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("--rho", type=float, default=0.0)
 103
 104 parser.add_argument("--nb_blocks", type=int, default=None)
 105
 106 parser.add_argument("--dropout", type=float, default=0.1)
 107
 108 ########################################
 109
 110 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 111
 112 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 113
 114 parser.add_argument("--continue_training", action="store_true", default=False)
 115
 116 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 117
 118 ##############################
 119 # rpl options
 120
 121 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 122
 123 parser.add_argument("--rpl_max_input", type=int, default=9)
 124
 125 parser.add_argument("--rpl_prog_len", type=int, default=8)
 126
 127 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 128
 129 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 130
 131 ##############################
 132 # grid options
 133
 134 parser.add_argument("--grid_size", type=int, default=6)
 135
 136 ##############################
 137 # picoclvr options
 138
 139 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 140
 141 parser.add_argument("--picoclvr_height", type=int, default=12)
 142
 143 parser.add_argument("--picoclvr_width", type=int, default=16)
 144
 145 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 146
 147 ##############################
 148 # Maze options
 149
 150 parser.add_argument("--maze_height", type=int, default=13)
 151
 152 parser.add_argument("--maze_width", type=int, default=21)
 153
 154 parser.add_argument("--maze_nb_walls", type=int, default=15)
 155
 156 ##############################
 157 # Snake options
 158
 159 parser.add_argument("--snake_height", type=int, default=9)
 160
 161 parser.add_argument("--snake_width", type=int, default=12)
 162
 163 parser.add_argument("--snake_nb_colors", type=int, default=5)
 164
 165 parser.add_argument("--snake_length", type=int, default=200)
 166
 167 ##############################
 168 # Stack options
 169
 170 parser.add_argument("--stack_nb_steps", type=int, default=100)
 171
 172 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 173
 174 parser.add_argument("--stack_nb_digits", type=int, default=3)
 175
 176 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 177
 178 ##############################
 179 # Expr options
 180
 181 parser.add_argument("--expr_nb_variables", type=int, default=5)
 182
 183 parser.add_argument("--expr_sequence_length", type=int, default=40)
 184
 185 parser.add_argument("--expr_operand_max", type=int, default=9)
 186
 187 parser.add_argument("--expr_result_max", type=int, default=99)
 188
 189 parser.add_argument("--expr_input_file", type=str, default=None)
 190
 191 ##############################
 192 # Memory
 193
 194 parser.add_argument("--memory_len_total", type=int, default=32)
 195
 196 ##############################
 197 # Mixing
 198
 199 parser.add_argument("--mixing_hard", action="store_true", default=False)
 200
 201 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 202
 203 ######################################################################
 204
 205 # args = parser.parse_args()
 206
 207 args, sup_args = parser.parse_known_args()
 208
 209 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 210
 211 if args.result_dir is None:
 212     args.result_dir = f"results_{args.task}_{args.model}"
 213
 214 ######################################################################
 215
 216 if not args.force_cpu and torch.cuda.is_available():
 217     device = torch.device("cuda")
 218     torch.backends.cuda.matmul.allow_tf32 = True
 219 else:
 220     device = torch.device("cpu")
 221
 222 ######################################################################
 223
 224 default_task_args = {
 225     "addition": {
 226         "model": "352M",
 227         "batch_size": 25,
 228         "nb_train_samples": 250000,
 229         "nb_test_samples": 10000,
 230     },
 231     "byheart": {
 232         "model": "37M",
 233         "batch_size": 25,
 234         "nb_train_samples": 50000,
 235         "nb_test_samples": 10000,
 236     },
 237     "expr": {
 238         "model": "352M",
 239         "batch_size": 25,
 240         "nb_train_samples": 2500000,
 241         "nb_test_samples": 10000,
 242     },
 243     "grid": {
 244         "model": "37M",
 245         "batch_size": 25,
 246         "nb_train_samples": 250000,
 247         "nb_test_samples": 10000,
 248     },
 249     "qmlp": {
 250         "model": "37M",
