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("--gate_dropout_proba", type=float, default=0.0)
 103
 104 parser.add_argument("--gate_dropout_sync", type=bool, default=False)
 105
 106 parser.add_argument("--rho_inner_loss", type=float, default=0.0)
 107
 108 parser.add_argument("--nb_blocks", type=int, default=None)
 109
 110 parser.add_argument("--dropout", type=float, default=0.1)
 111
 112 ########################################
 113
 114 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 115
 116 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 117
 118 parser.add_argument("--continue_training", action="store_true", default=False)
 119
 120 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 121
 122 ##############################
 123 # rpl options
 124
 125 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 126
 127 parser.add_argument("--rpl_max_input", type=int, default=9)
 128
 129 parser.add_argument("--rpl_prog_len", type=int, default=8)
 130
 131 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 132
 133 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 134
 135 ##############################
 136 # grid options
 137
 138 parser.add_argument("--grid_size", type=int, default=6)
 139
 140 parser.add_argument("--grid_nb_colors", type=int, default=6)
 141
 142 parser.add_argument("--grid_nb_shapes", type=int, default=6)
 143
 144 ##############################
 145 # picoclvr options
 146
 147 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 148
 149 parser.add_argument("--picoclvr_height", type=int, default=12)
 150
 151 parser.add_argument("--picoclvr_width", type=int, default=16)
 152
 153 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 154
 155 ##############################
 156 # Maze options
 157
 158 parser.add_argument("--maze_height", type=int, default=13)
 159
 160 parser.add_argument("--maze_width", type=int, default=21)
 161
 162 parser.add_argument("--maze_nb_walls", type=int, default=15)
 163
 164 ##############################
 165 # Snake options
 166
 167 parser.add_argument("--snake_height", type=int, default=9)
 168
 169 parser.add_argument("--snake_width", type=int, default=12)
 170
 171 parser.add_argument("--snake_nb_colors", type=int, default=5)
 172
 173 parser.add_argument("--snake_length", type=int, default=200)
 174
 175 ##############################
 176 # Stack options
 177
 178 parser.add_argument("--stack_nb_steps", type=int, default=100)
 179
 180 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 181
 182 parser.add_argument("--stack_nb_digits", type=int, default=3)
 183
 184 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 185
 186 ##############################
 187 # Expr options
 188
 189 parser.add_argument("--expr_nb_variables", type=int, default=5)
 190
 191 parser.add_argument("--expr_sequence_length", type=int, default=40)
 192
 193 parser.add_argument("--expr_operand_max", type=int, default=9)
 194
 195 parser.add_argument("--expr_result_max", type=int, default=99)
 196
 197 parser.add_argument("--expr_input_file", type=str, default=None)
 198
 199 ##############################
 200 # Memory
 201
 202 parser.add_argument("--memory_len_total", type=int, default=32)
 203
 204 ##############################
 205 # Mixing
 206
 207 parser.add_argument("--mixing_hard", action="store_true", default=False)
 208
 209 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 210
 211 ######################################################################
 212
 213 # args = parser.parse_args()
 214
 215 args, sup_args = parser.parse_known_args()
 216
 217 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 218
 219 if args.result_dir is None:
 220     args.result_dir = f"results_{args.task}_{args.model}"
 221
 222 ######################################################################
 223
 224 if not args.force_cpu and torch.cuda.is_available():
 225     device = torch.device("cuda")
 226     torch.backends.cuda.matmul.allow_tf32 = True
 227 else:
 228     device = torch.device("cpu")
 229
 230 ######################################################################
 231
 232 default_task_args = {
 233     "addition": {
 234         "model": "352M",
 235         "batch_size": 25,
 236         "nb_train_samples": 250000,
 237         "nb_test_samples": 10000,
 238     },
