
 141 ##############################
 142 # grid options
 143
 144 parser.add_argument("--grid_size", type=int, default=6)
 145
 146 parser.add_argument("--grid_nb_colors", type=int, default=6)
 147
 148 parser.add_argument("--grid_nb_shapes", type=int, default=6)
 149
 150 ##############################
 151 # picoclvr options
 152
 153 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 154
 155 parser.add_argument("--picoclvr_height", type=int, default=12)
 156
 157 parser.add_argument("--picoclvr_width", type=int, default=16)
 158
 159 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 160
 161 ##############################
 162 # Maze options
 163
 164 parser.add_argument("--maze_height", type=int, default=13)
 165
 166 parser.add_argument("--maze_width", type=int, default=21)
 167
 168 parser.add_argument("--maze_nb_walls", type=int, default=15)
 169
 170 ##############################
 171 # Snake options
 172
 173 parser.add_argument("--snake_height", type=int, default=9)
 174
 175 parser.add_argument("--snake_width", type=int, default=12)
 176
 177 parser.add_argument("--snake_nb_colors", type=int, default=5)
 178
 179 parser.add_argument("--snake_length", type=int, default=200)
 180
 181 ##############################
 182 # Stack options
 183
 184 parser.add_argument("--stack_nb_steps", type=int, default=100)
 185
 186 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 187
 188 parser.add_argument("--stack_nb_digits", type=int, default=3)
 189
 190 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 191
 192 ##############################
 193 # Expr options
 194
 195 parser.add_argument("--expr_nb_variables", type=int, default=5)
 196
 197 parser.add_argument("--expr_sequence_length", type=int, default=40)
 198
 199 parser.add_argument("--expr_operand_max", type=int, default=9)
 200
 201 parser.add_argument("--expr_result_max", type=int, default=99)
 202
 203 parser.add_argument("--expr_input_file", type=str, default=None)
 204
 205 ##############################
 206 # Memory
 207
 208 parser.add_argument("--memory_len_total", type=int, default=32)
 209
 210 ##############################
 211 # Mixing
 212
 213 parser.add_argument("--mixing_hard", action="store_true", default=False)
 214
 215 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 216
 217 ######################################################################
 218
 219 # args = parser.parse_args()
 220
 221 args, sup_args = parser.parse_known_args()
 222
 223 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 224
 225 if args.result_dir is None:
 226     args.result_dir = f"results_{args.task}_{args.model}"
 227
 228 ######################################################################
 229
 230 if not args.force_cpu and torch.cuda.is_available():
 231     device = torch.device("cuda")
 232     torch.backends.cuda.matmul.allow_tf32 = True
 233 else:
 234     device = torch.device("cpu")
 235
 236 ######################################################################
 237
 238 default_task_args = {
 239     "addition": {
 240         "model": "352M",
 241         "batch_size": 25,
 242         "nb_train_samples": 250000,
 243         "nb_test_samples": 10000,
 244     },
 245     "byheart": {
 246         "model": "37M",
 247         "batch_size": 25,
 248         "nb_train_samples": 50000,
 249         "nb_test_samples": 10000,
 250     },
 251     "expr": {
 252         "model": "352M",
 253         "batch_size": 25,
 254         "nb_train_samples": 2500000,
 255         "nb_test_samples": 10000,
 256     },
 257     "grid": {
 258         "model": "37M",
 259         "batch_size": 25,
 260         "nb_train_samples": 250000,
 261         "nb_test_samples": 10000,
 262     },
 263     "qmlp": {
 264         "model": "37M",
 265         "batch_size": 10,
 266         "nb_train_samples": 100000,
 267         "nb_test_samples": 1000,
 268     },
 269     "guessop": {
 270         "model": "352M",
 271         "batch_size": 25,
 272         "nb_train_samples": 1000000,
 273         "nb_test_samples": 10000,
 274     },
 275     "learnop": {
 276         "model": "37M",
 277         "batch_size": 25,
 278         "nb_train_samples": 50000,
 279         "nb_test_samples": 10000,
 280     },
 281     "maze": {
 282         "model": "37M",
 283         "batch_size": 5,
 284         "nb_train_samples": 100000,
 285         "nb_test_samples": 10000,
 286     },
 287     "picoclvr": {
 288         "model": "37M",
 289         "batch_size": 25,
 290         "nb_train_samples": 250000,
 291         "nb_test_samples": 10000,
 292     },
 293     "rpl": {
 294         "model": "352M",
 295         "batch_size": 5,
 296         "nb_train_samples": 2500000,
 297         "nb_test_samples": 10000,
 298     },
 299     "snake": {
