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 parser = argparse.ArgumentParser(
  28     description="An implementation of GPT with cache.",
  29     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  30 )
  31
  32 parser.add_argument(
  33     "--task",
  34     type=str,
  35     default="twotargets",
  36     help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
  37 )
  38
  39 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  40
  41 parser.add_argument("--result_dir", type=str, default=None)
  42
  43 parser.add_argument("--seed", type=int, default=0)
  44
  45 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  46
  47 ########################################
  48
  49 parser.add_argument("--nb_epochs", type=int, default=50)
  50
  51 parser.add_argument("--batch_size", type=int, default=None)
  52
  53 parser.add_argument("--nb_train_samples", type=int, default=None)
  54
  55 parser.add_argument("--nb_test_samples", type=int, default=None)
  56
  57 parser.add_argument("--optim", type=str, default="adam")
  58
  59 ########################################
  60
  61 parser.add_argument("--nb_warmup_iter", type=int, default=100)
  62
  63 parser.add_argument("--nb_decay_iter", type=int, default=5000)
  64
  65 parser.add_argument("--learning_rate", type=float, default=6e-4)
  66
  67 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
  68
  69 # legacy
  70
  71 parser.add_argument("--legacy_lr_schedule", action="store_true", default=False)
  72
  73 parser.add_argument("--legacy_learning_rate", type=float, default=1e-4)
  74
  75 parser.add_argument("--legacy_min_learning_rate", type=float, default=2e-5)
  76
  77 parser.add_argument("--nb_large_lr_epochs", type=float, default=10)
  78
  79 ########################################
  80
  81 parser.add_argument("--model", type=str, default=None)
  82
  83 parser.add_argument("--attention", type=str, default=None)
  84
  85 parser.add_argument("--dim_model", type=int, default=None)
  86
  87 parser.add_argument("--dim_keys", type=int, default=None)
  88
  89 parser.add_argument("--dim_hidden", type=int, default=None)
  90
  91 parser.add_argument("--nb_heads", type=int, default=None)
  92
  93 parser.add_argument("--nb_lines", type=int, default=None)
  94
  95 parser.add_argument("--caterpillar_height", type=int, default=None)
  96
  97 parser.add_argument("--rho", type=float, default=0.0)
  98
  99 parser.add_argument("--dim_rec_v", type=int, default=None)
 100
 101 parser.add_argument("--nb_blocks", type=int, default=None)
 102
 103 parser.add_argument("--dropout", type=float, default=0.1)
 104
 105 ########################################
 106
 107 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 108
 109 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 110
 111 parser.add_argument("--overwrite_results", action="store_true", default=False)
 112
 113 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 114
 115 ##############################
 116 # rpl options
 117
 118 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 119
 120 parser.add_argument("--rpl_max_input", type=int, default=9)
 121
 122 parser.add_argument("--rpl_prog_len", type=int, default=8)
 123
 124 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 125
 126 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 127
 128 ##############################
 129 # grid options
 130
 131 parser.add_argument("--grid_size", type=int, default=6)
 132
 133 ##############################
 134 # picoclvr options
 135
 136 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 137
 138 parser.add_argument("--picoclvr_height", type=int, default=12)
 139
 140 parser.add_argument("--picoclvr_width", type=int, default=16)
 141
 142 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 143
 144 ##############################
 145 # Maze options
 146
 147 parser.add_argument("--maze_height", type=int, default=13)
 148
 149 parser.add_argument("--maze_width", type=int, default=21)
 150
 151 parser.add_argument("--maze_nb_walls", type=int, default=15)
 152
 153 ##############################
 154 # Snake options
 155
 156 parser.add_argument("--snake_height", type=int, default=9)
 157
 158 parser.add_argument("--snake_width", type=int, default=12)
 159
 160 parser.add_argument("--snake_nb_colors", type=int, default=5)
 161
 162 parser.add_argument("--snake_length", type=int, default=200)
 163
 164 ##############################
 165 # Stack options
 166
 167 parser.add_argument("--stack_nb_steps", type=int, default=100)