 239     "byheart": {
 240         "model": "37M",
 241         "batch_size": 25,
 242         "nb_train_samples": 50000,
 243         "nb_test_samples": 10000,
 244     },
 245     "expr": {
 246         "model": "352M",
 247         "batch_size": 25,
 248         "nb_train_samples": 2500000,
 249         "nb_test_samples": 10000,
 250     },
 251     "grid": {
 252         "model": "37M",
 253         "batch_size": 25,
 254         "nb_train_samples": 250000,
 255         "nb_test_samples": 10000,
 256     },
 257     "qmlp": {
 258         "model": "37M",
 259         "batch_size": 10,
 260         "nb_train_samples": 100000,
 261         "nb_test_samples": 1000,
 262     },
 263     "guessop": {
 264         "model": "352M",
 265         "batch_size": 25,
 266         "nb_train_samples": 1000000,
 267         "nb_test_samples": 10000,
 268     },
 269     "learnop": {
 270         "model": "37M",
 271         "batch_size": 25,
 272         "nb_train_samples": 50000,
 273         "nb_test_samples": 10000,
 274     },
 275     "maze": {
 276         "model": "37M",
 277         "batch_size": 5,
 278         "nb_train_samples": 100000,
 279         "nb_test_samples": 10000,
 280     },
 281     "picoclvr": {
 282         "model": "37M",
 283         "batch_size": 25,
 284         "nb_train_samples": 250000,
 285         "nb_test_samples": 10000,
 286     },
 287     "rpl": {
 288         "model": "352M",
 289         "batch_size": 5,
 290         "nb_train_samples": 2500000,
 291         "nb_test_samples": 10000,
 292     },
 293     "snake": {
 294         "model": "37M",
 295         "batch_size": 25,
 296         "nb_train_samples": 250000,
 297         "nb_test_samples": 10000,
 298     },
 299     "stack": {
 300         "model": "37M",
 301         "batch_size": 25,
 302         "nb_train_samples": 100000,
 303         "nb_test_samples": 1000,
 304     },
 305     "twotargets": {
 306         "model": "37M",
 307         "batch_size": 25,
 308         "nb_train_samples": 50000,
 309         "nb_test_samples": 10000,
 310     },
 311     "memory": {
 312         "model": "37M",
 313         "batch_size": 25,
 314         "nb_train_samples": 25000,
 315         "nb_test_samples": 10000,
 316     },
 317     "mixing": {
 318         "model": "37M",
 319         "batch_size": 25,
 320         "nb_train_samples": 250000,
 321         "nb_test_samples": 10000,
 322     },
 323     "mnist": {
 324         "model": "37M",
 325         "batch_size": 10,
 326         "nb_train_samples": 60000,
 327         "nb_test_samples": 10000,
 328     },
 329 }
 330
 331 if args.task in default_task_args:
 332     for k, v in default_task_args[args.task].items():
 333         if getattr(args, k) is None:
 334             setattr(args, k, v)
 335
 336 ######################################################################
 337
 338 default_model_args = {
 339     "17K": {
 340         "attention": "mha",
 341         "dim_model": 32,
 342         "dim_keys": 32,
 343         "dim_hidden": 32,
 344         "nb_heads": 2,
 345         "nb_blocks": 2,
 346     },
 347     "17K-C": {
 348         "attention": "caterpillar",
 349         "dim_model": 32,
 350         "dim_keys": 32,
 351         "dim_hidden": 32,
 352         "nb_heads": 2,
 353         "nb_lines": 16,
 354         "caterpillar_height": 4,
 355         "nb_blocks": 2,
 356     },
 357     "4M": {
 358         "attention": "mha",
 359         "dim_model": 256,
 360         "dim_keys": 32,
 361         "dim_hidden": 1024,
 362         "nb_heads": 4,
 363         "nb_blocks": 6,
 364     },
 365     "4M-C": {
 366         "attention": "caterpillar",
 367         "dim_model": 256,
 368         "dim_keys": 32,
 369         "dim_hidden": 1024,
 370         "nb_heads": 4,
 371         "nb_lines": 32,
 372         "caterpillar_height": 4,
 373         "nb_blocks": 6,
 374     },
 375     "37M": {
 376         "attention": "mha",
 377         "dim_model": 512,
 378         "dim_keys": 64,
 379         "dim_hidden": 2048,
 380         "nb_heads": 8,
 381         "nb_blocks": 12,
 382     },
 383     "37M-C": {
 384         "attention": "caterpillar",
 385         "dim_model": 512,
 386         "dim_keys": 64,
 387         "dim_hidden": 2048,
 388         "nb_heads": 8,
 389         "nb_lines": 256,
 390         "caterpillar_height": 32,
