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="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
  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("--physical_batch_size", type=int, default=None)
  54
  55 parser.add_argument("--nb_train_samples", type=int, default=None)
  56
  57 parser.add_argument("--nb_test_samples", type=int, default=None)
  58
  59 parser.add_argument("--optim", type=str, default="adam")
  60
  61 parser.add_argument("--learning_rate", type=float, default=1e-4)
  62
  63 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  64
  65 ########################################
  66
  67 parser.add_argument("--model", type=str, default=None)
  68
  69 parser.add_argument("--dim_model", type=int, default=None)
  70
  71 parser.add_argument("--dim_keys", type=int, default=None)
  72
  73 parser.add_argument("--dim_hidden", type=int, default=None)
  74
  75 parser.add_argument("--nb_heads", type=int, default=None)
  76
  77 parser.add_argument("--nb_blocks", type=int, default=None)
  78
  79 parser.add_argument("--dropout", type=float, default=0.1)
  80
  81 ########################################
  82
  83 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  84
  85 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  86
  87 parser.add_argument("--resume", action="store_true", default=False)
  88
  89 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  90
  91 ##############################
  92 # filetask
  93
  94 parser.add_argument("--filetask_train_file", type=str, default=None)
  95
  96 parser.add_argument("--filetask_test_file", type=str, default=None)
  97
  98 ##############################
  99 # rpl options
 100
 101 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 102
 103 parser.add_argument("--rpl_max_input", type=int, default=9)
 104
 105 parser.add_argument("--rpl_prog_len", type=int, default=8)
 106
 107 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 108
 109 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 110
 111 ##############################
 112 # grid options
 113
 114 parser.add_argument("--grid_size", type=int, default=6)
 115
 116 parser.add_argument("--grid_fraction_play", type=float, default=0)
 117
 118 ##############################
 119 # picoclvr options
 120
 121 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 122
 123 parser.add_argument("--picoclvr_height", type=int, default=12)
 124
 125 parser.add_argument("--picoclvr_width", type=int, default=16)
 126
 127 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 128
 129 ##############################
 130 # Maze options
 131
 132 parser.add_argument("--maze_height", type=int, default=13)
 133
 134 parser.add_argument("--maze_width", type=int, default=21)
 135
 136 parser.add_argument("--maze_nb_walls", type=int, default=15)
 137
 138 ##############################
 139 # Snake options
 140
 141 parser.add_argument("--snake_height", type=int, default=9)
 142
 143 parser.add_argument("--snake_width", type=int, default=12)
 144
 145 parser.add_argument("--snake_nb_colors", type=int, default=5)
 146
 147 parser.add_argument("--snake_length", type=int, default=200)
 148
 149 ##############################
 150 # ByHeart options
 151
 152 parser.add_argument("--byheart_separation", type=int, default=1)
 153
 154 ##############################
 155 # Stack options
 156
 157 parser.add_argument("--stack_nb_steps", type=int, default=100)
 158
 159 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 160
 161 parser.add_argument("--stack_nb_digits", type=int, default=3)
 162
 163 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 164
 165 ##############################
 166 # Expr options
 167
 168 parser.add_argument("--expr_nb_variables", type=int, default=5)
 169
 170 parser.add_argument("--expr_sequence_length", type=int, default=40)
 171
 172 parser.add_argument("--expr_operand_max", type=int, default=9)
 173
 174 parser.add_argument("--expr_result_max", type=int, default=99)
 175
 176 parser.add_argument("--expr_input_file", type=str, default=None)
 177
 178 ##############################
 179 # Mixing
 180
 181 parser.add_argument("--mixing_hard", action="store_true", default=False)
