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 if torch.cuda.is_available():
  20     device = torch.device("cuda")
  21     torch.backends.cuda.matmul.allow_tf32 = True
  22 else:
  23     device = torch.device("cpu")
  24
  25 ######################################################################
  26
  27
  28 def str2bool(x):
  29     x = x.lower()
  30     if x in {"1", "true", "yes"}:
  31         return True
  32     elif x in {"0", "false", "no"}:
  33         return False
  34     else:
  35         raise ValueError
  36
  37
  38 parser = argparse.ArgumentParser(
  39     description="An implementation of GPT with cache.",
  40     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  41 )
  42
  43 parser.add_argument(
  44     "--task",
  45     type=str,
  46     default="twotargets",
  47     help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
  48 )
  49
  50 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  51
  52 parser.add_argument("--result_dir", type=str, default=None)
  53
  54 parser.add_argument("--seed", type=int, default=0)
  55
  56 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  57
  58 ########################################
  59
  60 parser.add_argument("--nb_epochs", type=int, default=50)
  61
  62 parser.add_argument("--batch_size", type=int, default=None)
  63
  64 parser.add_argument("--nb_train_samples", type=int, default=None)
  65
  66 parser.add_argument("--nb_test_samples", type=int, default=None)
  67
  68 parser.add_argument("--optim", type=str, default="adam")
  69
  70 ########################################
  71
  72 parser.add_argument("--nb_warmup_iter", type=int, default=100)
  73
  74 parser.add_argument("--nb_decay_iter", type=int, default=5000)
  75
  76 parser.add_argument("--learning_rate", type=float, default=6e-4)
  77
  78 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
  79
  80 # legacy
  81
  82 parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
  83
  84 parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
  85
  86 parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
  87
  88 parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
  89
  90 ########################################
  91
  92 parser.add_argument("--model", type=str, default=None)
  93
  94 parser.add_argument("--attention", type=str, default=None)
  95
  96 parser.add_argument("--dim_model", type=int, default=None)
  97
  98 parser.add_argument("--dim_keys", type=int, default=None)
  99
 100 parser.add_argument("--dim_hidden", type=int, default=None)
 101
 102 parser.add_argument("--nb_heads", type=int, default=None)
 103
 104 parser.add_argument("--nb_lines", type=int, default=None)
 105
 106 parser.add_argument("--caterpillar_height", type=int, default=None)
 107
 108 parser.add_argument("--rho", type=float, default=0.0)
 109
 110 parser.add_argument("--dim_rec_v", type=int, default=None)
 111
 112 parser.add_argument("--nb_blocks", type=int, default=None)
 113
 114 parser.add_argument("--dropout", type=float, default=0.1)
 115
 116 ########################################
 117
 118 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 119
 120 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 121
 122 parser.add_argument("--overwrite_results", action="store_true", default=False)
 123
 124 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 125
 126 ##############################
 127 # rpl options
 128
 129 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 130
 131 parser.add_argument("--rpl_max_input", type=int, default=9)
 132
 133 parser.add_argument("--rpl_prog_len", type=int, default=8)
 134
 135 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 136
 137 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 138
 139 ##############################
 140 # grid options
 141
 142 parser.add_argument("--grid_size", type=int, default=6)
 143
 144 ##############################
 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 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 216
 217 if args.result_dir is None:
 218     args.result_dir = f"results_{args.task}_{args.model}"
 219
 220 ######################################################################