 391         "nb_blocks": 12,
 392     },
 393     "122M": {
 394         "attention": "mha",
 395         "dim_model": 768,
 396         "dim_keys": 64,
 397         "dim_hidden": 2048,
 398         "nb_heads": 8,
 399         "nb_blocks": 24,
 400     },
 401     "122M-C": {
 402         "attention": "caterpillar",
 403         "dim_model": 768,
 404         "dim_keys": 64,
 405         "dim_hidden": 2048,
 406         "nb_heads": 8,
 407         "nb_lines": 128,
 408         "nb_blocks": 24,
 409     },
 410     "352M": {
 411         "attention": "mha",
 412         "dim_model": 1024,
 413         "dim_keys": 64,
 414         "dim_hidden": 2048,
 415         "nb_heads": 8,
 416         "nb_blocks": 48,
 417     },
 418     "352M-C": {
 419         "attention": "caterpillar",
 420         "dim_model": 1024,
 421         "dim_keys": 64,
 422         "dim_hidden": 2048,
 423         "nb_heads": 8,
 424         "nb_lines": 128,
 425         "nb_blocks": 48,
 426     },
 427 }
 428
 429 if args.model in default_model_args:
 430     for k, v in default_model_args[args.model].items():
 431         if getattr(args, k) is None:
 432             setattr(args, k, v)
 433 else:
 434     raise ValueError(f"Unknown model {args.model}")
 435
 436 ######################################################################
 437
 438 try:
 439     os.mkdir(args.result_dir)
 440 except FileExistsError:
 441     if not args.continue_training:
 442         print(f"result directory {args.result_dir} already exists")
 443         exit(1)
 444
 445 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
 446
 447 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 448
 449 if args.seed >= 0:
 450     # torch.backends.cudnn.deterministic = True
 451     # torch.backends.cudnn.benchmark = False
 452     # torch.use_deterministic_algorithms(True)
 453     torch.manual_seed(args.seed)
 454     if torch.cuda.is_available():
 455         torch.cuda.manual_seed_all(args.seed)
 456
 457 ######################################################################
 458
 459
 460 def log_string(s):
 461     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 462
 463     if log_file is not None:
 464         log_file.write(t + s + "\n")
 465         log_file.flush()
 466
 467     print(t + s)
 468     sys.stdout.flush()
 469
 470
 471 with os.popen("sha256sum *.py") as f:
 472     for l in f:
 473         log_string(f"sha256sum {l.strip()}")
 474
 475 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 476 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 477
 478 log_string(f"argv {' '.join(sys.argv)}")
 479
 480 for n in vars(args):
 481     log_string(f"args.{n} {getattr(args, n)}")
 482
 483 for k, v in sup_args.items():
 484     log_string(f'sup_args["{k}"] "{v}"')
 485
 486
 487 ######################################################################
 488
 489
 490 def get_lr(n_epoch, it):
 491     if args.legacy_lr_schedule:
 492         # my crude scheduling to compare to previous baseline, added
 493         # warmup though
 494
 495         if it < args.nb_warmup_iter:
 496             return args.legacy_large_lr * it / args.nb_warmup_iter
 497         elif n_epoch < args.legacy_nb_epoch_large_lr:
 498             return args.legacy_large_lr
 499         else:
 500             return args.legacy_small_lr
 501
 502     # from nanoGPT
 503
 504     # 1) linear warmup for warmup_iter steps
 505     if it < args.nb_warmup_iter:
 506         return args.learning_rate * it / args.nb_warmup_iter
 507     # 2) if it > nb_decay_iter, return min learning rate
 508     if it > args.nb_decay_iter:
 509         return args.min_learning_rate
 510     # 3) in between, use cosine decay down to min learning rate
 511     decay_ratio = (it - args.nb_warmup_iter) / (
 512         args.nb_decay_iter - args.nb_warmup_iter
 513     )
 514     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 515     return args.min_learning_rate + coeff * (
 516         args.learning_rate - args.min_learning_rate
 517     )
 518
 519
 520 ######################################################################
 521
 522
 523 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 524
 525
 526 def picoclvr_pruner_horizontal_green(p):
 527     return not ("green" in p and ("left" in p or "right" in p))
 528
 529
