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=str2bool, default=True)
 105
 106 parser.add_argument("--gate_dropout_replace", type=str2bool, default=True)
 107
 108 parser.add_argument("--rho_inner_loss", type=float, default=0.0)
 109
 110 parser.add_argument("--nb_blocks", type=int, default=None)
 111
 112 parser.add_argument("--dropout", type=float, default=0.1)
 113
 114 ########################################
 115
 116 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 117
 118 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 119
 120 parser.add_argument("--continue_training", action="store_true", default=False)
 121
 122 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 123
 124 ##############################
 125 # rpl options
 126
 127 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 128
 129 parser.add_argument("--rpl_max_input", type=int, default=9)
 130
 131 parser.add_argument("--rpl_prog_len", type=int, default=8)
 132
 133 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 134
 135 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 136
 137 ##############################
 138 # grid options
 139
 140 parser.add_argument("--grid_size", type=int, default=6)
 141
 142 parser.add_argument("--grid_nb_colors", type=int, default=6)
 143
 144 parser.add_argument("--grid_nb_shapes", type=int, default=6)
 145
 146 ##############################
 147 # picoclvr options
 148
 149 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 150
 151 parser.add_argument("--picoclvr_height", type=int, default=12)
 152
 153 parser.add_argument("--picoclvr_width", type=int, default=16)
 154
 155 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 156
 157 ##############################
 158 # Maze options
 159
 160 parser.add_argument("--maze_height", type=int, default=13)
 161
 162 parser.add_argument("--maze_width", type=int, default=21)
 163
 164 parser.add_argument("--maze_nb_walls", type=int, default=15)
 165
 166 ##############################
 167 # Snake options
 168
 169 parser.add_argument("--snake_height", type=int, default=9)
 170
 171 parser.add_argument("--snake_width", type=int, default=12)
 172
 173 parser.add_argument("--snake_nb_colors", type=int, default=5)
 174
 175 parser.add_argument("--snake_length", type=int, default=200)
 176
 177 ##############################
 178 # Stack options
 179
 180 parser.add_argument("--stack_nb_steps", type=int, default=100)
 181
 182 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 183
 184 parser.add_argument("--stack_nb_digits", type=int, default=3)
 185
 186 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 187
 188 ##############################
 189 # Expr options
 190
 191 parser.add_argument("--expr_nb_variables", type=int, default=5)
 192
 193 parser.add_argument("--expr_sequence_length", type=int, default=40)
 194
 195 parser.add_argument("--expr_operand_max", type=int, default=9)
 196
 197 parser.add_argument("--expr_result_max", type=int, default=99)
 198
 199 parser.add_argument("--expr_input_file", type=str, default=None)
 200
 201 ##############################
 202 # Memory
 203
 204 parser.add_argument("--memory_len_total", type=int, default=32)
 205
 206 ##############################
 207 # Mixing
 208
 209 parser.add_argument("--mixing_hard", action="store_true", default=False)
 210
 211 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 212
 213 ######################################################################
 214
 215 # args = parser.parse_args()
 216
 217 args, sup_args = parser.parse_known_args()
 218
 219 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 220
 221 if args.result_dir is None:
 222     args.result_dir = f"results_{args.task}_{args.model}"
 223
 224 ######################################################################
 225
 226 if not args.force_cpu and torch.cuda.is_available():
 227     device = torch.device("cuda")
 228     torch.backends.cuda.matmul.allow_tf32 = True
 229 else:
 230     device = torch.device("cpu")
 231
 232 ######################################################################
 233
 234 default_task_args = {
 235     "addition": {
 236         "model": "352M",
 237         "batch_size": 25,
 238         "nb_train_samples": 250000,
 239         "nb_test_samples": 10000,
 240     },
 241     "byheart": {