 300         "model": "37M",
 301         "batch_size": 25,
 302         "nb_train_samples": 250000,
 303         "nb_test_samples": 10000,
 304     },
 305     "stack": {
 306         "model": "37M",
 307         "batch_size": 25,
 308         "nb_train_samples": 100000,
 309         "nb_test_samples": 1000,
 310     },
 311     "twotargets": {
 312         "model": "37M",
 313         "batch_size": 25,
 314         "nb_train_samples": 50000,
 315         "nb_test_samples": 10000,
 316     },
 317     "memory": {
 318         "model": "37M",
 319         "batch_size": 25,
 320         "nb_train_samples": 25000,
 321         "nb_test_samples": 10000,
 322     },
 323     "mixing": {
 324         "model": "37M",
 325         "batch_size": 25,
 326         "nb_train_samples": 250000,
 327         "nb_test_samples": 10000,
 328     },
 329     "mnist": {
 330         "model": "37M",
 331         "batch_size": 10,
 332         "nb_train_samples": 60000,
 333         "nb_test_samples": 10000,
 334     },
 335 }
 336
 337 if args.task in default_task_args:
 338     for k, v in default_task_args[args.task].items():
 339         if getattr(args, k) is None:
 340             setattr(args, k, v)
 341
 342 ######################################################################
 343
 344 default_model_args = {
 345     "17K": {
 346         "attention": "mha",
 347         "dim_model": 32,
 348         "dim_keys": 32,
 349         "dim_hidden": 32,
 350         "nb_heads": 2,
 351         "nb_blocks": 2,
 352     },
 353     "17K-C": {
 354         "attention": "caterpillar",
 355         "dim_model": 32,
 356         "dim_keys": 32,
 357         "dim_hidden": 32,
 358         "nb_heads": 2,
 359         "nb_lines": 16,
 360         "caterpillar_height": 4,
 361         "nb_blocks": 2,
 362     },
 363     "4M": {
 364         "attention": "mha",
 365         "dim_model": 256,
 366         "dim_keys": 32,
 367         "dim_hidden": 1024,
 368         "nb_heads": 4,
 369         "nb_blocks": 6,
 370     },
 371     "4M-C": {
 372         "attention": "caterpillar",
 373         "dim_model": 256,
 374         "dim_keys": 32,
 375         "dim_hidden": 1024,
 376         "nb_heads": 4,
 377         "nb_lines": 32,
 378         "caterpillar_height": 4,
 379         "nb_blocks": 6,
 380     },
 381     "37M": {
 382         "attention": "mha",
 383         "dim_model": 512,
 384         "dim_keys": 64,
 385         "dim_hidden": 2048,
 386         "nb_heads": 8,
 387         "nb_blocks": 12,
 388     },
 389     "37M-C": {
 390         "attention": "caterpillar",
 391         "dim_model": 512,
 392         "dim_keys": 64,
 393         "dim_hidden": 2048,
 394         "nb_heads": 8,
 395         "nb_lines": 256,
 396         "caterpillar_height": 32,
 397         "nb_blocks": 12,
 398     },
 399     "122M": {
 400         "attention": "mha",
 401         "dim_model": 768,
 402         "dim_keys": 64,
 403         "dim_hidden": 2048,
 404         "nb_heads": 8,
 405         "nb_blocks": 24,
 406     },
 407     "122M-C": {
 408         "attention": "caterpillar",
 409         "dim_model": 768,
 410         "dim_keys": 64,
 411         "dim_hidden": 2048,
 412         "nb_heads": 8,
 413         "nb_lines": 128,
 414         "nb_blocks": 24,
 415     },
 416     "352M": {
 417         "attention": "mha",
 418         "dim_model": 1024,
 419         "dim_keys": 64,
 420         "dim_hidden": 2048,
 421         "nb_heads": 8,
 422         "nb_blocks": 48,
 423     },
 424     "352M-C": {
 425         "attention": "caterpillar",
 426         "dim_model": 1024,
 427         "dim_keys": 64,
 428         "dim_hidden": 2048,
 429         "nb_heads": 8,
 430         "nb_lines": 128,
 431         "nb_blocks": 48,
 432     },
 433 }
 434
 435 if args.model in default_model_args:
 436     for k, v in default_model_args[args.model].items():
 437         if getattr(args, k) is None:
 438             setattr(args, k, v)
 439 else:
 440     raise ValueError(f"Unknown model {args.model}")
 441
 442 ######################################################################
 443
 444 try:
 445     os.mkdir(args.result_dir)
 446 except FileExistsError:
 447     if not args.continue_training:
 448         print(f"result directory {args.result_dir} already exists")
 449         exit(1)
 450
 451 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
 452
 453 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 454
 455 if args.seed >= 0:
 456     # torch.backends.cudnn.deterministic = True
 457     # torch.backends.cudnn.benchmark = False
 458     # torch.use_deterministic_algorithms(True)
 459     torch.manual_seed(args.seed)
 460     if torch.cuda.is_available():
 461         torch.cuda.manual_seed_all(args.seed)
 462
 463 ######################################################################