 182
 183 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 184
 185 ##############################
 186 # greed options
 187
 188 parser.add_argument("--greed_height", type=int, default=5)
 189
 190 parser.add_argument("--greed_width", type=int, default=7)
 191
 192 parser.add_argument("--greed_T", type=int, default=25)
 193
 194 parser.add_argument("--greed_nb_walls", type=int, default=5)
 195
 196 parser.add_argument("--greed_nb_coins", type=int, default=2)
 197
 198 ######################################################################
 199
 200 args = parser.parse_args()
 201
 202 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 203
 204 if args.result_dir is None:
 205     args.result_dir = f"results_{args.task}"
 206
 207 ######################################################################
 208
 209 default_task_args = {
 210     "file": {
 211         "model": "37M",
 212         "batch_size": 25,
 213         "nb_train_samples": 250000,
 214         "nb_test_samples": 10000,
 215     },
 216     "addition": {
 217         "model": "352M",
 218         "batch_size": 25,
 219         "nb_train_samples": 250000,
 220         "nb_test_samples": 10000,
 221     },
 222     "byheart": {
 223         "model": "37M",
 224         "batch_size": 25,
 225         "nb_train_samples": 50000,
 226         "nb_test_samples": 10000,
 227     },
 228     "expr": {
 229         "model": "352M",
 230         "batch_size": 25,
 231         "nb_train_samples": 2500000,
 232         "nb_test_samples": 10000,
 233     },
 234     "grid": {
 235         "model": "37M",
 236         "batch_size": 25,
 237         "nb_train_samples": 250000,
 238         "nb_test_samples": 10000,
 239     },
 240     "qmlp": {
 241         "model": "37M",
 242         "batch_size": 10,
 243         "nb_train_samples": 100000,
 244         "nb_test_samples": 1000,
 245     },
 246     "guessop": {
 247         "model": "352M",
 248         "batch_size": 25,
 249         "nb_train_samples": 1000000,
 250         "nb_test_samples": 10000,
 251     },
 252     "learnop": {
 253         "model": "37M",
 254         "batch_size": 25,
 255         "nb_train_samples": 50000,
 256         "nb_test_samples": 10000,
 257     },
 258     "maze": {
 259         "model": "37M",
 260         "batch_size": 5,
 261         "nb_train_samples": 100000,
 262         "nb_test_samples": 10000,
 263     },
 264     "picoclvr": {
 265         "model": "37M",
 266         "batch_size": 25,
 267         "nb_train_samples": 250000,
 268         "nb_test_samples": 10000,
 269     },
 270     "rpl": {
 271         "model": "352M",
 272         "batch_size": 5,
 273         "nb_train_samples": 2500000,
 274         "nb_test_samples": 10000,
 275     },
 276     "snake": {
 277         "model": "37M",
 278         "batch_size": 25,
 279         "nb_train_samples": 250000,
 280         "nb_test_samples": 10000,
 281     },
 282     "stack": {
 283         "model": "37M",
 284         "batch_size": 25,
 285         "nb_train_samples": 100000,
 286         "nb_test_samples": 1000,
 287     },
 288     "twotargets": {
 289         "model": "37M",
 290         "batch_size": 25,
 291         "nb_train_samples": 50000,
 292         "nb_test_samples": 10000,
 293     },
 294     "memory": {
 295         "model": "37M",
 296         "batch_size": 100,
 297         "nb_train_samples": 25000,
 298         "nb_test_samples": 1000,
 299     },
 300     "mixing": {
 301         "model": "37M",
 302         "batch_size": 25,
 303         "nb_train_samples": 250000,
 304         "nb_test_samples": 10000,
 305     },
 306     "mnist": {
 307         "model": "37M",
 308         "batch_size": 10,
 309         "nb_train_samples": 60000,
 310         "nb_test_samples": 10000,
 311     },
 312     "greed": {
 313         "model": "37M",
 314         "batch_size": 25,
 315         "nb_train_samples": 25000,
 316         "nb_test_samples": 10000,
 317     },
 318 }
 319
 320 if args.task in default_task_args:
 321     for k, v in default_task_args[args.task].items():
 322         if getattr(args, k) is None:
 323             setattr(args, k, v)