 168
 169 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 170
 171 parser.add_argument("--stack_nb_digits", type=int, default=3)
 172
 173 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 174
 175 ##############################
 176 # Expr options
 177
 178 parser.add_argument("--expr_nb_variables", type=int, default=5)
 179
 180 parser.add_argument("--expr_sequence_length", type=int, default=40)
 181
 182 parser.add_argument("--expr_operand_max", type=int, default=9)
 183
 184 parser.add_argument("--expr_result_max", type=int, default=99)
 185
 186 parser.add_argument("--expr_input_file", type=str, default=None)
 187
 188 ##############################
 189 # Memory
 190
 191 parser.add_argument("--memory_len_total", type=int, default=32)
 192
 193 ##############################
 194 # Mixing
 195
 196 parser.add_argument("--mixing_hard", action="store_true", default=False)
 197
 198 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 199
 200 ######################################################################
 201
 202 args = parser.parse_args()
 203
 204 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 205
 206 if args.result_dir is None:
 207     args.result_dir = f"results_{args.task}_{args.model}"
 208
 209 ######################################################################
 210
 211 default_task_args = {
 212     "addition": {
 213         "model": "352M",
 214         "batch_size": 25,
 215         "nb_train_samples": 250000,
 216         "nb_test_samples": 10000,
 217     },
 218     "byheart": {
 219         "model": "37M",
 220         "batch_size": 25,
 221         "nb_train_samples": 50000,
 222         "nb_test_samples": 10000,
 223     },
 224     "expr": {
 225         "model": "352M",
 226         "batch_size": 25,
 227         "nb_train_samples": 2500000,
 228         "nb_test_samples": 10000,
 229     },
 230     "grid": {
 231         "model": "37M",
 232         "batch_size": 25,
 233         "nb_train_samples": 250000,
 234         "nb_test_samples": 10000,
 235     },
 236     "qmlp": {
 237         "model": "37M",
 238         "batch_size": 10,
 239         "nb_train_samples": 100000,
 240         "nb_test_samples": 1000,
 241     },
 242     "guessop": {
 243         "model": "352M",
 244         "batch_size": 25,
 245         "nb_train_samples": 1000000,
 246         "nb_test_samples": 10000,
 247     },
 248     "learnop": {
 249         "model": "37M",
 250         "batch_size": 25,
 251         "nb_train_samples": 50000,
 252         "nb_test_samples": 10000,
 253     },
 254     "maze": {
 255         "model": "37M",
 256         "batch_size": 5,
 257         "nb_train_samples": 100000,
 258         "nb_test_samples": 10000,
 259     },
 260     "picoclvr": {
 261         "model": "37M",
 262         "batch_size": 25,
 263         "nb_train_samples": 250000,
 264         "nb_test_samples": 10000,
 265     },
 266     "rpl": {
 267         "model": "352M",
 268         "batch_size": 5,
 269         "nb_train_samples": 2500000,
 270         "nb_test_samples": 10000,
 271     },
 272     "snake": {
 273         "model": "37M",
 274         "batch_size": 25,
 275         "nb_train_samples": 250000,
 276         "nb_test_samples": 10000,
 277     },
 278     "stack": {
 279         "model": "37M",
 280         "batch_size": 25,
 281         "nb_train_samples": 100000,
 282         "nb_test_samples": 1000,
 283     },
 284     "twotargets": {
 285         "model": "37M",
 286         "batch_size": 25,
 287         "nb_train_samples": 50000,
 288         "nb_test_samples": 10000,
 289     },
 290     "memory": {
 291         "model": "37M",
 292         "batch_size": 25,
 293         "nb_train_samples": 25000,
 294         "nb_test_samples": 10000,
 295     },
 296     "mixing": {
 297         "model": "37M",
 298         "batch_size": 25,
 299         "nb_train_samples": 250000,
 300         "nb_test_samples": 10000,
 301     },
 302     "mnist": {
 303         "model": "37M",
 304         "batch_size": 10,
 305         "nb_train_samples": 60000,
 306         "nb_test_samples": 10000,
 307     },
 308 }
 309
 310 if args.task in default_task_args:
 311     for k, v in default_task_args[args.task].items():
 312         if getattr(args, k) is None:
 313             setattr(args, k, v)
 314
 315 ######################################################################
 316
 317 default_model_args = {
 318     "17K": {
 319         "attention": "mha",
 320         "dim_model": 32,
