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
   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, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl",
  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=None)
  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 parser.add_argument("--learning_rate", type=float, default=1e-4)
  60
  61 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  62
  63 ########################################
  64
  65 parser.add_argument("--model", type=str, default="37M")
  66
  67 parser.add_argument("--dim_model", type=int, default=None)
  68
  69 parser.add_argument("--dim_keys", type=int, default=None)
  70
  71 parser.add_argument("--dim_hidden", type=int, default=None)
  72
  73 parser.add_argument("--nb_heads", type=int, default=None)
  74
  75 parser.add_argument("--nb_blocks", type=int, default=None)
  76
  77 parser.add_argument("--dropout", type=float, default=0.1)
  78
  79 ########################################
  80
  81 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  82
  83 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  84
  85 parser.add_argument("--overwrite_results", action="store_true", default=False)
  86
  87 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  88
  89 ##############################
  90 # rpl options
  91
  92 parser.add_argument("--rpl_nb_starting_values", type=int, default=5)
  93
  94 parser.add_argument("--rpl_max_input", type=int, default=9)
  95
  96 parser.add_argument("--rpl_prog_len", type=int, default=10)
  97
  98 parser.add_argument("--rpl_nb_runs", type=int, default=8)
  99
 100 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 101
 102 ##############################
 103 # picoclvr options
 104
 105 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 106
 107 parser.add_argument("--picoclvr_height", type=int, default=12)
 108
 109 parser.add_argument("--picoclvr_width", type=int, default=16)
 110
 111 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 112
 113 ##############################
 114 # Maze options
 115
 116 parser.add_argument("--maze_height", type=int, default=23)
 117
 118 parser.add_argument("--maze_width", type=int, default=39)
 119
 120 parser.add_argument("--maze_nb_walls", type=int, default=45)
 121
 122 ##############################
 123 # Snake options
 124
 125 parser.add_argument("--snake_height", type=int, default=9)
 126
 127 parser.add_argument("--snake_width", type=int, default=12)
 128
 129 parser.add_argument("--snake_nb_colors", type=int, default=5)
 130
 131 parser.add_argument("--snake_length", type=int, default=200)
 132
 133 ##############################
 134 # Stack options
 135
 136 parser.add_argument("--stack_nb_steps", type=int, default=100)
 137
 138 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 139
 140 parser.add_argument("--stack_nb_digits", type=int, default=3)
 141
 142 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 143
 144 ##############################
 145 # Expr options
 146
 147 parser.add_argument("--expr_nb_variables", type=int, default=5)
 148
 149 parser.add_argument("--expr_sequence_length", type=int, default=40)
 150
 151 parser.add_argument("--expr_operand_max", type=int, default=9)
 152
 153 parser.add_argument("--expr_result_max", type=int, default=99)
 154
 155 parser.add_argument("--expr_input_file", type=str, default=None)
 156
 157 ##############################
 158 # World options
 159
 160 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 161
 162 ######################################################################
 163
 164 args = parser.parse_args()
 165
 166 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 167
 168 if args.result_dir is None:
 169     args.result_dir = f"results_{args.task}"
 170
 171 ######################################################################
 172
 173 default_task_args = {
 174     "byheart": {
 175         "nb_epochs": 5,
 176         "batch_size": 25,
 177         "nb_train_samples": 50000,
 178         "nb_test_samples": 10000,
 179     },
 180     "learnop": {
 181         "nb_epochs": 5,
 182         "batch_size": 25,
 183         "nb_train_samples": 50000,
 184         "nb_test_samples": 10000,
 185     },
 186     "guessop": {
 187         "nb_epochs": 5,
 188         "batch_size": 25,
 189         "nb_train_samples": 50000,
 190         "nb_test_samples": 10000,
 191     },
 192     "twotargets": {
 193         "nb_epochs": 5,
 194         "batch_size": 25,
 195         "nb_train_samples": 50000,
 196         "nb_test_samples": 10000,
 197     },
 198     "addition": {
 199         "nb_epochs": 5,
 200         "batch_size": 25,
 201         "nb_train_samples": 50000,
 202         "nb_test_samples": 10000,
 203     },
 204     "picoclvr": {
 205         "nb_epochs": 25,
 206         "batch_size": 25,
 207         "nb_train_samples": 250000,
 208         "nb_test_samples": 10000,
 209     },
 210     "mnist": {
 211         "nb_epochs": 25,
 212         "batch_size": 10,
 213         "nb_train_samples": 60000,