 464
 465
 466 def log_string(s):
 467     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 468
 469     if log_file is not None:
 470         log_file.write(t + s + "\n")
 471         log_file.flush()
 472
 473     print(t + s)
 474     sys.stdout.flush()
 475
 476
 477 with os.popen("sha256sum *.py") as f:
 478     for l in f:
 479         log_string(f"sha256sum {l.strip()}")
 480
 481 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 482 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 483
 484 log_string(f"argv {' '.join(sys.argv)}")
 485
 486 for n in vars(args):
 487     log_string(f"args.{n} {getattr(args, n)}")
 488
 489 for k, v in sup_args.items():
 490     log_string(f'sup_args["{k}"] "{v}"')
 491
 492
 493 ######################################################################
 494
 495
 496 def get_lr(n_epoch, it):
 497     if args.legacy_lr_schedule:
 498         # my crude scheduling to compare to previous baseline, added
 499         # warmup though
 500
 501         if it < args.nb_warmup_iter:
 502             return args.legacy_large_lr * it / args.nb_warmup_iter
 503         elif n_epoch < args.legacy_nb_epoch_large_lr:
 504             return args.legacy_large_lr
 505         else:
 506             return args.legacy_small_lr
 507
 508     # from nanoGPT
 509
 510     # 1) linear warmup for warmup_iter steps
 511     if it < args.nb_warmup_iter:
 512         return args.learning_rate * it / args.nb_warmup_iter
 513     # 2) if it > nb_decay_iter, return min learning rate
 514     if it > args.nb_decay_iter:
 515         return args.min_learning_rate
 516     # 3) in between, use cosine decay down to min learning rate
 517     decay_ratio = (it - args.nb_warmup_iter) / (
 518         args.nb_decay_iter - args.nb_warmup_iter
 519     )
 520     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 521     return args.min_learning_rate + coeff * (
 522         args.learning_rate - args.min_learning_rate
 523     )
 524
 525
 526 ######################################################################
 527
 528
 529 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 530
 531
 532 def picoclvr_pruner_horizontal_green(p):
 533     return not ("green" in p and ("left" in p or "right" in p))
 534
 535
 536 picoclvr_pruner_train = (
 537     picoclvr_pruner_horizontal_green
 538     if args.picocvlr_prune_properties in {"train+eval"}
 539     else None
 540 )
 541
 542 picoclvr_pruner_eval = (
 543     (lambda p: not picoclvr_pruner_horizontal_green(p))
 544     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 545     else None
 546 )
 547
 548 ######################################################################
 549
 550 device_data = device
 551
 552 if args.task == "byheart":
 553     task = tasks.SandBox(
 554         problem=problems.ProblemByHeart(),
 555         nb_train_samples=args.nb_train_samples,
 556         nb_test_samples=args.nb_test_samples,
 557         batch_size=args.batch_size,
 558         logger=log_string,
 559         device=device_data,
 560     )
 561     args.max_percents_of_test_in_train = -1
 562
 563 elif args.task == "learnop":
 564     task = tasks.SandBox(
 565         problem=problems.ProblemLearnOperator(),
 566         nb_train_samples=args.nb_train_samples,
 567         nb_test_samples=args.nb_test_samples,
 568         batch_size=args.batch_size,
 569         logger=log_string,
 570         device=device_data,
 571     )
 572
 573
 574 elif args.task == "guessop":
 575     task = tasks.SandBox(
 576         problem=problems.ProblemGuessOperator(),
 577         nb_train_samples=args.nb_train_samples,
 578         nb_test_samples=args.nb_test_samples,
 579         batch_size=args.batch_size,
 580         logger=log_string,
 581         device=device_data,
 582     )
 583
 584
 585 elif args.task == "twotargets":
 586     task = tasks.SandBox(
 587         problem=problems.ProblemTwoTargets(),
 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 == "memory":
 596     task = tasks.SandBox(
 597         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 598         nb_train_samples=args.nb_train_samples,
 599         nb_test_samples=args.nb_test_samples,
 600         batch_size=args.batch_size,
 601         logger=log_string,
 602         device=device_data,
 603     )
 604
 605 elif args.task == "mixing":
 606     task = tasks.SandBox(
 607         problem=problems.ProblemMixing(
 608             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 609         ),
 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 == "addition":
 618     task = tasks.SandBox(
 619         problem=problems.ProblemAddition(),
 620         nb_train_samples=args.nb_train_samples,
 621         nb_test_samples=args.nb_test_samples,