 321         "dim_keys": 32,
 322         "dim_hidden": 32,
 323         "nb_heads": 2,
 324         "dim_rec_v": 16,
 325         "nb_blocks": 2,
 326     },
 327     "17K-C": {
 328         "attention": "caterpillar",
 329         "dim_model": 32,
 330         "dim_keys": 32,
 331         "dim_hidden": 32,
 332         "nb_heads": 2,
 333         "nb_lines": 16,
 334         "caterpillar_height": 4,
 335         "dim_rec_v": 16,
 336         "nb_blocks": 2,
 337     },
 338     "4M": {
 339         "attention": "mha",
 340         "dim_model": 256,
 341         "dim_keys": 32,
 342         "dim_hidden": 1024,
 343         "nb_heads": 4,
 344         "dim_rec_v": 64,
 345         "nb_blocks": 6,
 346     },
 347     "4M-C": {
 348         "attention": "caterpillar",
 349         "dim_model": 256,
 350         "dim_keys": 32,
 351         "dim_hidden": 1024,
 352         "nb_heads": 4,
 353         "nb_lines": 32,
 354         "caterpillar_height": 4,
 355         "dim_rec_v": 64,  # dim_model / nb_heads
 356         "nb_blocks": 6,
 357     },
 358     "37M": {
 359         "attention": "mha",
 360         "dim_model": 512,
 361         "dim_keys": 64,
 362         "dim_hidden": 2048,
 363         "nb_heads": 8,
 364         "dim_rec_v": 64,
 365         "nb_blocks": 12,
 366     },
 367     "37M-C": {
 368         "attention": "caterpillar",
 369         "dim_model": 512,
 370         "dim_keys": 64,
 371         "dim_hidden": 2048,
 372         "nb_heads": 8,
 373         "nb_lines": 256,
 374         "caterpillar_height": 32,
 375         "dim_rec_v": 64,
 376         "nb_blocks": 12,
 377     },
 378     "122M": {
 379         "attention": "mha",
 380         "dim_model": 768,
 381         "dim_keys": 64,
 382         "dim_hidden": 2048,
 383         "nb_heads": 8,
 384         "dim_rec_v": 96,
 385         "nb_blocks": 24,
 386     },
 387     "122M-C": {
 388         "attention": "caterpillar",
 389         "dim_model": 768,
 390         "dim_keys": 64,
 391         "dim_hidden": 2048,
 392         "nb_heads": 8,
 393         "nb_lines": 128,
 394         "dim_rec_v": 96,
 395         "nb_blocks": 24,
 396     },
 397     "352M": {
 398         "attention": "mha",
 399         "dim_model": 1024,
 400         "dim_keys": 64,
 401         "dim_hidden": 2048,
 402         "nb_heads": 8,
 403         "dim_rec_v": 128,
 404         "nb_blocks": 48,
 405     },
 406     "352M-C": {
 407         "attention": "caterpillar",
 408         "dim_model": 1024,
 409         "dim_keys": 64,
 410         "dim_hidden": 2048,
 411         "nb_heads": 8,
 412         "nb_lines": 128,
 413         "dim_rec_v": 128,
 414         "nb_blocks": 48,
 415     },
 416 }
 417
 418 if args.model in default_model_args:
 419     for k, v in default_model_args[args.model].items():
 420         if getattr(args, k) is None:
 421             setattr(args, k, v)
 422 else:
 423     raise ValueError(f"Unknown model {args.model}")
 424
 425 ######################################################################
 426
 427 try:
 428     os.mkdir(args.result_dir)
 429 except FileExistsError:
 430     if not args.overwrite_results:
 431         print(f"result directory {args.result_dir} already exists")
 432         exit(1)
 433
 434 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 435
 436 if args.seed >= 0:
 437     # torch.backends.cudnn.deterministic = True
 438     # torch.backends.cudnn.benchmark = False
 439     # torch.use_deterministic_algorithms(True)
 440     torch.manual_seed(args.seed)
 441     if torch.cuda.is_available():
 442         torch.cuda.manual_seed_all(args.seed)
 443
 444 ######################################################################
 445
 446
 447 def log_string(s):
 448     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 449
 450     if log_file is not None:
 451         log_file.write(t + s + "\n")
 452         log_file.flush()
 453
 454     print(t + s)
 455     sys.stdout.flush()
 456
 457
 458 with os.popen("sha256sum *.py") as f:
 459     for l in f:
 460         log_string(f"sha256sum {l.strip()}")
 461
 462 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 463 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 464
 465 log_string(f"argv {' '.join(sys.argv)}")
 466
 467 for n in vars(args):
 468     log_string(f"args.{n} {getattr(args, n)}")
 469
 470
 471 ######################################################################
 472
 473
 474 def get_lr(n_epoch, it):
 475     if args.legacy_lr_schedule:
 476         # my crude scheduling to compare to previous baseline, added
 477         # warmup though
 478
 479         if it < args.nb_warmup_iter:
 480             return args.legacy_learning_rate * it / args.nb_warmup_iter
 481         elif it < args.nb_large_lr_epochs:
 482             return args.legacy_learning_rate
 483         else:
 484             return args.legacy_min_learning_rate
 485
 486     # from nanoGPT
 487
 488     # 1) linear warmup for warmup_iter steps
 489     if it < args.nb_warmup_iter:
 490         return args.learning_rate * it / args.nb_warmup_iter
 491     # 2) if it > nb_decay_iter, return min learning rate
 492     if it > args.nb_decay_iter:
 493         return args.min_learning_rate
 494     # 3) in between, use cosine decay down to min learning rate
 495     decay_ratio = (it - args.nb_warmup_iter) / (
 496         args.nb_decay_iter - args.nb_warmup_iter
 497     )
 498     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 499     return args.min_learning_rate + coeff * (
 500         args.learning_rate - args.min_learning_rate
 501     )
 502
 503
 504 ######################################################################
 505
 506
 507 def picoclvr_pruner_horizontal_green(p):
 508     return not ("green" in p and ("left" in p or "right" in p))
 509
 510
 511 picoclvr_pruner_train = (
 512     picoclvr_pruner_horizontal_green
 513     if args.picocvlr_prune_properties in {"train+eval"}
 514     else None
 515 )
 516
 517 picoclvr_pruner_eval = (
 518     (lambda p: not picoclvr_pruner_horizontal_green(p))
 519     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 520     else None
 521 )
 522
 523 ######################################################################
 524
 525 device_data = device
 526
 527 if args.task == "byheart":
 528     task = tasks.SandBox(
 529         problem=problems.ProblemByHeart(),
 530         nb_train_samples=args.nb_train_samples,
 531         nb_test_samples=args.nb_test_samples,
 532         batch_size=args.batch_size,
 533         logger=log_string,
 534         device=device_data,
 535     )
 536     args.max_percents_of_test_in_train = -1
 537
 538 elif args.task == "learnop":
 539     task = tasks.SandBox(
 540         problem=problems.ProblemLearnOperator(),
 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
 548
 549 elif args.task == "guessop":
 550     task = tasks.SandBox(
 551         problem=problems.ProblemGuessOperator(),
 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 == "twotargets":
 561     task = tasks.SandBox(
 562         problem=problems.ProblemTwoTargets(),
 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 elif args.task == "memory":
 571     task = tasks.SandBox(
 572         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 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 elif args.task == "mixing":
 581     task = tasks.SandBox(
 582         problem=problems.ProblemMixing(
 583             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 584         ),
 585         nb_train_samples=args.nb_train_samples,
 586         nb_test_samples=args.nb_test_samples,
 587         batch_size=args.batch_size,
 588         logger=log_string,
 589         device=device_data,
 590     )
 591
 592 elif args.task == "addition":
 593     task = tasks.SandBox(
 594         problem=problems.ProblemAddition(),
 595         nb_train_samples=args.nb_train_samples,
 596         nb_test_samples=args.nb_test_samples,
 597         batch_size=args.batch_size,
 598         logger=log_string,
 599         device=device_data,
 600     )
 601
 602 elif args.task == "picoclvr":
 603     task = tasks.PicoCLVR(
 604         nb_train_samples=args.nb_train_samples,
 605         nb_test_samples=args.nb_test_samples,
 606         batch_size=args.batch_size,
 607         height=args.picoclvr_height,
 608         width=args.picoclvr_width,
 609         nb_colors=args.picoclvr_nb_colors,
 610         logger=log_string,
 611         device=device_data,
 612         pruner_train=picoclvr_pruner_train,
 613         pruner_eval=picoclvr_pruner_eval,
 614     )
 615
 616 elif args.task == "mnist":
 617     task = tasks.MNIST(
 618         nb_train_samples=args.nb_train_samples,
 619         nb_test_samples=args.nb_test_samples,
 620         batch_size=args.batch_size,
 621         device=device_data,
 622     )
 623
 624 elif args.task == "maze":
 625     task = tasks.Maze(
 626         nb_train_samples=args.nb_train_samples,
 627         nb_test_samples=args.nb_test_samples,