 242         "model": "37M",
 243         "batch_size": 25,
 244         "nb_train_samples": 50000,
 245         "nb_test_samples": 10000,
 246     },
 247     "expr": {
 248         "model": "352M",
 249         "batch_size": 25,
 250         "nb_train_samples": 2500000,
 251         "nb_test_samples": 10000,
 252     },
 253     "grid": {
 254         "model": "37M",
 255         "batch_size": 25,
 256         "nb_train_samples": 250000,
 257         "nb_test_samples": 10000,
 258     },
 259     "qmlp": {
 260         "model": "37M",
 261         "batch_size": 10,
 262         "nb_train_samples": 100000,
 263         "nb_test_samples": 1000,
 264     },
 265     "guessop": {
 266         "model": "352M",
 267         "batch_size": 25,
 268         "nb_train_samples": 1000000,
 269         "nb_test_samples": 10000,
 270     },
 271     "learnop": {
 272         "model": "37M",
 273         "batch_size": 25,
 274         "nb_train_samples": 50000,
 275         "nb_test_samples": 10000,
 276     },
 277     "maze": {
 278         "model": "37M",
 279         "batch_size": 5,
 280         "nb_train_samples": 100000,
 281         "nb_test_samples": 10000,
 282     },
 283     "picoclvr": {
 284         "model": "37M",
 285         "batch_size": 25,
 286         "nb_train_samples": 250000,
 287         "nb_test_samples": 10000,
 288     },
 289     "rpl": {
 290         "model": "352M",
 291         "batch_size": 5,
 292         "nb_train_samples": 2500000,
 293         "nb_test_samples": 10000,
 294     },
 295     "snake": {
 296         "model": "37M",
 297         "batch_size": 25,
 298         "nb_train_samples": 250000,
 299         "nb_test_samples": 10000,
 300     },
 301     "stack": {
 302         "model": "37M",
 303         "batch_size": 25,
 304         "nb_train_samples": 100000,
 305         "nb_test_samples": 1000,
 306     },
 307     "twotargets": {
 308         "model": "37M",
 309         "batch_size": 25,
 310         "nb_train_samples": 50000,
 311         "nb_test_samples": 10000,
 312     },
 313     "memory": {
 314         "model": "37M",
 315         "batch_size": 25,
 316         "nb_train_samples": 25000,
 317         "nb_test_samples": 10000,
 318     },
 319     "mixing": {
 320         "model": "37M",
 321         "batch_size": 25,
 322         "nb_train_samples": 250000,
 323         "nb_test_samples": 10000,
 324     },
 325     "mnist": {
 326         "model": "37M",
 327         "batch_size": 10,
 328         "nb_train_samples": 60000,
 329         "nb_test_samples": 10000,
 330     },
 331 }
 332
 333 if args.task in default_task_args:
 334     for k, v in default_task_args[args.task].items():
 335         if getattr(args, k) is None:
 336             setattr(args, k, v)
 337
 338 ######################################################################
 339
 340 default_model_args = {
 341     "17K": {
 342         "attention": "mha",
 343         "dim_model": 32,
 344         "dim_keys": 32,
 345         "dim_hidden": 32,
 346         "nb_heads": 2,
 347         "nb_blocks": 2,
 348     },
 349     "17K-C": {
 350         "attention": "caterpillar",
 351         "dim_model": 32,
 352         "dim_keys": 32,
 353         "dim_hidden": 32,
 354         "nb_heads": 2,
 355         "nb_lines": 16,
 356         "caterpillar_height": 4,
 357         "nb_blocks": 2,
 358     },
 359     "4M": {
 360         "attention": "mha",
 361         "dim_model": 256,
 362         "dim_keys": 32,
 363         "dim_hidden": 1024,
 364         "nb_heads": 4,
 365         "nb_blocks": 6,
 366     },
 367     "4M-C": {
 368         "attention": "caterpillar",
 369         "dim_model": 256,
 370         "dim_keys": 32,
 371         "dim_hidden": 1024,
 372         "nb_heads": 4,
 373         "nb_lines": 32,
 374         "caterpillar_height": 4,
 375         "nb_blocks": 6,
 376     },
 377     "37M": {
 378         "attention": "mha",
 379         "dim_model": 512,
 380         "dim_keys": 64,
 381         "dim_hidden": 2048,
 382         "nb_heads": 8,
 383         "nb_blocks": 12,
 384     },
 385     "37M-C": {
 386         "attention": "caterpillar",
 387         "dim_model": 512,
 388         "dim_keys": 64,
 389         "dim_hidden": 2048,
 390         "nb_heads": 8,
 391         "nb_lines": 256,
 392         "caterpillar_height": 32,