 251         "batch_size": 10,
 252         "nb_train_samples": 100000,
 253         "nb_test_samples": 1000,
 254     },
 255     "guessop": {
 256         "model": "352M",
 257         "batch_size": 25,
 258         "nb_train_samples": 1000000,
 259         "nb_test_samples": 10000,
 260     },
 261     "learnop": {
 262         "model": "37M",
 263         "batch_size": 25,
 264         "nb_train_samples": 50000,
 265         "nb_test_samples": 10000,
 266     },
 267     "maze": {
 268         "model": "37M",
 269         "batch_size": 5,
 270         "nb_train_samples": 100000,
 271         "nb_test_samples": 10000,
 272     },
 273     "picoclvr": {
 274         "model": "37M",
 275         "batch_size": 25,
 276         "nb_train_samples": 250000,
 277         "nb_test_samples": 10000,
 278     },
 279     "rpl": {
 280         "model": "352M",
 281         "batch_size": 5,
 282         "nb_train_samples": 2500000,
 283         "nb_test_samples": 10000,
 284     },
 285     "snake": {
 286         "model": "37M",
 287         "batch_size": 25,
 288         "nb_train_samples": 250000,
 289         "nb_test_samples": 10000,
 290     },
 291     "stack": {
 292         "model": "37M",
 293         "batch_size": 25,
 294         "nb_train_samples": 100000,
 295         "nb_test_samples": 1000,
 296     },
 297     "twotargets": {
 298         "model": "37M",
 299         "batch_size": 25,
 300         "nb_train_samples": 50000,
 301         "nb_test_samples": 10000,
 302     },
 303     "memory": {
 304         "model": "37M",
 305         "batch_size": 25,
 306         "nb_train_samples": 25000,
 307         "nb_test_samples": 10000,
 308     },
 309     "mixing": {
 310         "model": "37M",
 311         "batch_size": 25,
 312         "nb_train_samples": 250000,
 313         "nb_test_samples": 10000,
 314     },
 315     "mnist": {
 316         "model": "37M",
 317         "batch_size": 10,
 318         "nb_train_samples": 60000,
 319         "nb_test_samples": 10000,
 320     },
 321 }
 322
 323 if args.task in default_task_args:
 324     for k, v in default_task_args[args.task].items():
 325         if getattr(args, k) is None:
 326             setattr(args, k, v)
 327
 328 ######################################################################
 329
 330 default_model_args = {
 331     "17K": {
 332         "attention": "mha",
 333         "dim_model": 32,
 334         "dim_keys": 32,
 335         "dim_hidden": 32,
 336         "nb_heads": 2,
 337         "nb_blocks": 2,
 338     },
 339     "17K-C": {
 340         "attention": "caterpillar",
 341         "dim_model": 32,
 342         "dim_keys": 32,
 343         "dim_hidden": 32,
 344         "nb_heads": 2,
 345         "nb_lines": 16,
 346         "caterpillar_height": 4,
 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         "nb_blocks": 6,
 356     },
 357     "4M-C": {
 358         "attention": "caterpillar",
 359         "dim_model": 256,
 360         "dim_keys": 32,
 361         "dim_hidden": 1024,
 362         "nb_heads": 4,
 363         "nb_lines": 32,
 364         "caterpillar_height": 4,
 365         "nb_blocks": 6,
 366     },
 367     "37M": {
 368         "attention": "mha",
 369         "dim_model": 512,
 370         "dim_keys": 64,
 371         "dim_hidden": 2048,
 372         "nb_heads": 8,
 373         "nb_blocks": 12,
 374     },
 375     "37M-C": {
 376         "attention": "caterpillar",
 377         "dim_model": 512,
 378         "dim_keys": 64,
 379         "dim_hidden": 2048,
 380         "nb_heads": 8,
 381         "nb_lines": 256,
 382         "caterpillar_height": 32,
 383         "nb_blocks": 12,
 384     },
 385     "122M": {
 386         "attention": "mha",
 387         "dim_model": 768,
 388         "dim_keys": 64,
 389         "dim_hidden": 2048,
 390         "nb_heads": 8,
 391         "nb_blocks": 24,
 392     },
 393     "122M-C": {
 394         "attention": "caterpillar",
 395         "dim_model": 768,
 396         "dim_keys": 64,
 397         "dim_hidden": 2048,
 398         "nb_heads": 8,
 399         "nb_lines": 128,
 400         "nb_blocks": 24,
 401     },
 402     "352M": {
 403         "attention": "mha",
 404         "dim_model": 1024,
 405         "dim_keys": 64,
 406         "dim_hidden": 2048,
 407         "nb_heads": 8,
 408         "nb_blocks": 48,
 409     },
 410     "352M-C": {
 411         "attention": "caterpillar",
 412         "dim_model": 1024,
 413         "dim_keys": 64,
 414         "dim_hidden": 2048,
 415         "nb_heads": 8,
 416         "nb_lines": 128,