 628         batch_size=args.batch_size,
 629         height=args.maze_height,
 630         width=args.maze_width,
 631         nb_walls=args.maze_nb_walls,
 632         device=device_data,
 633     )
 634
 635 elif args.task == "snake":
 636     task = tasks.Snake(
 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.snake_height,
 641         width=args.snake_width,
 642         nb_colors=args.snake_nb_colors,
 643         length=args.snake_length,
 644         prompt_length=args.snake_length // 2,
 645         device=device_data,
 646     )
 647
 648 elif args.task == "stack":
 649     task = tasks.Stack(
 650         nb_train_samples=args.nb_train_samples,
 651         nb_test_samples=args.nb_test_samples,
 652         batch_size=args.batch_size,
 653         logger=log_string,
 654         nb_steps=args.stack_nb_steps,
 655         nb_stacks=args.stack_nb_stacks,
 656         nb_digits=args.stack_nb_digits,
 657         fraction_values_for_train=args.stack_fraction_values_for_train,
 658         device=device_data,
 659     )
 660
 661 elif args.task == "expr":
 662     task = tasks.Expr(
 663         nb_train_samples=args.nb_train_samples,
 664         nb_test_samples=args.nb_test_samples,
 665         nb_variables=args.expr_nb_variables,
 666         sequence_length=args.expr_sequence_length,
 667         operand_max=args.expr_operand_max,
 668         result_max=args.expr_result_max,
 669         batch_size=args.batch_size,
 670         device=device_data,
 671     )
 672
 673 elif args.task == "rpl":
 674     task = tasks.RPL(
 675         nb_train_samples=args.nb_train_samples,
 676         nb_test_samples=args.nb_test_samples,
 677         batch_size=args.batch_size,
 678         nb_starting_values=args.rpl_nb_starting_values,
 679         max_input=args.rpl_max_input,
 680         prog_len=args.rpl_prog_len,
 681         nb_runs=args.rpl_nb_runs,
 682         no_prog=args.rpl_no_prog,
 683         logger=log_string,
 684         device=device_data,
 685     )
 686
 687 elif args.task == "grid":
 688     task = tasks.Grid(
 689         nb_train_samples=args.nb_train_samples,
 690         nb_test_samples=args.nb_test_samples,
 691         batch_size=args.batch_size,
 692         size=args.grid_size,
 693         logger=log_string,
 694         device=device_data,
 695     )
 696
 697 elif args.task == "qmlp":
 698     task = tasks.QMLP(
 699         nb_train_samples=args.nb_train_samples,
 700         nb_test_samples=args.nb_test_samples,
 701         batch_size=args.batch_size,
 702         result_dir=args.result_dir,
 703         logger=log_string,
 704         device=device_data,
 705     )
 706
 707 else:
 708     raise ValueError(f"Unknown task {args.task}")
 709
 710 ######################################################################
 711
 712 log_string(f"device {device}")
 713
 714 vocabulary_size = task.vocabulary_size()
 715
 716 log_string(f"vocabulary_size {vocabulary_size}")
 717
 718 ##############################
 719
 720 model = mygpt.MyGPT(
 721     vocabulary_size=vocabulary_size,
 722     dim_model=args.dim_model,
 723     dim_keys=args.dim_keys,
 724     dim_hidden=args.dim_hidden,
 725     nb_heads=args.nb_heads,
 726     nb_lines=args.nb_lines,
 727     caterpillar_height=args.caterpillar_height,
 728     dim_rec_v=args.dim_rec_v,
 729     nb_blocks=args.nb_blocks,
 730     causal=True,
 731     dropout=args.dropout,
 732     attention_layer=args.attention,
 733 )
 734
 735 model.to(device)
 736
 737 nb_parameters = sum(p.numel() for p in model.parameters())
 738 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 739
 740 ######################################################################
 741
 742 nb_epochs_finished = 0
 743
 744 if args.no_checkpoint:
 745     log_string(f"not trying to load checkpoint.")
 746
 747 else:
 748     try:
 749         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 750         checkpoint = torch.load(checkpoint_name)
 751         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 752         model.load_state_dict(checkpoint["model_state"])
 753         torch.set_rng_state(checkpoint["rng_state"])
 754         if torch.cuda.is_available():
 755             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 756
 757         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 758
 759     except FileNotFoundError:
 760         log_string("starting from scratch.")
 761
 762     except:
 763         log_string("error when loading the checkpoint.")