 324
 325 ######################################################################
 326
 327 default_model_args = {
 328     "17K": {
 329         "dim_model": 32,
 330         "dim_keys": 32,
 331         "dim_hidden": 32,
 332         "nb_heads": 2,
 333         "nb_blocks": 2,
 334     },
 335     "4M": {
 336         "dim_model": 256,
 337         "dim_keys": 32,
 338         "dim_hidden": 1024,
 339         "nb_heads": 4,
 340         "nb_blocks": 6,
 341     },
 342     "37M": {
 343         "dim_model": 512,
 344         "dim_keys": 64,
 345         "dim_hidden": 2048,
 346         "nb_heads": 8,
 347         "nb_blocks": 12,
 348     },
 349     "122M": {
 350         "dim_model": 768,
 351         "dim_keys": 64,
 352         "dim_hidden": 2048,
 353         "nb_heads": 8,
 354         "nb_blocks": 24,
 355     },
 356     "352M": {
 357         "dim_model": 1024,
 358         "dim_keys": 64,
 359         "dim_hidden": 2048,
 360         "nb_heads": 8,
 361         "nb_blocks": 48,
 362     },
 363 }
 364
 365 if args.model in default_model_args:
 366     for k, v in default_model_args[args.model].items():
 367         if getattr(args, k) is None:
 368             setattr(args, k, v)
 369 else:
 370     raise ValueError(f"Unknown model {args.model}")
 371
 372 ######################################################################
 373
 374 try:
 375     os.mkdir(args.result_dir)
 376 except FileExistsError:
 377     if not args.resume:
 378         print(f"result directory {args.result_dir} already exists")
 379         exit(1)
 380
 381 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 382
 383 if args.seed >= 0:
 384     # torch.backends.cudnn.deterministic = True
 385     # torch.backends.cudnn.benchmark = False
 386     # torch.use_deterministic_algorithms(True)
 387     torch.manual_seed(args.seed)
 388     if torch.cuda.is_available():
 389         torch.cuda.manual_seed_all(args.seed)
 390
 391 ######################################################################
 392
 393
 394 def log_string(s):
 395     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 396
 397     if log_file is not None:
 398         log_file.write(t + s + "\n")
 399         log_file.flush()
 400
 401     print(t + s)
 402     sys.stdout.flush()
 403
 404
 405 log_string(f"argv {' '.join(sys.argv)}")
 406
 407 for n in vars(args):
 408     log_string(f"args.{n} {getattr(args, n)}")
 409
 410
 411 ######################################################################
 412
 413
 414 def picoclvr_pruner_horizontal_green(p):
 415     return not ("green" in p and ("left" in p or "right" in p))
 416
 417
 418 picoclvr_pruner_train = (
 419     picoclvr_pruner_horizontal_green
 420     if args.picocvlr_prune_properties in {"train+eval"}
 421     else None
 422 )
 423
 424 picoclvr_pruner_eval = (
 425     (lambda p: not picoclvr_pruner_horizontal_green(p))
 426     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 427     else None
 428 )
 429
 430 ######################################################################
 431
 432 if args.physical_batch_size is None:
 433     args.physical_batch_size = args.batch_size
 434 else:
 435     assert args.batch_size % args.physical_batch_size == 0
 436
 437 assert args.nb_train_samples % args.batch_size == 0
 438 assert args.nb_test_samples % args.batch_size == 0
 439
 440 if args.task == "file":
 441     assert (
 442         args.filetask_train_file is not None and args.filetask_test_file is not None
 443     ), "You have to specify the task train and test files"
 444     task = tasks.TaskFromFile(
 445         args.filetask_train_file,
 446         args.filetask_test_file,
 447         nb_train_samples=args.nb_train_samples,
 448         nb_test_samples=args.nb_test_samples,
 449         batch_size=args.physical_batch_size,
 450         shuffle=True,
 451         device=device,
 452     )
 453     args.max_percents_of_test_in_train = 0
 454
 455 elif args.task == "byheart":
 456     task = tasks.SandBox(
 457         problem=problems.ProblemByHeart(separation=args.byheart_separation),
 458         nb_train_samples=args.nb_train_samples,