 417         "nb_blocks": 48,
 418     },
 419 }
 420
 421 if args.model in default_model_args:
 422     for k, v in default_model_args[args.model].items():
 423         if getattr(args, k) is None:
 424             setattr(args, k, v)
 425 else:
 426     raise ValueError(f"Unknown model {args.model}")
 427
 428 ######################################################################
 429
 430 try:
 431     os.mkdir(args.result_dir)
 432 except FileExistsError:
 433     if not args.continue_training:
 434         print(f"result directory {args.result_dir} already exists")
 435         exit(1)
 436
 437 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
 438
 439 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 440
 441 if args.seed >= 0:
 442     # torch.backends.cudnn.deterministic = True
 443     # torch.backends.cudnn.benchmark = False
 444     # torch.use_deterministic_algorithms(True)
 445     torch.manual_seed(args.seed)
 446     if torch.cuda.is_available():
 447         torch.cuda.manual_seed_all(args.seed)
 448
 449 ######################################################################
 450
 451
 452 def log_string(s):
 453     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 454
 455     if log_file is not None:
 456         log_file.write(t + s + "\n")
 457         log_file.flush()
 458
 459     print(t + s)
 460     sys.stdout.flush()
 461
 462
 463 with os.popen("sha256sum *.py") as f:
 464     for l in f:
 465         log_string(f"sha256sum {l.strip()}")
 466
 467 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 468 os.system(f"tar --ignore-failed-read zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 469
 470 log_string(f"argv {' '.join(sys.argv)}")
 471
 472 for n in vars(args):
 473     log_string(f"args.{n} {getattr(args, n)}")
 474
 475 for n in vars(sup_args):
 476     log_string(f"sup_args.{n} {getattr(sup_args, n)}")
 477
 478
 479 ######################################################################
 480
 481
 482 def get_lr(n_epoch, it):
 483     if args.legacy_lr_schedule:
 484         # my crude scheduling to compare to previous baseline, added
 485         # warmup though
 486
 487         if it < args.nb_warmup_iter:
 488             return args.legacy_large_lr * it / args.nb_warmup_iter
 489         elif n_epoch < args.legacy_nb_epoch_large_lr:
 490             return args.legacy_large_lr
 491         else:
 492             return args.legacy_small_lr
 493
 494     # from nanoGPT
 495
 496     # 1) linear warmup for warmup_iter steps
 497     if it < args.nb_warmup_iter:
 498         return args.learning_rate * it / args.nb_warmup_iter
 499     # 2) if it > nb_decay_iter, return min learning rate
 500     if it > args.nb_decay_iter:
 501         return args.min_learning_rate
 502     # 3) in between, use cosine decay down to min learning rate
 503     decay_ratio = (it - args.nb_warmup_iter) / (
 504         args.nb_decay_iter - args.nb_warmup_iter
 505     )
 506     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 507     return args.min_learning_rate + coeff * (
 508         args.learning_rate - args.min_learning_rate
 509     )
 510
 511
 512 ######################################################################
 513
 514
 515 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 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     nb_blocks=args.nb_blocks,
 740     causal=True,
 741     dropout=args.dropout,
 742     attention_layer=args.attention,
 743     logger=log_string,
 744     **sup_args,
 745 )
 746
 747 model.to(device)
 748
 749 nb_parameters = sum(p.numel() for p in model.parameters())
 750 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 751
 752 ######################################################################
 753
 754 nb_epochs_finished = 0
 755
 756 if args.no_checkpoint:
 757     log_string(f"not trying to load checkpoint.")
 758
 759 else:
 760     try:
 761         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 762         checkpoint = torch.load(checkpoint_name)
 763         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 764         model.load_state_dict(checkpoint["model_state"])
 765         torch.set_rng_state(checkpoint["rng_state"])
 766         if torch.cuda.is_available():
 767             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 768
 769         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 770
 771     except FileNotFoundError:
 772         log_string("starting from scratch.")
 773
 774     except:
 775         log_string("error when loading the checkpoint.")