 221
 222 default_task_args = {
 223     "addition": {
 224         "model": "352M",
 225         "batch_size": 25,
 226         "nb_train_samples": 250000,
 227         "nb_test_samples": 10000,
 228     },
 229     "byheart": {
 230         "model": "37M",
 231         "batch_size": 25,
 232         "nb_train_samples": 50000,
 233         "nb_test_samples": 10000,
 234     },
 235     "expr": {
 236         "model": "352M",
 237         "batch_size": 25,
 238         "nb_train_samples": 2500000,
 239         "nb_test_samples": 10000,
 240     },
 241     "grid": {
 242         "model": "37M",
 243         "batch_size": 25,
 244         "nb_train_samples": 250000,
 245         "nb_test_samples": 10000,
 246     },
 247     "qmlp": {
 248         "model": "37M",
 249         "batch_size": 10,
 250         "nb_train_samples": 100000,
 251         "nb_test_samples": 1000,
 252     },
 253     "guessop": {
 254         "model": "352M",
 255         "batch_size": 25,
 256         "nb_train_samples": 1000000,
 257         "nb_test_samples": 10000,
 258     },
 259     "learnop": {
 260         "model": "37M",
 261         "batch_size": 25,
 262         "nb_train_samples": 50000,
 263         "nb_test_samples": 10000,
 264     },
 265     "maze": {
 266         "model": "37M",
 267         "batch_size": 5,
 268         "nb_train_samples": 100000,
 269         "nb_test_samples": 10000,
 270     },
 271     "picoclvr": {
 272         "model": "37M",
 273         "batch_size": 25,
 274         "nb_train_samples": 250000,
 275         "nb_test_samples": 10000,
 276     },
 277     "rpl": {
 278         "model": "352M",
 279         "batch_size": 5,
 280         "nb_train_samples": 2500000,
 281         "nb_test_samples": 10000,
 282     },
 283     "snake": {
 284         "model": "37M",
 285         "batch_size": 25,
 286         "nb_train_samples": 250000,
 287         "nb_test_samples": 10000,
 288     },
 289     "stack": {
 290         "model": "37M",
 291         "batch_size": 25,
 292         "nb_train_samples": 100000,
 293         "nb_test_samples": 1000,
 294     },
 295     "twotargets": {
 296         "model": "37M",
 297         "batch_size": 25,
 298         "nb_train_samples": 50000,
 299         "nb_test_samples": 10000,
 300     },
 301     "memory": {
 302         "model": "37M",
 303         "batch_size": 25,
 304         "nb_train_samples": 25000,
 305         "nb_test_samples": 10000,
 306     },
 307     "mixing": {
 308         "model": "37M",
 309         "batch_size": 25,
 310         "nb_train_samples": 250000,
 311         "nb_test_samples": 10000,
 312     },
 313     "mnist": {
 314         "model": "37M",
 315         "batch_size": 10,
 316         "nb_train_samples": 60000,
 317         "nb_test_samples": 10000,
 318     },
 319 }
 320
 321 if args.task in default_task_args:
 322     for k, v in default_task_args[args.task].items():
 323         if getattr(args, k) is None:
 324             setattr(args, k, v)
 325
 326 ######################################################################
 327
 328 default_model_args = {
 329     "17K": {
 330         "attention": "mha",
 331         "dim_model": 32,
 332         "dim_keys": 32,
 333         "dim_hidden": 32,
 334         "nb_heads": 2,
 335         "dim_rec_v": 16,
 336         "nb_blocks": 2,
 337     },
 338     "17K-C": {
 339         "attention": "caterpillar",
 340         "dim_model": 32,
 341         "dim_keys": 32,
 342         "dim_hidden": 32,
 343         "nb_heads": 2,
 344         "nb_lines": 16,
 345         "caterpillar_height": 4,
 346         "dim_rec_v": 16,
 347         "nb_blocks": 2,
 348     },
 349     "4M": {
 350         "attention": "mha",
 351         "dim_model": 256,
 352         "dim_keys": 32,
 353         "dim_hidden": 1024,
 354         "nb_heads": 4,
 355         "dim_rec_v": 64,
 356         "nb_blocks": 6,
 357     },
 358     "4M-C": {
 359         "attention": "caterpillar",
 360         "dim_model": 256,
 361         "dim_keys": 32,
 362         "dim_hidden": 1024,
 363         "nb_heads": 4,
 364         "nb_lines": 32,
 365         "caterpillar_height": 4,