 459         nb_test_samples=args.nb_test_samples,
 460         batch_size=args.physical_batch_size,
 461         logger=log_string,
 462         device=device,
 463     )
 464     args.max_percents_of_test_in_train = -1
 465
 466 elif args.task == "learnop":
 467     task = tasks.SandBox(
 468         problem=problems.ProblemLearnOperator(),
 469         nb_train_samples=args.nb_train_samples,
 470         nb_test_samples=args.nb_test_samples,
 471         batch_size=args.physical_batch_size,
 472         logger=log_string,
 473         device=device,
 474     )
 475
 476
 477 elif args.task == "guessop":
 478     task = tasks.SandBox(
 479         problem=problems.ProblemGuessOperator(),
 480         nb_train_samples=args.nb_train_samples,
 481         nb_test_samples=args.nb_test_samples,
 482         batch_size=args.physical_batch_size,
 483         logger=log_string,
 484         device=device,
 485     )
 486
 487
 488 elif args.task == "twotargets":
 489     task = tasks.SandBox(
 490         problem=problems.ProblemTwoTargets(),
 491         nb_train_samples=args.nb_train_samples,
 492         nb_test_samples=args.nb_test_samples,
 493         batch_size=args.physical_batch_size,
 494         logger=log_string,
 495         device=device,
 496     )
 497
 498 elif args.task == "memory":
 499     task = tasks.SandBox(
 500         problem=problems.ProblemMemory(),
 501         nb_train_samples=args.nb_train_samples,
 502         nb_test_samples=args.nb_test_samples,
 503         batch_size=args.physical_batch_size,
 504         logger=log_string,
 505         device=device,
 506     )
 507
 508 elif args.task == "mixing":
 509     task = tasks.SandBox(
 510         problem=problems.ProblemMixing(
 511             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 512         ),
 513         nb_train_samples=args.nb_train_samples,
 514         nb_test_samples=args.nb_test_samples,
 515         batch_size=args.physical_batch_size,
 516         logger=log_string,
 517         device=device,
 518     )
 519
 520 elif args.task == "addition":
 521     task = tasks.SandBox(
 522         problem=problems.ProblemAddition(),
 523         nb_train_samples=args.nb_train_samples,
 524         nb_test_samples=args.nb_test_samples,
 525         batch_size=args.physical_batch_size,
 526         logger=log_string,
 527         device=device,
 528     )
 529
 530 elif args.task == "picoclvr":
 531     task = tasks.PicoCLVR(
 532         nb_train_samples=args.nb_train_samples,
 533         nb_test_samples=args.nb_test_samples,
 534         batch_size=args.physical_batch_size,
 535         height=args.picoclvr_height,
 536         width=args.picoclvr_width,
 537         nb_colors=args.picoclvr_nb_colors,
 538         logger=log_string,
 539         device=device,
 540         pruner_train=picoclvr_pruner_train,
 541         pruner_eval=picoclvr_pruner_eval,
 542     )
 543
 544 elif args.task == "mnist":
 545     task = tasks.MNIST(
 546         nb_train_samples=args.nb_train_samples,
 547         nb_test_samples=args.nb_test_samples,
 548         batch_size=args.physical_batch_size,
 549         device=device,
 550     )
 551
 552 elif args.task == "maze":
 553     task = tasks.Maze(
 554         nb_train_samples=args.nb_train_samples,
 555         nb_test_samples=args.nb_test_samples,
 556         batch_size=args.physical_batch_size,
 557         height=args.maze_height,
 558         width=args.maze_width,
 559         nb_walls=args.maze_nb_walls,
 560         device="cpu",
 561     )
 562
 563 elif args.task == "snake":
 564     task = tasks.Snake(
 565         nb_train_samples=args.nb_train_samples,
 566         nb_test_samples=args.nb_test_samples,
 567         batch_size=args.physical_batch_size,
 568         height=args.snake_height,
 569         width=args.snake_width,
 570         nb_colors=args.snake_nb_colors,
 571         length=args.snake_length,
 572         prompt_length=args.snake_length // 2,
 573         device=device,
 574     )
 575
 576 elif args.task == "stack":
 577     task = tasks.Stack(
 578         nb_train_samples=args.nb_train_samples,
 579         nb_test_samples=args.nb_test_samples,