 622         batch_size=args.batch_size,
 623         logger=log_string,
 624         device=device_data,
 625     )
 626
 627 elif args.task == "picoclvr":
 628     task = tasks.PicoCLVR(
 629         nb_train_samples=args.nb_train_samples,
 630         nb_test_samples=args.nb_test_samples,
 631         batch_size=args.batch_size,
 632         height=args.picoclvr_height,
 633         width=args.picoclvr_width,
 634         nb_colors=args.picoclvr_nb_colors,
 635         logger=log_string,
 636         device=device_data,
 637         pruner_train=picoclvr_pruner_train,
 638         pruner_eval=picoclvr_pruner_eval,
 639     )
 640
 641 elif args.task == "mnist":
 642     task = tasks.MNIST(
 643         nb_train_samples=args.nb_train_samples,
 644         nb_test_samples=args.nb_test_samples,
 645         batch_size=args.batch_size,
 646         device=device_data,
 647     )
 648
 649 elif args.task == "maze":
 650     task = tasks.Maze(
 651         nb_train_samples=args.nb_train_samples,
 652         nb_test_samples=args.nb_test_samples,
 653         batch_size=args.batch_size,
 654         height=args.maze_height,
 655         width=args.maze_width,
 656         nb_walls=args.maze_nb_walls,
 657         device=device_data,
 658     )
 659
 660 elif args.task == "snake":
 661     task = tasks.Snake(
 662         nb_train_samples=args.nb_train_samples,
 663         nb_test_samples=args.nb_test_samples,
 664         batch_size=args.batch_size,
 665         height=args.snake_height,
 666         width=args.snake_width,
 667         nb_colors=args.snake_nb_colors,
 668         length=args.snake_length,
 669         prompt_length=args.snake_length // 2,
 670         device=device_data,
 671     )
 672
 673 elif args.task == "stack":
 674     task = tasks.Stack(
 675         nb_train_samples=args.nb_train_samples,
 676         nb_test_samples=args.nb_test_samples,
 677         batch_size=args.batch_size,
 678         logger=log_string,
 679         nb_steps=args.stack_nb_steps,
 680         nb_stacks=args.stack_nb_stacks,
 681         nb_digits=args.stack_nb_digits,
 682         fraction_values_for_train=args.stack_fraction_values_for_train,
 683         device=device_data,
 684     )
 685
 686 elif args.task == "expr":
 687     task = tasks.Expr(
 688         nb_train_samples=args.nb_train_samples,
 689         nb_test_samples=args.nb_test_samples,
 690         nb_variables=args.expr_nb_variables,
 691         sequence_length=args.expr_sequence_length,
 692         operand_max=args.expr_operand_max,
 693         result_max=args.expr_result_max,
 694         batch_size=args.batch_size,
 695         device=device_data,
 696     )
 697
 698 elif args.task == "rpl":
 699     task = tasks.RPL(
 700         nb_train_samples=args.nb_train_samples,
 701         nb_test_samples=args.nb_test_samples,
 702         batch_size=args.batch_size,
 703         nb_starting_values=args.rpl_nb_starting_values,
 704         max_input=args.rpl_max_input,
 705         prog_len=args.rpl_prog_len,
 706         nb_runs=args.rpl_nb_runs,
 707         no_prog=args.rpl_no_prog,
 708         logger=log_string,
 709         device=device_data,
 710     )
 711
 712 elif args.task == "grid":
 713     task = tasks.Grid(
 714         nb_train_samples=args.nb_train_samples,
 715         nb_test_samples=args.nb_test_samples,
 716         batch_size=args.batch_size,
 717         size=args.grid_size,
 718         nb_shapes=args.grid_nb_shapes,
 719         nb_colors=args.grid_nb_colors,
 720         logger=log_string,
 721         device=device_data,
 722     )
 723
 724 elif args.task == "qmlp":
 725     task = tasks.QMLP(
 726         nb_train_samples=args.nb_train_samples,
 727         nb_test_samples=args.nb_test_samples,
 728         batch_size=args.batch_size,
 729         result_dir=args.result_dir,
 730         logger=log_string,
 731         device=device_data,
 732     )
 733
 734 else:
 735     raise ValueError(f"Unknown task {args.task}")
 736
 737 ######################################################################
 738
 739 log_string(f"device {device}")
 740
 741 vocabulary_size = task.vocabulary_size()
 742
 743 if args.memex_proba > 0:
 744     vocabulary_size += 1
 745
 746 log_string(f"vocabulary_size {vocabulary_size}")
 747
 748 ##############################
 749
 750 model = mygpt.MyGPT(
 751     vocabulary_size=vocabulary_size,
 752     dim_model=args.dim_model,
 753     dim_keys=args.dim_keys,
 754     dim_hidden=args.dim_hidden,
 755     nb_heads=args.nb_heads,
 756     nb_lines=args.nb_lines,
 757     caterpillar_height=args.caterpillar_height,
 758     nb_blocks=args.nb_blocks,
 759     causal=True,
 760     dropout=args.dropout,
 761     attention_layer=args.attention,
 762     logger=log_string,