 214         "nb_test_samples": 10000,
 215     },
 216     "maze": {
 217         "nb_epochs": 25,
 218         "batch_size": 5,
 219         "nb_train_samples": 250000,
 220         "nb_test_samples": 10000,
 221     },
 222     "snake": {
 223         "nb_epochs": 5,
 224         "batch_size": 25,
 225         "nb_train_samples": 50000,
 226         "nb_test_samples": 10000,
 227     },
 228     "stack": {
 229         "nb_epochs": 5,
 230         "batch_size": 25,
 231         "nb_train_samples": 100000,
 232         "nb_test_samples": 1000,
 233     },
 234     "expr": {
 235         "nb_epochs": 40,
 236         "batch_size": 25,
 237         "nb_train_samples": 1000000,
 238         "nb_test_samples": 10000,
 239     },
 240     "rpl": {
 241         "nb_epochs": 40,
 242         "batch_size": 25,
 243         "nb_train_samples": 100000,
 244         "nb_test_samples": 10000,
 245     },
 246     "world": {
 247         "nb_epochs": 10,
 248         "batch_size": 25,
 249         "nb_train_samples": 25000,
 250         "nb_test_samples": 1000,
 251     },
 252 }
 253
 254 if args.task in default_task_args:
 255     for k, v in default_task_args[args.task].items():
 256         if getattr(args, k) is None:
 257             setattr(args, k, v)
 258
 259 ######################################################################
 260
 261 default_model_args = {
 262     "17K": {
 263         "dim_model": 32,
 264         "dim_keys": 32,
 265         "dim_hidden": 32,
 266         "nb_heads": 2,
 267         "nb_blocks": 2,
 268     },
 269     "37M": {
 270         "dim_model": 512,
 271         "dim_keys": 64,
 272         "dim_hidden": 2048,
 273         "nb_heads": 8,
 274         "nb_blocks": 12,
 275     },
 276     "122M": {
 277         "dim_model": 768,
 278         "dim_keys": 64,
 279         "dim_hidden": 2048,
 280         "nb_heads": 8,
 281         "nb_blocks": 24,
 282     },
 283     "352M": {
 284         "dim_model": 1024,
 285         "dim_keys": 64,
 286         "dim_hidden": 2048,
 287         "nb_heads": 8,
 288         "nb_blocks": 48,
 289     },
 290 }
 291
 292 if args.model in default_model_args:
 293     for k, v in default_model_args[args.model].items():
 294         if getattr(args, k) is None:
 295             setattr(args, k, v)
 296 else:
 297     raise ValueError(f"Unknown model {args.model}")
 298
 299 ######################################################################
 300
 301 try:
 302     os.mkdir(args.result_dir)
 303 except FileExistsError:
 304     if not args.overwrite_results:
 305         print(f"result directory {args.result_dir} already exists")
 306         exit(1)
 307
 308 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 309
 310 if args.seed >= 0:
 311     # torch.backends.cudnn.deterministic = True
 312     # torch.backends.cudnn.benchmark = False
 313     # torch.use_deterministic_algorithms(True)
 314     torch.manual_seed(args.seed)
 315     if torch.cuda.is_available():
 316         torch.cuda.manual_seed_all(args.seed)
 317
 318 ######################################################################
 319
 320
 321 def log_string(s):
 322     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 323
 324     if log_file is not None:
 325         log_file.write(t + s + "\n")
 326         log_file.flush()
 327
 328     print(t + s)
 329     sys.stdout.flush()
 330
 331
 332 for n in vars(args):
 333     log_string(f"args.{n} {getattr(args, n)}")
 334
 335
 336 ######################################################################
 337
 338
 339 def picoclvr_pruner_horizontal_green(p):
 340     return not ("green" in p and ("left" in p or "right" in p))
 341
 342
 343 picoclvr_pruner_train = (
 344     picoclvr_pruner_horizontal_green
 345     if args.picocvlr_prune_properties in {"train+eval"}
 346     else None
 347 )
 348
 349 picoclvr_pruner_eval = (
 350     (lambda p: not picoclvr_pruner_horizontal_green(p))
 351     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 352     else None
 353 )
 354
 355 ######################################################################
 356
 357 if args.task == "byheart":
 358     task = tasks.SandBox(
 359         problem=problems.ProblemByHeart(),
 360         nb_train_samples=args.nb_train_samples,
 361         nb_test_samples=args.nb_test_samples,
 362         batch_size=args.batch_size,
 363         logger=log_string,
 364         device=device,
 365     )
 366     args.max_percents_of_test_in_train = -1
 367
 368 elif args.task == "learnop":
 369     task = tasks.SandBox(
 370         problem=problems.ProblemLearnOperator(),
 371         nb_train_samples=args.nb_train_samples,
 372         nb_test_samples=args.nb_test_samples,
 373         batch_size=args.batch_size,
 374         logger=log_string,
 375         device=device,
 376     )
 377
 378
 379 elif args.task == "guessop":
 380     task = tasks.SandBox(
 381         problem=problems.ProblemGuessOperator(),
 382         nb_train_samples=args.nb_train_samples,
 383         nb_test_samples=args.nb_test_samples,