 580         batch_size=args.physical_batch_size,
 581         logger=log_string,
 582         nb_steps=args.stack_nb_steps,
 583         nb_stacks=args.stack_nb_stacks,
 584         nb_digits=args.stack_nb_digits,
 585         fraction_values_for_train=args.stack_fraction_values_for_train,
 586         device=device,
 587     )
 588
 589 elif args.task == "expr":
 590     task = tasks.Expr(
 591         nb_train_samples=args.nb_train_samples,
 592         nb_test_samples=args.nb_test_samples,
 593         nb_variables=args.expr_nb_variables,
 594         sequence_length=args.expr_sequence_length,
 595         operand_max=args.expr_operand_max,
 596         result_max=args.expr_result_max,
 597         batch_size=args.physical_batch_size,
 598         device=device,
 599     )
 600
 601 elif args.task == "rpl":
 602     task = tasks.RPL(
 603         nb_train_samples=args.nb_train_samples,
 604         nb_test_samples=args.nb_test_samples,
 605         batch_size=args.physical_batch_size,
 606         nb_starting_values=args.rpl_nb_starting_values,
 607         max_input=args.rpl_max_input,
 608         prog_len=args.rpl_prog_len,
 609         nb_runs=args.rpl_nb_runs,
 610         no_prog=args.rpl_no_prog,
 611         logger=log_string,
 612         device=device,
 613     )
 614
 615 elif args.task == "grid":
 616     task = tasks.Grid(
 617         nb_train_samples=args.nb_train_samples,
 618         nb_test_samples=args.nb_test_samples,
 619         batch_size=args.physical_batch_size,
 620         size=args.grid_size,
 621         fraction_play=args.grid_fraction_play,
 622         logger=log_string,
 623         device=device,
 624     )
 625
 626 elif args.task == "qmlp":
 627     task = tasks.QMLP(
 628         nb_train_samples=args.nb_train_samples,
 629         nb_test_samples=args.nb_test_samples,
 630         batch_size=args.physical_batch_size,
 631         result_dir=args.result_dir,
 632         logger=log_string,
 633         device=device,
 634     )
 635
 636 elif args.task == "greed":
 637     task = tasks.Greed(
 638         nb_train_samples=args.nb_train_samples,
 639         nb_test_samples=args.nb_test_samples,
 640         batch_size=args.physical_batch_size,
 641         height=args.greed_height,
 642         width=args.greed_width,
 643         T=args.greed_T,
 644         nb_walls=args.greed_nb_walls,
 645         nb_coins=args.greed_nb_coins,
 646         logger=log_string,
 647         device=device,
 648     )
 649
 650 else:
 651     raise ValueError(f"Unknown task {args.task}")
 652
 653 ######################################################################
 654
 655 log_string(f"device {device}")
 656
 657 vocabulary_size = task.vocabulary_size()
 658
 659 log_string(f"vocabulary_size {vocabulary_size}")
 660
 661 ##############################
 662
 663 model = mygpt.MyGPT(
 664     vocabulary_size=vocabulary_size,
 665     dim_model=args.dim_model,
 666     dim_keys=args.dim_keys,
 667     dim_hidden=args.dim_hidden,
 668     nb_heads=args.nb_heads,
 669     nb_blocks=args.nb_blocks,
 670     causal=True,
 671     dropout=args.dropout,
 672 )
 673
 674 model.to(device)
 675
 676 nb_parameters = sum(p.numel() for p in model.parameters())
 677 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 678
 679 ######################################################################
 680
 681 nb_epochs_finished = 0
 682
 683 if args.no_checkpoint:
 684     log_string(f"not trying to load checkpoint.")
 685
 686 else:
 687     try:
 688         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 689         checkpoint = torch.load(checkpoint_name)
 690         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 691         model.load_state_dict(checkpoint["model_state"])
 692         torch.set_rng_state(checkpoint["rng_state"])
 693         if torch.cuda.is_available():
 694             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 695
 696         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 697
 698     except FileNotFoundError:
 699         log_string("starting from scratch.")
 700
 701     except:
 702         log_string("error when loading the checkpoint.")