 393         "nb_blocks": 12,
 394     },
 395     "122M": {
 396         "attention": "mha",
 397         "dim_model": 768,
 398         "dim_keys": 64,
 399         "dim_hidden": 2048,
 400         "nb_heads": 8,
 401         "nb_blocks": 24,
 402     },
 403     "122M-C": {
 404         "attention": "caterpillar",
 405         "dim_model": 768,
 406         "dim_keys": 64,
 407         "dim_hidden": 2048,
 408         "nb_heads": 8,
 409         "nb_lines": 128,
 410         "nb_blocks": 24,
 411     },
 412     "352M": {
 413         "attention": "mha",
 414         "dim_model": 1024,
 415         "dim_keys": 64,
 416         "dim_hidden": 2048,
 417         "nb_heads": 8,
 418         "nb_blocks": 48,
 419     },
 420     "352M-C": {
 421         "attention": "caterpillar",
 422         "dim_model": 1024,
 423         "dim_keys": 64,
 424         "dim_hidden": 2048,
 425         "nb_heads": 8,
 426         "nb_lines": 128,
 427         "nb_blocks": 48,
 428     },
 429 }
 430
 431 if args.model in default_model_args:
 432     for k, v in default_model_args[args.model].items():
 433         if getattr(args, k) is None:
 434             setattr(args, k, v)
 435 else:
 436     raise ValueError(f"Unknown model {args.model}")
 437
 438 ######################################################################
 439
 440 try:
 441     os.mkdir(args.result_dir)
 442 except FileExistsError:
 443     if not args.continue_training:
 444         print(f"result directory {args.result_dir} already exists")
 445         exit(1)
 446
 447 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
 448
 449 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 450
 451 if args.seed >= 0:
 452     # torch.backends.cudnn.deterministic = True
 453     # torch.backends.cudnn.benchmark = False
 454     # torch.use_deterministic_algorithms(True)
 455     torch.manual_seed(args.seed)
 456     if torch.cuda.is_available():
 457         torch.cuda.manual_seed_all(args.seed)
 458
 459 ######################################################################
 460
 461
 462 def log_string(s):
 463     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 464
 465     if log_file is not None:
 466         log_file.write(t + s + "\n")
 467         log_file.flush()
 468
 469     print(t + s)
 470     sys.stdout.flush()
 471
 472
 473 with os.popen("sha256sum *.py") as f:
 474     for l in f:
 475         log_string(f"sha256sum {l.strip()}")
 476
 477 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 478 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 479
 480 log_string(f"argv {' '.join(sys.argv)}")
 481
 482 for n in vars(args):
 483     log_string(f"args.{n} {getattr(args, n)}")
 484
 485 for k, v in sup_args.items():
 486     log_string(f'sup_args["{k}"] "{v}"')
 487
 488
 489 ######################################################################
 490
 491
 492 def get_lr(n_epoch, it):
 493     if args.legacy_lr_schedule:
 494         # my crude scheduling to compare to previous baseline, added
 495         # warmup though
 496
 497         if it < args.nb_warmup_iter:
 498             return args.legacy_large_lr * it / args.nb_warmup_iter
 499         elif n_epoch < args.legacy_nb_epoch_large_lr:
 500             return args.legacy_large_lr
 501         else:
 502             return args.legacy_small_lr
 503
 504     # from nanoGPT
 505
 506     # 1) linear warmup for warmup_iter steps
 507     if it < args.nb_warmup_iter:
 508         return args.learning_rate * it / args.nb_warmup_iter
 509     # 2) if it > nb_decay_iter, return min learning rate
 510     if it > args.nb_decay_iter:
 511         return args.min_learning_rate
 512     # 3) in between, use cosine decay down to min learning rate
 513     decay_ratio = (it - args.nb_warmup_iter) / (
 514         args.nb_decay_iter - args.nb_warmup_iter
 515     )
 516     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 517     return args.min_learning_rate + coeff * (
 518         args.learning_rate - args.min_learning_rate
 519     )
 520
 521
 522 ######################################################################
 523
 524
 525 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 526
 527
 528 def picoclvr_pruner_horizontal_green(p):
 529     return not ("green" in p and ("left" in p or "right" in p))
 530
 531