 366         "dim_rec_v": 64,  # dim_model / nb_heads
 367         "nb_blocks": 6,
 368     },
 369     "37M": {
 370         "attention": "mha",
 371         "dim_model": 512,
 372         "dim_keys": 64,
 373         "dim_hidden": 2048,
 374         "nb_heads": 8,
 375         "dim_rec_v": 64,
 376         "nb_blocks": 12,
 377     },
 378     "37M-C": {
 379         "attention": "caterpillar",
 380         "dim_model": 512,
 381         "dim_keys": 64,
 382         "dim_hidden": 2048,
 383         "nb_heads": 8,
 384         "nb_lines": 256,
 385         "caterpillar_height": 32,
 386         "dim_rec_v": 64,
 387         "nb_blocks": 12,
 388     },
 389     "122M": {
 390         "attention": "mha",
 391         "dim_model": 768,
 392         "dim_keys": 64,
 393         "dim_hidden": 2048,
 394         "nb_heads": 8,
 395         "dim_rec_v": 96,
 396         "nb_blocks": 24,
 397     },
 398     "122M-C": {
 399         "attention": "caterpillar",
 400         "dim_model": 768,
 401         "dim_keys": 64,
 402         "dim_hidden": 2048,
 403         "nb_heads": 8,
 404         "nb_lines": 128,
 405         "dim_rec_v": 96,
 406         "nb_blocks": 24,
 407     },
 408     "352M": {
 409         "attention": "mha",
 410         "dim_model": 1024,
 411         "dim_keys": 64,
 412         "dim_hidden": 2048,
 413         "nb_heads": 8,
 414         "dim_rec_v": 128,
 415         "nb_blocks": 48,
 416     },
 417     "352M-C": {
 418         "attention": "caterpillar",
 419         "dim_model": 1024,
 420         "dim_keys": 64,
 421         "dim_hidden": 2048,
 422         "nb_heads": 8,
 423         "nb_lines": 128,
 424         "dim_rec_v": 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.overwrite_results:
 442         print(f"result directory {args.result_dir} already exists")
 443         exit(1)
 444
 445 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 446
 447 if args.seed >= 0:
 448     # torch.backends.cudnn.deterministic = True
 449     # torch.backends.cudnn.benchmark = False
 450     # torch.use_deterministic_algorithms(True)
 451     torch.manual_seed(args.seed)
 452     if torch.cuda.is_available():
 453         torch.cuda.manual_seed_all(args.seed)
 454
 455 ######################################################################
 456
 457
 458 def log_string(s):
 459     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 460
 461     if log_file is not None:
 462         log_file.write(t + s + "\n")
 463         log_file.flush()
 464
 465     print(t + s)
 466     sys.stdout.flush()
 467
 468
 469 with os.popen("sha256sum *.py") as f:
 470     for l in f:
 471         log_string(f"sha256sum {l.strip()}")
 472
 473 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 474 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 475
 476 log_string(f"argv {' '.join(sys.argv)}")
 477
 478 for n in vars(args):
 479     log_string(f"args.{n} {getattr(args, n)}")
 480
 481
 482 ######################################################################
 483
 484
 485 def get_lr(n_epoch, it):
 486     if args.legacy_lr_schedule:
 487         # my crude scheduling to compare to previous baseline, added
 488         # warmup though
 489
 490         if it < args.nb_warmup_iter:
 491             return args.legacy_large_lr * it / args.nb_warmup_iter
 492         elif n_epoch < args.legacy_nb_epoch_large_lr:
 493             return args.legacy_large_lr
 494         else:
 495             return args.legacy_small_lr
 496
 497     # from nanoGPT
 498
 499     # 1) linear warmup for warmup_iter steps
 500     if it < args.nb_warmup_iter:
 501         return args.learning_rate * it / args.nb_warmup_iter
 502     # 2) if it > nb_decay_iter, return min learning rate
 503     if it > args.nb_decay_iter:
 504         return args.min_learning_rate
 505     # 3) in between, use cosine decay down to min learning rate
 506     decay_ratio = (it - args.nb_warmup_iter) / (
 507         args.nb_decay_iter - args.nb_warmup_iter
 508     )
 509     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 510     return args.min_learning_rate + coeff * (
 511         args.learning_rate - args.min_learning_rate
 512     )
 513
 514