 530 picoclvr_pruner_train = (
 531     picoclvr_pruner_horizontal_green
 532     if args.picocvlr_prune_properties in {"train+eval"}
 533     else None
 534 )
 535
 536 picoclvr_pruner_eval = (
 537     (lambda p: not picoclvr_pruner_horizontal_green(p))
 538     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 539     else None
 540 )
 541
 542 ######################################################################
 543
 544 device_data = device
 545
 546 if args.task == "byheart":
 547     task = tasks.SandBox(
 548         problem=problems.ProblemByHeart(),
 549         nb_train_samples=args.nb_train_samples,
 550         nb_test_samples=args.nb_test_samples,
 551         batch_size=args.batch_size,
 552         logger=log_string,
 553         device=device_data,
 554     )
 555     args.max_percents_of_test_in_train = -1
 556
 557 elif args.task == "learnop":
 558     task = tasks.SandBox(
 559         problem=problems.ProblemLearnOperator(),
 560         nb_train_samples=args.nb_train_samples,
 561         nb_test_samples=args.nb_test_samples,
 562         batch_size=args.batch_size,
 563         logger=log_string,
 564         device=device_data,
 565     )
 566
 567
 568 elif args.task == "guessop":
 569     task = tasks.SandBox(
 570         problem=problems.ProblemGuessOperator(),
 571         nb_train_samples=args.nb_train_samples,
 572         nb_test_samples=args.nb_test_samples,
 573         batch_size=args.batch_size,
 574         logger=log_string,
 575         device=device_data,
 576     )
 577
 578
 579 elif args.task == "twotargets":
 580     task = tasks.SandBox(
 581         problem=problems.ProblemTwoTargets(),
 582         nb_train_samples=args.nb_train_samples,
 583         nb_test_samples=args.nb_test_samples,
 584         batch_size=args.batch_size,
 585         logger=log_string,
 586         device=device_data,
 587     )
 588
 589 elif args.task == "memory":
 590     task = tasks.SandBox(
 591         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 592         nb_train_samples=args.nb_train_samples,
 593         nb_test_samples=args.nb_test_samples,
 594         batch_size=args.batch_size,
 595         logger=log_string,
 596         device=device_data,
 597     )
 598
 599 elif args.task == "mixing":
 600     task = tasks.SandBox(
 601         problem=problems.ProblemMixing(
 602             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 603         ),
 604         nb_train_samples=args.nb_train_samples,
 605         nb_test_samples=args.nb_test_samples,
 606         batch_size=args.batch_size,
 607         logger=log_string,
 608         device=device_data,
 609     )
 610
 611 elif args.task == "addition":
 612     task = tasks.SandBox(
 613         problem=problems.ProblemAddition(),
 614         nb_train_samples=args.nb_train_samples,
 615         nb_test_samples=args.nb_test_samples,
 616         batch_size=args.batch_size,
 617         logger=log_string,
 618         device=device_data,
 619     )
 620
 621 elif args.task == "picoclvr":
 622     task = tasks.PicoCLVR(
 623         nb_train_samples=args.nb_train_samples,
 624         nb_test_samples=args.nb_test_samples,
 625         batch_size=args.batch_size,
 626         height=args.picoclvr_height,
 627         width=args.picoclvr_width,
 628         nb_colors=args.picoclvr_nb_colors,
 629         logger=log_string,
 630         device=device_data,
 631         pruner_train=picoclvr_pruner_train,
 632         pruner_eval=picoclvr_pruner_eval,
 633     )
 634
 635 elif args.task == "mnist":
 636     task = tasks.MNIST(
 637         nb_train_samples=args.nb_train_samples,
 638         nb_test_samples=args.nb_test_samples,
 639         batch_size=args.batch_size,
 640         device=device_data,
 641     )
 642
 643 elif args.task == "maze":
 644     task = tasks.Maze(
 645         nb_train_samples=args.nb_train_samples,
 646         nb_test_samples=args.nb_test_samples,
 647         batch_size=args.batch_size,
 648         height=args.maze_height,
 649         width=args.maze_width,
 650         nb_walls=args.maze_nb_walls,
 651         device=device_data,
 652     )
 653
 654 elif args.task == "snake":
 655     task = tasks.Snake(
 656         nb_train_samples=args.nb_train_samples,
 657         nb_test_samples=args.nb_test_samples,
 658         batch_size=args.batch_size,