 532 picoclvr_pruner_train = (
 533     picoclvr_pruner_horizontal_green
 534     if args.picocvlr_prune_properties in {"train+eval"}
 535     else None
 536 )
 537
 538 picoclvr_pruner_eval = (
 539     (lambda p: not picoclvr_pruner_horizontal_green(p))
 540     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 541     else None
 542 )
 543
 544 ######################################################################
 545
 546 device_data = device
 547
 548 if args.task == "byheart":
 549     task = tasks.SandBox(
 550         problem=problems.ProblemByHeart(),
 551         nb_train_samples=args.nb_train_samples,
 552         nb_test_samples=args.nb_test_samples,
 553         batch_size=args.batch_size,
 554         logger=log_string,
 555         device=device_data,
 556     )
 557     args.max_percents_of_test_in_train = -1
 558
 559 elif args.task == "learnop":
 560     task = tasks.SandBox(
 561         problem=problems.ProblemLearnOperator(),
 562         nb_train_samples=args.nb_train_samples,
 563         nb_test_samples=args.nb_test_samples,
 564         batch_size=args.batch_size,
 565         logger=log_string,
 566         device=device_data,
 567     )
 568
 569
 570 elif args.task == "guessop":
 571     task = tasks.SandBox(
 572         problem=problems.ProblemGuessOperator(),
 573         nb_train_samples=args.nb_train_samples,
 574         nb_test_samples=args.nb_test_samples,
 575         batch_size=args.batch_size,
 576         logger=log_string,
 577         device=device_data,
 578     )
 579
 580
 581 elif args.task == "twotargets":
 582     task = tasks.SandBox(
 583         problem=problems.ProblemTwoTargets(),
 584         nb_train_samples=args.nb_train_samples,
 585         nb_test_samples=args.nb_test_samples,
 586         batch_size=args.batch_size,
 587         logger=log_string,
 588         device=device_data,
 589     )
 590
 591 elif args.task == "memory":
 592     task = tasks.SandBox(
 593         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 594         nb_train_samples=args.nb_train_samples,
 595         nb_test_samples=args.nb_test_samples,
 596         batch_size=args.batch_size,
 597         logger=log_string,
 598         device=device_data,
 599     )
 600
 601 elif args.task == "mixing":
 602     task = tasks.SandBox(
 603         problem=problems.ProblemMixing(
 604             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 605         ),
 606         nb_train_samples=args.nb_train_samples,
 607         nb_test_samples=args.nb_test_samples,
 608         batch_size=args.batch_size,
 609         logger=log_string,
 610         device=device_data,
 611     )
 612
 613 elif args.task == "addition":
 614     task = tasks.SandBox(
 615         problem=problems.ProblemAddition(),
 616         nb_train_samples=args.nb_train_samples,
 617         nb_test_samples=args.nb_test_samples,
 618         batch_size=args.batch_size,
 619         logger=log_string,
 620         device=device_data,
 621     )
 622
 623 elif args.task == "picoclvr":
 624     task = tasks.PicoCLVR(
 625         nb_train_samples=args.nb_train_samples,
 626         nb_test_samples=args.nb_test_samples,
 627         batch_size=args.batch_size,
 628         height=args.picoclvr_height,
 629         width=args.picoclvr_width,
 630         nb_colors=args.picoclvr_nb_colors,
 631         logger=log_string,
 632         device=device_data,
 633         pruner_train=picoclvr_pruner_train,
 634         pruner_eval=picoclvr_pruner_eval,
 635     )
 636
 637 elif args.task == "mnist":
 638     task = tasks.MNIST(
 639         nb_train_samples=args.nb_train_samples,
 640         nb_test_samples=args.nb_test_samples,
 641         batch_size=args.batch_size,
 642         device=device_data,
 643     )
 644
 645 elif args.task == "maze":
 646     task = tasks.Maze(
 647         nb_train_samples=args.nb_train_samples,
 648         nb_test_samples=args.nb_test_samples,
 649         batch_size=args.batch_size,
 650         height=args.maze_height,
 651         width=args.maze_width,
 652         nb_walls=args.maze_nb_walls,
 653         device=device_data,
 654     )
 655
 656 elif args.task == "snake":
 657     task = tasks.Snake(
 658         nb_train_samples=args.nb_train_samples,
 659         nb_test_samples=args.nb_test_samples,
 660         batch_size=args.batch_size,