 515 ######################################################################
 516
 517
 518 def picoclvr_pruner_horizontal_green(p):
 519     return not ("green" in p and ("left" in p or "right" in p))
 520
 521
 522 picoclvr_pruner_train = (
 523     picoclvr_pruner_horizontal_green
 524     if args.picocvlr_prune_properties in {"train+eval"}
 525     else None
 526 )
 527
 528 picoclvr_pruner_eval = (
 529     (lambda p: not picoclvr_pruner_horizontal_green(p))
 530     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 531     else None
 532 )
 533
 534 ######################################################################
 535
 536 device_data = device
 537
 538 if args.task == "byheart":
 539     task = tasks.SandBox(
 540         problem=problems.ProblemByHeart(),
 541         nb_train_samples=args.nb_train_samples,
 542         nb_test_samples=args.nb_test_samples,
 543         batch_size=args.batch_size,
 544         logger=log_string,
 545         device=device_data,
 546     )
 547     args.max_percents_of_test_in_train = -1
 548
 549 elif args.task == "learnop":
 550     task = tasks.SandBox(
 551         problem=problems.ProblemLearnOperator(),
 552         nb_train_samples=args.nb_train_samples,
 553         nb_test_samples=args.nb_test_samples,
 554         batch_size=args.batch_size,
 555         logger=log_string,
 556         device=device_data,
 557     )
 558
 559
 560 elif args.task == "guessop":
 561     task = tasks.SandBox(
 562         problem=problems.ProblemGuessOperator(),
 563         nb_train_samples=args.nb_train_samples,
 564         nb_test_samples=args.nb_test_samples,
 565         batch_size=args.batch_size,
 566         logger=log_string,
 567         device=device_data,
 568     )
 569
 570
 571 elif args.task == "twotargets":
 572     task = tasks.SandBox(
 573         problem=problems.ProblemTwoTargets(),
 574         nb_train_samples=args.nb_train_samples,
 575         nb_test_samples=args.nb_test_samples,
 576         batch_size=args.batch_size,
 577         logger=log_string,
 578         device=device_data,
 579     )
 580
 581 elif args.task == "memory":
 582     task = tasks.SandBox(
 583         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 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 == "mixing":
 592     task = tasks.SandBox(
 593         problem=problems.ProblemMixing(
 594             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 595         ),
 596         nb_train_samples=args.nb_train_samples,
 597         nb_test_samples=args.nb_test_samples,
 598         batch_size=args.batch_size,
 599         logger=log_string,
 600         device=device_data,
 601     )
 602
 603 elif args.task == "addition":
 604     task = tasks.SandBox(
 605         problem=problems.ProblemAddition(),
 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 == "picoclvr":
 614     task = tasks.PicoCLVR(
 615         nb_train_samples=args.nb_train_samples,
 616         nb_test_samples=args.nb_test_samples,
 617         batch_size=args.batch_size,
 618         height=args.picoclvr_height,
 619         width=args.picoclvr_width,
 620         nb_colors=args.picoclvr_nb_colors,
 621         logger=log_string,
 622         device=device_data,
 623         pruner_train=picoclvr_pruner_train,
 624         pruner_eval=picoclvr_pruner_eval,
 625     )
 626
 627 elif args.task == "mnist":
 628     task = tasks.MNIST(
 629         nb_train_samples=args.nb_train_samples,
 630         nb_test_samples=args.nb_test_samples,
 631         batch_size=args.batch_size,
 632         device=device_data,
 633     )
 634
 635 elif args.task == "maze":
 636     task = tasks.Maze(
 637         nb_train_samples=args.nb_train_samples,
 638         nb_test_samples=args.nb_test_samples,
 639         batch_size=args.batch_size,
 640         height=args.maze_height,
 641         width=args.maze_width,
 642         nb_walls=args.maze_nb_walls,
 643         device=device_data,
 644     )
 645
 646 elif args.task == "snake":
 647     task = tasks.Snake(
 648         nb_train_samples=args.nb_train_samples,
 649         nb_test_samples=args.nb_test_samples,
 650         batch_size=args.batch_size,
 651         height=args.snake_height,