 659         height=args.snake_height,
 660         width=args.snake_width,
 661         nb_colors=args.snake_nb_colors,
 662         length=args.snake_length,
 663         prompt_length=args.snake_length // 2,
 664         device=device_data,
 665     )
 666
 667 elif args.task == "stack":
 668     task = tasks.Stack(
 669         nb_train_samples=args.nb_train_samples,
 670         nb_test_samples=args.nb_test_samples,
 671         batch_size=args.batch_size,
 672         logger=log_string,
 673         nb_steps=args.stack_nb_steps,
 674         nb_stacks=args.stack_nb_stacks,
 675         nb_digits=args.stack_nb_digits,
 676         fraction_values_for_train=args.stack_fraction_values_for_train,
 677         device=device_data,
 678     )
 679
 680 elif args.task == "expr":
 681     task = tasks.Expr(
 682         nb_train_samples=args.nb_train_samples,
 683         nb_test_samples=args.nb_test_samples,
 684         nb_variables=args.expr_nb_variables,
 685         sequence_length=args.expr_sequence_length,
 686         operand_max=args.expr_operand_max,
 687         result_max=args.expr_result_max,
 688         batch_size=args.batch_size,
 689         device=device_data,
 690     )
 691
 692 elif args.task == "rpl":
 693     task = tasks.RPL(
 694         nb_train_samples=args.nb_train_samples,
 695         nb_test_samples=args.nb_test_samples,
 696         batch_size=args.batch_size,
 697         nb_starting_values=args.rpl_nb_starting_values,
 698         max_input=args.rpl_max_input,
 699         prog_len=args.rpl_prog_len,
 700         nb_runs=args.rpl_nb_runs,
 701         no_prog=args.rpl_no_prog,
 702         logger=log_string,
 703         device=device_data,
 704     )
 705
 706 elif args.task == "grid":
 707     task = tasks.Grid(
 708         nb_train_samples=args.nb_train_samples,
 709         nb_test_samples=args.nb_test_samples,
 710         batch_size=args.batch_size,
 711         size=args.grid_size,
 712         nb_shapes=args.grid_nb_shapes,
 713         nb_colors=args.grid_nb_colors,
 714         logger=log_string,
 715         device=device_data,
 716     )
 717
 718 elif args.task == "qmlp":
 719     task = tasks.QMLP(
 720         nb_train_samples=args.nb_train_samples,
 721         nb_test_samples=args.nb_test_samples,
 722         batch_size=args.batch_size,
 723         result_dir=args.result_dir,
 724         logger=log_string,
 725         device=device_data,
 726     )
 727
 728 else:
 729     raise ValueError(f"Unknown task {args.task}")
 730
 731 ######################################################################
 732
 733 log_string(f"device {device}")
 734
 735 vocabulary_size = task.vocabulary_size()
 736
 737 log_string(f"vocabulary_size {vocabulary_size}")
 738
 739 ##############################
 740
 741 model = mygpt.MyGPT(
 742     vocabulary_size=vocabulary_size,
 743     dim_model=args.dim_model,
 744     dim_keys=args.dim_keys,
 745     dim_hidden=args.dim_hidden,
 746     nb_heads=args.nb_heads,
 747     nb_lines=args.nb_lines,
 748     caterpillar_height=args.caterpillar_height,
 749     nb_blocks=args.nb_blocks,
 750     causal=True,
 751     dropout=args.dropout,
 752     attention_layer=args.attention,
 753     logger=log_string,
 754     args=args,
 755 )
 756
 757 model.to(device)
 758
 759 nb_parameters = sum(p.numel() for p in model.parameters())
 760 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 761
 762 ######################################################################
 763
 764 nb_epochs_finished = 0
 765
 766 if args.no_checkpoint:
 767     log_string(f"not trying to load checkpoint.")
 768
 769 else:
 770     try:
 771         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 772         checkpoint = torch.load(checkpoint_name)
 773         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 774         model.load_state_dict(checkpoint["model_state"])
 775         torch.set_rng_state(checkpoint["rng_state"])
 776         if torch.cuda.is_available():
 777             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 778
 779         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 780
 781     except FileNotFoundError:
 782         log_string("starting from scratch.")
 783
 784     except:
 785         log_string("error when loading the checkpoint.")