 776         exit(1)
 777
 778 ######################################################################
 779
 780 if args.task == "expr" and args.expr_input_file is not None:
 781     task.produce_results(
 782         n_epoch=nb_epochs_finished,
 783         model=model,
 784         result_dir=args.result_dir,
 785         logger=log_string,
 786         deterministic_synthesis=args.deterministic_synthesis,
 787         input_file=args.expr_input_file,
 788     )
 789
 790     exit(0)
 791
 792 ######################################################################
 793
 794 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 795
 796 # Compute the entropy of the training tokens
 797
 798 token_count = 0
 799 for input in task.batches(split="train"):
 800     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 801 token_probas = token_count / token_count.sum()
 802 entropy = -torch.xlogy(token_probas, token_probas).sum()
 803 train_set_perplexity = math.exp(entropy)
 804
 805 ######################################################################
 806 # A bit of paranoia never hurts
 807
 808 if args.max_percents_of_test_in_train >= 0:
 809
 810     def subsets_as_tuples(batches, cs):
 811         s = set()
 812         for batch in batches:
 813             for x in batch:
 814                 s.add(tuple([v.item() for v in x]))
 815                 if len(s) == cs:
 816                     yield s
 817                     s = set()
 818         yield s
 819
 820     nb_test, nb_in_train = 0, 0
 821     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 822         in_train = set()
 823         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 824             in_train.update(test_subset.intersection(train_subset))
 825         nb_in_train += len(in_train)
 826         nb_test += len(test_subset)
 827
 828     log_string(
 829         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 830     )
 831
 832     assert (
 833         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 834     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 835
 836 ##############################
 837
 838 nb_samples_seen = 0
 839
 840 if nb_epochs_finished >= nb_epochs:
 841     task.produce_results(
 842         n_epoch=nb_epochs_finished,
 843         model=model,
 844         result_dir=args.result_dir,
 845         logger=log_string,
 846         deterministic_synthesis=args.deterministic_synthesis,
 847     )
 848
 849 time_pred_result = datetime.datetime.now()
 850
 851 it = 0
 852
 853 n_batch = 0
 854
 855 for n_epoch in range(nb_epochs_finished, nb_epochs):
 856     if args.optim == "sgd":
 857         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 858     elif args.optim == "adam":
 859         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 860     elif args.optim == "adamw":
 861         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 862     else:
 863         raise ValueError(f"Unknown optimizer {args.optim}.")
 864
 865     model.train()
 866
 867     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 868
 869     for input in task.batches(split="train"):
 870         model.reset_inner_loss()
 871         input = input.to(device)
 872
 873         output = model(mygpt.BracketedSequence(input)).x
 874         loss = F.cross_entropy(output.transpose(1, 2), input)
 875         inner_loss = model.get_inner_loss()
 876
 877         acc_train_loss += loss.item() * input.size(0)
 878         acc_train_inner_loss += inner_loss.item() * input.size(0)
 879
 880         nb_train_samples += input.size(0)
 881         nb_samples_seen += input.size(0)
 882
 883         total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
 884
 885         it += 1
 886         lr = get_lr(n_epoch, it)
 887         for param_group in optimizer.param_groups:
 888             param_group["lr"] = lr
 889
 890         # log_string(f"learning_rate {lr}")
 891
 892         optimizer.zero_grad()
 893         total_loss.backward()
 894         optimizer.step()
 895
 896         grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
 897
 898         loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
 899
 900         n_batch += 1
 901
 902     with torch.autograd.no_grad():
 903         model.eval()
 904
 905         nb_test_samples, acc_test_loss = 0, 0.0
 906
 907         for input in task.batches(split="test"):
 908             input = input.to(device)
 909
 910             output = model(mygpt.BracketedSequence(input)).x
 911             loss = F.cross_entropy(output.transpose(1, 2), input)
 912             acc_test_loss += loss.item() * input.size(0)
 913             nb_test_samples += input.size(0)
 914
 915         log_string(
 916             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}"
 917         )
 918
 919         task.produce_results(
 920             n_epoch=n_epoch,
 921             model=model,
 922             result_dir=args.result_dir,
 923             logger=log_string,
 924             deterministic_synthesis=args.deterministic_synthesis,
 925         )
 926
 927         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 928         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 929
 930         log_string(
 931             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 932         )
 933
 934         time_current_result = datetime.datetime.now()
 935         log_string(
 936             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 937         )
 938         time_pred_result = time_current_result
 939
 940     checkpoint = {
 941         "nb_epochs_finished": n_epoch + 1,
 942         "model_state": model.state_dict(),
 943         "rng_state": torch.get_rng_state(),
 944     }
 945
 946     if torch.cuda.is_available():
 947         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 948
 949     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 950     torch.save(checkpoint, checkpoint_name)
 951     log_string(f"saved checkpoint {checkpoint_name}")
 952
 953 ######################################################################