 763     args=args,
 764 )
 765
 766 model.to(device)
 767
 768 nb_parameters = sum(p.numel() for p in model.parameters())
 769 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 770
 771 ######################################################################
 772
 773 nb_epochs_finished = 0
 774
 775 if args.no_checkpoint:
 776     log_string(f"not trying to load checkpoint.")
 777
 778 else:
 779     try:
 780         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 781         checkpoint = torch.load(checkpoint_name)
 782         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 783         model.load_state_dict(checkpoint["model_state"])
 784         torch.set_rng_state(checkpoint["rng_state"])
 785         if torch.cuda.is_available():
 786             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 787
 788         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 789
 790     except FileNotFoundError:
 791         log_string("starting from scratch.")
 792
 793     except:
 794         log_string("error when loading the checkpoint.")
 795         exit(1)
 796
 797 ######################################################################
 798
 799 if args.task == "expr" and args.expr_input_file is not None:
 800     task.produce_results(
 801         n_epoch=nb_epochs_finished,
 802         model=model,
 803         result_dir=args.result_dir,
 804         logger=log_string,
 805         deterministic_synthesis=args.deterministic_synthesis,
 806         input_file=args.expr_input_file,
 807     )
 808
 809     exit(0)
 810
 811 ######################################################################
 812
 813 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 814
 815 # Compute the entropy of the training tokens
 816
 817 token_count = 0
 818 for input in task.batches(split="train"):
 819     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 820 token_probas = token_count / token_count.sum()
 821 entropy = -torch.xlogy(token_probas, token_probas).sum()
 822 train_set_perplexity = math.exp(entropy)
 823
 824 ######################################################################
 825 # A bit of paranoia never hurts
 826
 827 if args.max_percents_of_test_in_train >= 0:
 828
 829     def subsets_as_tuples(batches, cs):
 830         s = set()
 831         for batch in batches:
 832             for x in batch:
 833                 s.add(tuple([v.item() for v in x]))
 834                 if len(s) == cs:
 835                     yield s
 836                     s = set()
 837         yield s
 838
 839     nb_test, nb_in_train = 0, 0
 840     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 841         in_train = set()
 842         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 843             in_train.update(test_subset.intersection(train_subset))
 844         nb_in_train += len(in_train)
 845         nb_test += len(test_subset)
 846
 847     log_string(
 848         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 849     )
 850
 851     assert (
 852         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 853     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 854
 855 ##############################
 856
 857 if "calibrate" in sup_args:
 858     for input in task.batches(split="train", desc="calibrate"):
 859         input = input.to(device)
 860         output = model(mygpt.BracketedSequence(input)).x
 861
 862     for n, m in model.named_modules():
 863         for a in dir(m):
 864             x = getattr(m, a)
 865             if isinstance(x, mygpt.Calibrator):
 866                 print(f"####### ${n} | ${a} ########################")
 867                 mean, std = x.moments()
 868                 print("mean\n", mean, "\n")
 869                 print("std\n", std, "\n")
 870                 print(f"############################################\n\n")
 871
 872     exit(0)
 873
 874 ##############################
 875
 876 nb_samples_seen = 0
 877
 878 if nb_epochs_finished >= nb_epochs:
 879     task.produce_results(
 880         n_epoch=nb_epochs_finished,
 881         model=model,
 882         result_dir=args.result_dir,
 883         logger=log_string,
 884         deterministic_synthesis=args.deterministic_synthesis,
 885     )
 886
 887 time_pred_result = datetime.datetime.now()
 888
 889 it = 0
 890
 891 n_batch = 0
 892
 893 for n_epoch in range(nb_epochs_finished, nb_epochs):
 894     if args.optim == "sgd":
 895         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 896     elif args.optim == "adam":
 897         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 898     elif args.optim == "adamw":
 899         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 900     else:
 901         raise ValueError(f"Unknown optimizer {args.optim}.")