 786         exit(1)
 787
 788 ######################################################################
 789
 790 if args.task == "expr" and args.expr_input_file is not None:
 791     task.produce_results(
 792         n_epoch=nb_epochs_finished,
 793         model=model,
 794         result_dir=args.result_dir,
 795         logger=log_string,
 796         deterministic_synthesis=args.deterministic_synthesis,
 797         input_file=args.expr_input_file,
 798     )
 799
 800     exit(0)
 801
 802 ######################################################################
 803
 804 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 805
 806 # Compute the entropy of the training tokens
 807
 808 token_count = 0
 809 for input in task.batches(split="train"):
 810     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 811 token_probas = token_count / token_count.sum()
 812 entropy = -torch.xlogy(token_probas, token_probas).sum()
 813 train_set_perplexity = math.exp(entropy)
 814
 815 ######################################################################
 816 # A bit of paranoia never hurts
 817
 818 if args.max_percents_of_test_in_train >= 0:
 819
 820     def subsets_as_tuples(batches, cs):
 821         s = set()
 822         for batch in batches:
 823             for x in batch:
 824                 s.add(tuple([v.item() for v in x]))
 825                 if len(s) == cs:
 826                     yield s
 827                     s = set()
 828         yield s
 829
 830     nb_test, nb_in_train = 0, 0
 831     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 832         in_train = set()
 833         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 834             in_train.update(test_subset.intersection(train_subset))
 835         nb_in_train += len(in_train)
 836         nb_test += len(test_subset)
 837
 838     log_string(
 839         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 840     )
 841
 842     assert (
 843         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 844     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 845
 846 ##############################
 847
 848 if "calibrate" in sup_args:
 849     for input in task.batches(split="train", desc="calibrate"):
 850         input = input.to(device)
 851         output = model(mygpt.BracketedSequence(input)).x
 852
 853     for n, m in model.named_modules():
 854         for a in dir(m):
 855             x = getattr(m, a)
 856             if isinstance(x, mygpt.Calibrator):
 857                 print(f"####### ${n} | ${a} ########################")
 858                 mean, std = x.moments()
 859                 print("mean\n", mean, "\n")
 860                 print("std\n", std, "\n")
 861                 print(f"############################################\n\n")
 862
 863     exit(0)
 864
 865 ##############################
 866
 867 nb_samples_seen = 0
 868
 869 if nb_epochs_finished >= nb_epochs:
 870     task.produce_results(
 871         n_epoch=nb_epochs_finished,
 872         model=model,
 873         result_dir=args.result_dir,
 874         logger=log_string,
 875         deterministic_synthesis=args.deterministic_synthesis,
 876     )
 877
 878 time_pred_result = datetime.datetime.now()
 879
 880 it = 0
 881
 882 n_batch = 0
 883
 884 for n_epoch in range(nb_epochs_finished, nb_epochs):
 885     if args.optim == "sgd":
 886         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 887     elif args.optim == "adam":
 888         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 889     elif args.optim == "adamw":
 890         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 891     else:
 892         raise ValueError(f"Unknown optimizer {args.optim}.")
 893
 894     model.train()
 895
 896     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 897
 898     for input in task.batches(split="train"):
 899         model.reset_inner_loss()
 900         input = input.to(device)
 901
 902         output = model(mygpt.BracketedSequence(input)).x
 903         loss = F.cross_entropy(output.transpose(1, 2), input)
 904         inner_loss = model.get_inner_loss()
 905
 906         acc_train_loss += loss.item() * input.size(0)
 907         acc_train_inner_loss += inner_loss.item() * input.size(0)
 908
 909         nb_train_samples += input.size(0)
 910         nb_samples_seen += input.size(0)
 911
 912         total_loss = loss + (
 913             args.rho_inner_loss * inner_loss if args.rho_inner_loss > 0 else 0.0
 914         )
 915
 916         it += 1
 917         lr = get_lr(n_epoch, it)
 918         for param_group in optimizer.param_groups:
 919             param_group["lr"] = lr
 920
 921         # log_string(f"learning_rate {lr}")
 922
 923         optimizer.zero_grad()
 924         total_loss.backward()
 925         optimizer.step()
 926
 927         grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
 928
 929         loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
 930
 931         n_batch += 1
 932
 933     with torch.autograd.no_grad():
 934         model.eval()
 935
 936         nb_test_samples, acc_test_loss = 0, 0.0
 937
 938         for input in task.batches(split="test"):
 939             input = input.to(device)
 940
 941             output = model(mygpt.BracketedSequence(input)).x
 942             loss = F.cross_entropy(output.transpose(1, 2), input)
 943             acc_test_loss += loss.item() * input.size(0)
 944             nb_test_samples += input.size(0)
 945
 946         log_string(
 947             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}"
 948         )
 949
 950         task.produce_results(
 951             n_epoch=n_epoch,
 952             model=model,
 953             result_dir=args.result_dir,
 954             logger=log_string,
 955             deterministic_synthesis=args.deterministic_synthesis,
 956         )
 957
 958         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 959         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 960
 961         log_string(
 962             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 963         )
 964
 965         time_current_result = datetime.datetime.now()
 966         log_string(
 967             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 968         )
 969         time_pred_result = time_current_result
 970
 971     checkpoint = {
 972         "nb_epochs_finished": n_epoch + 1,
 973         "model_state": model.state_dict(),
 974         "rng_state": torch.get_rng_state(),
 975     }
 976
 977     if torch.cuda.is_available():
 978         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 979
 980     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 981     torch.save(checkpoint, checkpoint_name)
 982     log_string(f"saved checkpoint {checkpoint_name}")
 983
 984 ######################################################################