 764         exit(1)
 765
 766 ######################################################################
 767
 768 if args.task == "expr" and args.expr_input_file is not None:
 769     task.produce_results(
 770         n_epoch=nb_epochs_finished,
 771         model=model,
 772         result_dir=args.result_dir,
 773         logger=log_string,
 774         deterministic_synthesis=args.deterministic_synthesis,
 775         input_file=args.expr_input_file,
 776     )
 777
 778     exit(0)
 779
 780 ######################################################################
 781
 782 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 783
 784 # Compute the entropy of the training tokens
 785
 786 token_count = 0
 787 for input in task.batches(split="train"):
 788     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 789 token_probas = token_count / token_count.sum()
 790 entropy = -torch.xlogy(token_probas, token_probas).sum()
 791 train_set_perplexity = math.exp(entropy)
 792
 793 ######################################################################
 794 # A bit of paranoia never hurts
 795
 796 if args.max_percents_of_test_in_train >= 0:
 797
 798     def subsets_as_tuples(batches, cs):
 799         s = set()
 800         for batch in batches:
 801             for x in batch:
 802                 s.add(tuple([v.item() for v in x]))
 803                 if len(s) == cs:
 804                     yield s
 805                     s = set()
 806         yield s
 807
 808     nb_test, nb_in_train = 0, 0
 809     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 810         in_train = set()
 811         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 812             in_train.update(test_subset.intersection(train_subset))
 813         nb_in_train += len(in_train)
 814         nb_test += len(test_subset)
 815
 816     log_string(
 817         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 818     )
 819
 820     assert (
 821         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 822     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 823
 824 ##############################
 825
 826 nb_samples_seen = 0
 827
 828 if nb_epochs_finished >= nb_epochs:
 829     task.produce_results(
 830         n_epoch=nb_epochs_finished,
 831         model=model,
 832         result_dir=args.result_dir,
 833         logger=log_string,
 834         deterministic_synthesis=args.deterministic_synthesis,
 835     )
 836
 837 time_pred_result = None
 838
 839 it = 0
 840
 841 for n_epoch in range(nb_epochs_finished, nb_epochs):
 842     if args.optim == "sgd":
 843         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 844     elif args.optim == "adam":
 845         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 846     elif args.optim == "adamw":
 847         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 848     else:
 849         raise ValueError(f"Unknown optimizer {args.optim}.")
 850
 851     model.train()
 852
 853     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 854
 855     for input in task.batches(split="train"):
 856         model.reset_inner_loss()
 857         input = input.to(device)
 858
 859         output = model(mygpt.BracketedSequence(input)).x
 860         loss = F.cross_entropy(output.transpose(1, 2), input)
 861         inner_loss = model.get_inner_loss()
 862
 863         acc_train_loss += loss.item() * input.size(0)
 864         acc_train_inner_loss += inner_loss.item() * input.size(0)
 865
 866         nb_train_samples += input.size(0)
 867         nb_samples_seen += input.size(0)
 868
 869         total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
 870
 871         it += 1
 872         lr = get_lr(n_epoch, it)
 873         for param_group in optimizer.param_groups:
 874             param_group["lr"] = lr
 875
 876         # log_string(f"learning_rate {lr}")
 877
 878         optimizer.zero_grad()
 879         total_loss.backward()
 880         optimizer.step()
 881
 882     with torch.autograd.no_grad():
 883         model.eval()
 884
 885         nb_test_samples, acc_test_loss = 0, 0.0
 886
 887         for input in task.batches(split="test"):
 888             input = input.to(device)
 889
 890             output = model(mygpt.BracketedSequence(input)).x
 891             loss = F.cross_entropy(output.transpose(1, 2), input)
 892             acc_test_loss += loss.item() * input.size(0)
 893             nb_test_samples += input.size(0)
 894
 895         log_string(
 896             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}"
 897         )
 898
 899         task.produce_results(
 900             n_epoch=n_epoch,
 901             model=model,
 902             result_dir=args.result_dir,
 903             logger=log_string,
 904             deterministic_synthesis=args.deterministic_synthesis,
 905         )
 906
 907         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 908         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 909
 910         log_string(
 911             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 912         )
 913
 914         time_current_result = datetime.datetime.now()
 915         if time_pred_result is not None:
 916             log_string(
 917                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 918             )
 919         time_pred_result = time_current_result
 920
 921     checkpoint = {
 922         "nb_epochs_finished": n_epoch + 1,
 923         "model_state": model.state_dict(),
 924         "rng_state": torch.get_rng_state(),
 925     }
 926
 927     if torch.cuda.is_available():
 928         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 929
 930     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 931     torch.save(checkpoint, checkpoint_name)
 932     log_string(f"saved checkpoint {checkpoint_name}")
 933
 934 ######################################################################