 703         exit(1)
 704
 705 ######################################################################
 706
 707 if args.task == "expr" and args.expr_input_file is not None:
 708     task.produce_results(
 709         n_epoch=nb_epochs_finished,
 710         model=model,
 711         result_dir=args.result_dir,
 712         logger=log_string,
 713         deterministic_synthesis=args.deterministic_synthesis,
 714         input_file=args.expr_input_file,
 715     )
 716
 717     exit(0)
 718
 719 ######################################################################
 720
 721 # Compute the entropy of the training tokens
 722
 723 token_count = 0
 724 for input in task.batches(split="train", desc="train-entropy"):
 725     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 726 token_probas = token_count / token_count.sum()
 727 entropy = -torch.xlogy(token_probas, token_probas).sum()
 728 train_set_perplexity = math.exp(entropy)
 729
 730 ######################################################################
 731 # A bit of paranoia never hurts
 732
 733 if args.max_percents_of_test_in_train >= 0:
 734
 735     def subsets_as_tuples(batches, cs):
 736         s = set()
 737         for batch in batches:
 738             for x in batch:
 739                 s.add(tuple([v.item() for v in x]))
 740                 if len(s) == cs:
 741                     yield s
 742                     s = set()
 743         yield s
 744
 745     nb_test, nb_in_train = 0, 0
 746     for test_subset in subsets_as_tuples(
 747         task.batches(split="test", desc="test-check"), 25000
 748     ):
 749         in_train = set()
 750         for train_subset in subsets_as_tuples(
 751             task.batches(split="train", desc="train-check"), 25000
 752         ):
 753             in_train.update(test_subset.intersection(train_subset))
 754         nb_in_train += len(in_train)
 755         nb_test += len(test_subset)
 756
 757     log_string(
 758         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 759     )
 760
 761     assert (
 762         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 763     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 764
 765 ##############################
 766
 767 if args.learning_rate_schedule == "cos":
 768     learning_rate_schedule = {}
 769     for n_epoch in range(args.nb_epochs):
 770         u = n_epoch / args.nb_epochs * math.pi
 771         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 772 else:
 773     u = {
 774         int(k): float(v)
 775         for k, v in [
 776             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 777         ]
 778     }
 779
 780     learning_rate_schedule = {}
 781     learning_rate = args.learning_rate
 782     for n_epoch in range(args.nb_epochs):
 783         if n_epoch in u:
 784             learning_rate = u[n_epoch]
 785         learning_rate_schedule[n_epoch] = learning_rate
 786
 787 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 788
 789 ##############################
 790
 791 if nb_epochs_finished >= args.nb_epochs:
 792     task.produce_results(
 793         n_epoch=nb_epochs_finished,
 794         model=model,
 795         result_dir=args.result_dir,
 796         logger=log_string,
 797         deterministic_synthesis=args.deterministic_synthesis,
 798     )
 799
 800 time_pred_result = None
 801
 802 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 803     learning_rate = learning_rate_schedule[n_epoch]
 804
 805     log_string(f"learning_rate {learning_rate}")
 806
 807     if args.optim == "sgd":
 808         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 809     elif args.optim == "adam":
 810         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 811     elif args.optim == "adamw":
 812         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 813     else:
 814         raise ValueError(f"Unknown optimizer {args.optim}.")
 815
 816     model.train()
 817
 818     nb_train_samples, acc_train_loss = 0, 0.0
 819
 820     for input in task.batches(split="train"):
 821         input = input.to(device)
 822
 823         if nb_train_samples % args.batch_size == 0:
 824             optimizer.zero_grad()
 825
 826         output = model(mygpt.BracketedSequence(input)).x
 827         loss = F.cross_entropy(output.transpose(1, 2), input)
 828         acc_train_loss += loss.item() * input.size(0)
 829
 830         nb_train_samples += input.size(0)
 831
 832         loss.backward()
 833
 834         if nb_train_samples % args.batch_size == 0:
 835             optimizer.step()
 836
 837     with torch.autograd.no_grad():
 838         model.eval()
 839
 840         nb_test_samples, acc_test_loss = 0, 0.0
 841         nb_samples_accumulated = 0
 842
 843         for input in task.batches(split="test"):
 844             input = input.to(device)
 845
 846             bs = model(mygpt.BracketedSequence(input))
 847             output = bs.x
 848
 849             loss = F.cross_entropy(output.transpose(1, 2), input)
 850
 851             acc_test_loss += loss.item() * input.size(0)
 852
 853             nb_test_samples += input.size(0)
 854
 855         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 856         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 857
 858         log_string(
 859             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 860         )
 861
 862         task.produce_results(
 863             n_epoch=n_epoch,
 864             model=model,
 865             result_dir=args.result_dir,
 866             logger=log_string,
 867             deterministic_synthesis=args.deterministic_synthesis,
 868         )
 869
 870         time_current_result = datetime.datetime.now()
 871         if time_pred_result is not None:
 872             log_string(
 873                 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
 874             )
 875         time_pred_result = time_current_result
 876
 877     checkpoint = {
 878         "nb_epochs_finished": n_epoch + 1,
 879         "model_state": model.state_dict(),
 880         "rng_state": torch.get_rng_state(),
 881     }
 882
 883     if torch.cuda.is_available():
 884         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 885
 886     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 887     torch.save(checkpoint, checkpoint_name)
 888     log_string(f"saved checkpoint {checkpoint_name}")
 889
 890 ######################################################################