 661         height=args.snake_height,
 662         width=args.snake_width,
 663         nb_colors=args.snake_nb_colors,
 664         length=args.snake_length,
 665         prompt_length=args.snake_length // 2,
 666         device=device_data,
 667     )
 668
 669 elif args.task == "stack":
 670     task = tasks.Stack(
 671         nb_train_samples=args.nb_train_samples,
 672         nb_test_samples=args.nb_test_samples,
 673         batch_size=args.batch_size,
 674         logger=log_string,
 675         nb_steps=args.stack_nb_steps,
 676         nb_stacks=args.stack_nb_stacks,
 677         nb_digits=args.stack_nb_digits,
 678         fraction_values_for_train=args.stack_fraction_values_for_train,
 679         device=device_data,
 680     )
 681
 682 elif args.task == "expr":
 683     task = tasks.Expr(
 684         nb_train_samples=args.nb_train_samples,
 685         nb_test_samples=args.nb_test_samples,
 686         nb_variables=args.expr_nb_variables,
 687         sequence_length=args.expr_sequence_length,
 688         operand_max=args.expr_operand_max,
 689         result_max=args.expr_result_max,
 690         batch_size=args.batch_size,
 691         device=device_data,
 692     )
 693
 694 elif args.task == "rpl":
 695     task = tasks.RPL(
 696         nb_train_samples=args.nb_train_samples,
 697         nb_test_samples=args.nb_test_samples,
 698         batch_size=args.batch_size,
 699         nb_starting_values=args.rpl_nb_starting_values,
 700         max_input=args.rpl_max_input,
 701         prog_len=args.rpl_prog_len,
 702         nb_runs=args.rpl_nb_runs,
 703         no_prog=args.rpl_no_prog,
 704         logger=log_string,
 705         device=device_data,
 706     )
 707
 708 elif args.task == "grid":
 709     task = tasks.Grid(
 710         nb_train_samples=args.nb_train_samples,
 711         nb_test_samples=args.nb_test_samples,
 712         batch_size=args.batch_size,
 713         size=args.grid_size,
 714         nb_shapes=args.grid_nb_shapes,
 715         nb_colors=args.grid_nb_colors,
 716         logger=log_string,
 717         device=device_data,
 718     )
 719
 720 elif args.task == "qmlp":
 721     task = tasks.QMLP(
 722         nb_train_samples=args.nb_train_samples,
 723         nb_test_samples=args.nb_test_samples,
 724         batch_size=args.batch_size,
 725         result_dir=args.result_dir,
 726         logger=log_string,
 727         device=device_data,
 728     )
 729
 730 else:
 731     raise ValueError(f"Unknown task {args.task}")
 732
 733 ######################################################################
 734
 735 log_string(f"device {device}")
 736
 737 vocabulary_size = task.vocabulary_size()
 738
 739 log_string(f"vocabulary_size {vocabulary_size}")
 740
 741 ##############################
 742
 743 model = mygpt.MyGPT(
 744     vocabulary_size=vocabulary_size,
 745     dim_model=args.dim_model,
 746     dim_keys=args.dim_keys,
 747     dim_hidden=args.dim_hidden,
 748     nb_heads=args.nb_heads,
 749     nb_lines=args.nb_lines,
 750     caterpillar_height=args.caterpillar_height,
 751     nb_blocks=args.nb_blocks,
 752     causal=True,
 753     dropout=args.dropout,
 754     attention_layer=args.attention,
 755     logger=log_string,
 756     args=args,
 757 )
 758
 759 model.to(device)
 760
 761 nb_parameters = sum(p.numel() for p in model.parameters())
 762 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 763
 764 ######################################################################
 765
 766 nb_epochs_finished = 0
 767
 768 if args.no_checkpoint:
 769     log_string(f"not trying to load checkpoint.")
 770
 771 else:
 772     try:
 773         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 774         checkpoint = torch.load(checkpoint_name)
 775         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 776         model.load_state_dict(checkpoint["model_state"])
 777         torch.set_rng_state(checkpoint["rng_state"])
 778         if torch.cuda.is_available():
 779             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 780
 781         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 782
 783     except FileNotFoundError:
 784         log_string("starting from scratch.")
 785
 786     except:
 787         log_string("error when loading the checkpoint.")