 652         width=args.snake_width,
 653         nb_colors=args.snake_nb_colors,
 654         length=args.snake_length,
 655         prompt_length=args.snake_length // 2,
 656         device=device_data,
 657     )
 658
 659 elif args.task == "stack":
 660     task = tasks.Stack(
 661         nb_train_samples=args.nb_train_samples,
 662         nb_test_samples=args.nb_test_samples,
 663         batch_size=args.batch_size,
 664         logger=log_string,
 665         nb_steps=args.stack_nb_steps,
 666         nb_stacks=args.stack_nb_stacks,
 667         nb_digits=args.stack_nb_digits,
 668         fraction_values_for_train=args.stack_fraction_values_for_train,
 669         device=device_data,
 670     )
 671
 672 elif args.task == "expr":
 673     task = tasks.Expr(
 674         nb_train_samples=args.nb_train_samples,
 675         nb_test_samples=args.nb_test_samples,
 676         nb_variables=args.expr_nb_variables,
 677         sequence_length=args.expr_sequence_length,
 678         operand_max=args.expr_operand_max,
 679         result_max=args.expr_result_max,
 680         batch_size=args.batch_size,
 681         device=device_data,
 682     )
 683
 684 elif args.task == "rpl":
 685     task = tasks.RPL(
 686         nb_train_samples=args.nb_train_samples,
 687         nb_test_samples=args.nb_test_samples,
 688         batch_size=args.batch_size,
 689         nb_starting_values=args.rpl_nb_starting_values,
 690         max_input=args.rpl_max_input,
 691         prog_len=args.rpl_prog_len,
 692         nb_runs=args.rpl_nb_runs,
 693         no_prog=args.rpl_no_prog,
 694         logger=log_string,
 695         device=device_data,
 696     )
 697
 698 elif args.task == "grid":
 699     task = tasks.Grid(
 700         nb_train_samples=args.nb_train_samples,
 701         nb_test_samples=args.nb_test_samples,
 702         batch_size=args.batch_size,
 703         size=args.grid_size,
 704         logger=log_string,
 705         device=device_data,
 706     )
 707
 708 elif args.task == "qmlp":
 709     task = tasks.QMLP(
 710         nb_train_samples=args.nb_train_samples,
 711         nb_test_samples=args.nb_test_samples,
 712         batch_size=args.batch_size,
 713         result_dir=args.result_dir,
 714         logger=log_string,
 715         device=device_data,
 716     )
 717
 718 else:
 719     raise ValueError(f"Unknown task {args.task}")
 720
 721 ######################################################################
 722
 723 log_string(f"device {device}")
 724
 725 vocabulary_size = task.vocabulary_size()
 726
 727 log_string(f"vocabulary_size {vocabulary_size}")
 728
 729 ##############################
 730
 731 model = mygpt.MyGPT(
 732     vocabulary_size=vocabulary_size,
 733     dim_model=args.dim_model,
 734     dim_keys=args.dim_keys,
 735     dim_hidden=args.dim_hidden,
 736     nb_heads=args.nb_heads,
 737     nb_lines=args.nb_lines,
 738     caterpillar_height=args.caterpillar_height,
 739     dim_rec_v=args.dim_rec_v,
 740     nb_blocks=args.nb_blocks,
 741     causal=True,
 742     dropout=args.dropout,
 743     attention_layer=args.attention,
 744 )
 745
 746 model.to(device)
 747
 748 nb_parameters = sum(p.numel() for p in model.parameters())
 749 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 750
 751 ######################################################################
 752
 753 nb_epochs_finished = 0
 754
 755 if args.no_checkpoint:
 756     log_string(f"not trying to load checkpoint.")
 757
 758 else:
 759     try:
 760         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 761         checkpoint = torch.load(checkpoint_name)
 762         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 763         model.load_state_dict(checkpoint["model_state"])
 764         torch.set_rng_state(checkpoint["rng_state"])
 765         if torch.cuda.is_available():
 766             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 767
 768         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 769
 770     except FileNotFoundError:
 771         log_string("starting from scratch.")
 772
 773     except:
 774         log_string("error when loading the checkpoint.")