 902
 903     model.train()
 904
 905     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 906
 907     def add_memex(batches, memex_proba):
 908         for input in batches:
 909             if torch.rand(1).item() < memex_proba:
 910                 sep = torch.full(
 911                     (input.size(0), 1), vocabulary_size - 1, device=input.device
 912                 )
 913
 914                 yield torch.cat(
 915                     [
 916                         input,
 917                         sep,
 918                         input,
 919                     ],
 920                     dim=1,
 921                 )
 922             yield input
 923
 924     train_batches = add_memex(
 925         task.batches(split="train"),
 926         args.memex_proba if n_epoch < args.memex_nb_epochs else 0.0,
 927     )
 928
 929     for input in train_batches:
 930         model.reset_inner_loss()
 931         input = input.to(device)
 932
 933         output = model(mygpt.BracketedSequence(input)).x
 934         loss = F.cross_entropy(output.transpose(1, 2), input)
 935         inner_loss = model.get_inner_loss()
 936
 937         acc_train_loss += loss.item() * input.size(0)
 938         acc_train_inner_loss += inner_loss.item() * input.size(0)
 939
 940         nb_train_samples += input.size(0)
 941         nb_samples_seen += input.size(0)
 942
 943         total_loss = loss + (
 944             args.rho_inner_loss * inner_loss if args.rho_inner_loss > 0 else 0.0
 945         )
 946
 947         it += 1
 948         lr = get_lr(n_epoch, it)
 949         for param_group in optimizer.param_groups:
 950             param_group["lr"] = lr
 951
 952         # log_string(f"learning_rate {lr}")
 953
 954         optimizer.zero_grad()
 955         total_loss.backward()
 956         optimizer.step()
 957
 958         grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
 959
 960         loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
 961
 962         n_batch += 1
 963
 964     with torch.autograd.no_grad():
 965         model.eval()
 966
 967         nb_test_samples, acc_test_loss = 0, 0.0
 968
 969         for input in task.batches(split="test"):
 970             input = input.to(device)
 971
 972             output = model(mygpt.BracketedSequence(input)).x
 973             loss = F.cross_entropy(output.transpose(1, 2), input)
 974             acc_test_loss += loss.item() * input.size(0)
 975             nb_test_samples += input.size(0)
 976
 977         log_string(
 978             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}"
 979         )
 980
 981         task.produce_results(
 982             n_epoch=n_epoch,
 983             model=model,
 984             result_dir=args.result_dir,
 985             logger=log_string,
 986             deterministic_synthesis=args.deterministic_synthesis,
 987         )
 988
 989         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 990         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 991
 992         log_string(
 993             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 994         )
 995
 996         time_current_result = datetime.datetime.now()
 997         log_string(
 998             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 999         )
1000         time_pred_result = time_current_result
1001
1002     checkpoint = {
1003         "nb_epochs_finished": n_epoch + 1,
1004         "model_state": model.state_dict(),
1005         "rng_state": torch.get_rng_state(),
1006     }
1007
1008     if torch.cuda.is_available():
1009         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1010
1011     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1012     torch.save(checkpoint, checkpoint_name)
1013     log_string(f"saved checkpoint {checkpoint_name}")
1014
1015 ######################################################################