 788         exit(1)
 789
 790 ######################################################################
 791
 792 if args.task == "expr" and args.expr_input_file is not None:
 793     task.produce_results(
 794         n_epoch=nb_epochs_finished,
 795         model=model,
 796         result_dir=args.result_dir,
 797         logger=log_string,
 798         deterministic_synthesis=args.deterministic_synthesis,
 799         input_file=args.expr_input_file,
 800     )
 801
 802     exit(0)
 803
 804 ######################################################################
 805
 806 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 807
 808 # Compute the entropy of the training tokens
 809
 810 token_count = 0
 811 for input in task.batches(split="train"):
 812     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 813 token_probas = token_count / token_count.sum()
 814 entropy = -torch.xlogy(token_probas, token_probas).sum()
 815 train_set_perplexity = math.exp(entropy)
 816
 817 ######################################################################
 818 # A bit of paranoia never hurts
 819
 820 if args.max_percents_of_test_in_train >= 0:
 821
 822     def subsets_as_tuples(batches, cs):
 823         s = set()
 824         for batch in batches:
 825             for x in batch:
 826                 s.add(tuple([v.item() for v in x]))
 827                 if len(s) == cs:
 828                     yield s
 829                     s = set()
 830         yield s
 831
 832     nb_test, nb_in_train = 0, 0
 833     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 834         in_train = set()
 835         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 836             in_train.update(test_subset.intersection(train_subset))
 837         nb_in_train += len(in_train)
 838         nb_test += len(test_subset)
 839
 840     log_string(
 841         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 842     )
 843
 844     assert (
 845         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 846     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 847
 848 ##############################
 849
 850 if "calibrate" in sup_args:
 851     for input in task.batches(split="train", desc="calibrate"):
 852         input = input.to(device)
 853         output = model(mygpt.BracketedSequence(input)).x
 854
 855     for n, m in model.named_modules():
 856         for a in dir(m):
 857             x = getattr(m, a)
 858             if isinstance(x, mygpt.Calibrator):
 859                 print(f"####### ${n} | ${a} ########################")
 860                 mean, std = x.moments()
 861                 print("mean\n", mean, "\n")
 862                 print("std\n", std, "\n")
 863                 print(f"############################################\n\n")
 864
 865     exit(0)
 866
 867 ##############################
 868
 869 nb_samples_seen = 0
 870
 871 if nb_epochs_finished >= nb_epochs:
 872     task.produce_results(
 873         n_epoch=nb_epochs_finished,
 874         model=model,
 875         result_dir=args.result_dir,
 876         logger=log_string,
 877         deterministic_synthesis=args.deterministic_synthesis,
 878     )
 879
 880 time_pred_result = datetime.datetime.now()
 881
 882 it = 0
 883
 884 n_batch = 0
 885
 886 for n_epoch in range(nb_epochs_finished, nb_epochs):
 887     if args.optim == "sgd":
 888         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 889     elif args.optim == "adam":
 890         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 891     elif args.optim == "adamw":
 892         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 893     else:
 894         raise ValueError(f"Unknown optimizer {args.optim}.")
 895
 896     model.train()
 897
 898     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 899
 900     for input in task.batches(split="train"):
 901         model.reset_inner_loss()
 902         input = input.to(device)
 903
 904         output = model(mygpt.BracketedSequence(input)).x
 905         loss = F.cross_entropy(output.transpose(1, 2), input)
 906         inner_loss = model.get_inner_loss()
 907
 908         acc_train_loss += loss.item() * input.size(0)
 909         acc_train_inner_loss += inner_loss.item() * input.size(0)
 910
 911         nb_train_samples += input.size(0)
 912         nb_samples_seen += input.size(0)
 913
 914         total_loss = loss + (
 915             args.rho_inner_loss * inner_loss if args.rho_inner_loss > 0 else 0.0
 916         )
 917
 918         it += 1
 919         lr = get_lr(n_epoch, it)
 920         for param_group in optimizer.param_groups:
 921             param_group["lr"] = lr
 922
 923         # log_string(f"learning_rate {lr}")
 924
 925         optimizer.zero_grad()
 926         total_loss.backward()
 927         optimizer.step()
 928
 929         grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
 930
 931         loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
 932
 933         n_batch += 1
 934
 935     with torch.autograd.no_grad():
 936         model.eval()
 937
 938         nb_test_samples, acc_test_loss = 0, 0.0
 939
 940         for input in task.batches(split="test"):
 941             input = input.to(device)
 942
 943             output = model(mygpt.BracketedSequence(input)).x
 944             loss = F.cross_entropy(output.transpose(1, 2), input)
 945             acc_test_loss += loss.item() * input.size(0)
 946             nb_test_samples += input.size(0)
 947
 948         log_string(
 949             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}"
 950         )
 951
 952         task.produce_results(
 953             n_epoch=n_epoch,
 954             model=model,
 955             result_dir=args.result_dir,
 956             logger=log_string,
 957             deterministic_synthesis=args.deterministic_synthesis,
 958         )
 959
 960         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 961         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 962
 963         log_string(
 964             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 965         )
 966
 967         time_current_result = datetime.datetime.now()
 968         log_string(
 969             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 970         )
 971         time_pred_result = time_current_result
 972
 973     checkpoint = {
 974         "nb_epochs_finished": n_epoch + 1,
 975         "model_state": model.state_dict(),
 976         "rng_state": torch.get_rng_state(),
 977     }
 978
 979     if torch.cuda.is_available():
 980         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 981
 982     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 983     torch.save(checkpoint, checkpoint_name)
 984     log_string(f"saved checkpoint {checkpoint_name}")
 985
 986 ######################################################################