 775         exit(1)
 776
 777 ######################################################################
 778
 779 if args.task == "expr" and args.expr_input_file is not None:
 780     task.produce_results(
 781         n_epoch=nb_epochs_finished,
 782         model=model,
 783         result_dir=args.result_dir,
 784         logger=log_string,
 785         deterministic_synthesis=args.deterministic_synthesis,
 786         input_file=args.expr_input_file,
 787     )
 788
 789     exit(0)
 790
 791 ######################################################################
 792
 793 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 794
 795 # Compute the entropy of the training tokens
 796
 797 token_count = 0
 798 for input in task.batches(split="train"):
 799     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 800 token_probas = token_count / token_count.sum()
 801 entropy = -torch.xlogy(token_probas, token_probas).sum()
 802 train_set_perplexity = math.exp(entropy)
 803
 804 ######################################################################
 805 # A bit of paranoia never hurts
 806
 807 if args.max_percents_of_test_in_train >= 0:
 808
 809     def subsets_as_tuples(batches, cs):
 810         s = set()
 811         for batch in batches:
 812             for x in batch:
 813                 s.add(tuple([v.item() for v in x]))
 814                 if len(s) == cs:
 815                     yield s
 816                     s = set()
 817         yield s
 818
 819     nb_test, nb_in_train = 0, 0
 820     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 821         in_train = set()
 822         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 823             in_train.update(test_subset.intersection(train_subset))
 824         nb_in_train += len(in_train)
 825         nb_test += len(test_subset)
 826
 827     log_string(
 828         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 829     )
 830
 831     assert (
 832         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 833     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 834
 835 ##############################
 836
 837 nb_samples_seen = 0
 838
 839 if nb_epochs_finished >= nb_epochs:
 840     task.produce_results(
 841         n_epoch=nb_epochs_finished,
 842         model=model,
 843         result_dir=args.result_dir,
 844         logger=log_string,
 845         deterministic_synthesis=args.deterministic_synthesis,
 846     )
 847
 848 time_pred_result = None
 849
 850 it = 0
 851
 852 for n_epoch in range(nb_epochs_finished, nb_epochs):
 853     if args.optim == "sgd":
 854         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 855     elif args.optim == "adam":
 856         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 857     elif args.optim == "adamw":
 858         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 859     else:
 860         raise ValueError(f"Unknown optimizer {args.optim}.")
 861
 862     model.train()
 863
 864     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 865
 866     for input in task.batches(split="train"):
 867         model.reset_inner_loss()
 868         input = input.to(device)
 869
 870         output = model(mygpt.BracketedSequence(input)).x
 871         loss = F.cross_entropy(output.transpose(1, 2), input)
 872         inner_loss = model.get_inner_loss()
 873
 874         acc_train_loss += loss.item() * input.size(0)
 875         acc_train_inner_loss += inner_loss.item() * input.size(0)
 876
 877         nb_train_samples += input.size(0)
 878         nb_samples_seen += input.size(0)
 879
 880         total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
 881
 882         it += 1
 883         lr = get_lr(n_epoch, it)
 884         for param_group in optimizer.param_groups:
 885             param_group["lr"] = lr
 886
 887         # log_string(f"learning_rate {lr}")
 888
 889         optimizer.zero_grad()
 890         total_loss.backward()
 891         optimizer.step()
 892
 893     with torch.autograd.no_grad():
 894         model.eval()
 895
 896         nb_test_samples, acc_test_loss = 0, 0.0
 897
 898         for input in task.batches(split="test"):
 899             input = input.to(device)
 900
 901             output = model(mygpt.BracketedSequence(input)).x
 902             loss = F.cross_entropy(output.transpose(1, 2), input)
 903             acc_test_loss += loss.item() * input.size(0)
 904             nb_test_samples += input.size(0)
 905
 906         log_string(
 907             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}"
 908         )
 909
 910         task.produce_results(
 911             n_epoch=n_epoch,
 912             model=model,
 913             result_dir=args.result_dir,
 914             logger=log_string,
 915             deterministic_synthesis=args.deterministic_synthesis,
 916         )
 917
 918         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 919         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 920
 921         log_string(
 922             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 923         )
 924
 925         time_current_result = datetime.datetime.now()
 926         if time_pred_result is not None:
 927             log_string(
 928                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 929             )
 930         time_pred_result = time_current_result
 931
 932     checkpoint = {
 933         "nb_epochs_finished": n_epoch + 1,
 934         "model_state": model.state_dict(),
 935         "rng_state": torch.get_rng_state(),
 936     }
 937
 938     if torch.cuda.is_available():
 939         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 940
 941     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 942     torch.save(checkpoint, checkpoint_name)
 943     log_string(f"saved checkpoint {checkpoint_name}")
 944
 945 ######################################################################