 384         batch_size=args.batch_size,
 385         logger=log_string,
 386         device=device,
 387     )
 388
 389
 390 elif args.task == "twotargets":
 391     task = tasks.SandBox(
 392         problem=problems.ProblemTwoTargets(),
 393         nb_train_samples=args.nb_train_samples,
 394         nb_test_samples=args.nb_test_samples,
 395         batch_size=args.batch_size,
 396         logger=log_string,
 397         device=device,
 398     )
 399
 400 elif args.task == "addition":
 401     task = tasks.SandBox(
 402         problem=problems.ProblemAddition(),
 403         nb_train_samples=args.nb_train_samples,
 404         nb_test_samples=args.nb_test_samples,
 405         batch_size=args.batch_size,
 406         logger=log_string,
 407         device=device,
 408     )
 409
 410 elif args.task == "picoclvr":
 411     task = tasks.PicoCLVR(
 412         nb_train_samples=args.nb_train_samples,
 413         nb_test_samples=args.nb_test_samples,
 414         batch_size=args.batch_size,
 415         height=args.picoclvr_height,
 416         width=args.picoclvr_width,
 417         nb_colors=args.picoclvr_nb_colors,
 418         logger=log_string,
 419         device=device,
 420         pruner_train=picoclvr_pruner_train,
 421         pruner_eval=picoclvr_pruner_eval,
 422     )
 423
 424 elif args.task == "mnist":
 425     task = tasks.MNIST(
 426         nb_train_samples=args.nb_train_samples,
 427         nb_test_samples=args.nb_test_samples,
 428         batch_size=args.batch_size,
 429         device=device,
 430     )
 431
 432 elif args.task == "maze":
 433     task = tasks.Maze(
 434         nb_train_samples=args.nb_train_samples,
 435         nb_test_samples=args.nb_test_samples,
 436         batch_size=args.batch_size,
 437         height=args.maze_height,
 438         width=args.maze_width,
 439         nb_walls=args.maze_nb_walls,
 440         device=device,
 441     )
 442
 443 elif args.task == "snake":
 444     task = tasks.Snake(
 445         nb_train_samples=args.nb_train_samples,
 446         nb_test_samples=args.nb_test_samples,
 447         batch_size=args.batch_size,
 448         height=args.snake_height,
 449         width=args.snake_width,
 450         nb_colors=args.snake_nb_colors,
 451         length=args.snake_length,
 452         prompt_length=args.snake_length // 2,
 453         device=device,
 454     )
 455
 456 elif args.task == "stack":
 457     task = tasks.Stack(
 458         nb_train_samples=args.nb_train_samples,
 459         nb_test_samples=args.nb_test_samples,
 460         batch_size=args.batch_size,
 461         logger=log_string,
 462         nb_steps=args.stack_nb_steps,
 463         nb_stacks=args.stack_nb_stacks,
 464         nb_digits=args.stack_nb_digits,
 465         fraction_values_for_train=args.stack_fraction_values_for_train,
 466         device=device,
 467     )
 468
 469 elif args.task == "expr":
 470     task = tasks.Expr(
 471         nb_train_samples=args.nb_train_samples,
 472         nb_test_samples=args.nb_test_samples,
 473         nb_variables=args.expr_nb_variables,
 474         sequence_length=args.expr_sequence_length,
 475         operand_max=args.expr_operand_max,
 476         result_max=args.expr_result_max,
 477         batch_size=args.batch_size,
 478         device=device,
 479     )
 480
 481 elif args.task == "rpl":
 482     task = tasks.RPL(
 483         nb_train_samples=args.nb_train_samples,
 484         nb_test_samples=args.nb_test_samples,
 485         batch_size=args.batch_size,
 486         nb_starting_values=args.rpl_nb_starting_values,
 487         max_input=args.rpl_max_input,
 488         prog_len=args.rpl_prog_len,
 489         nb_runs=args.rpl_nb_runs,
 490         no_prog=args.rpl_no_prog,
 491         logger=log_string,
 492         device=device,
 493     )
 494
 495 elif args.task == "world":
 496     task = tasks.World(
 497         nb_train_samples=args.nb_train_samples,
 498         nb_test_samples=args.nb_test_samples,
 499         batch_size=args.batch_size,
 500         vqae_nb_epochs=args.world_vqae_nb_epochs,
 501         logger=log_string,
 502         device=device,
 503     )
 504
 505 else:
 506     raise ValueError(f"Unknown task {args.task}")
 507
 508 ######################################################################
 509
 510 log_string(f"device {device}")
 511
 512 vocabulary_size = task.vocabulary_size()
 513
 514 log_string(f"vocabulary_size {vocabulary_size}")
 515
 516 ##############################
 517
 518 model = mygpt.MyGPT(
 519     vocabulary_size=vocabulary_size,
 520     dim_model=args.dim_model,
 521     dim_keys=args.dim_keys,
 522     dim_hidden=args.dim_hidden,
 523     nb_heads=args.nb_heads,
 524     nb_blocks=args.nb_blocks,
 525     causal=True,
 526     dropout=args.dropout,
 527 )
 528
 529 model.to(device)
 530
 531 nb_parameters = sum(p.numel() for p in model.parameters())
 532 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 533
 534 ######################################################################
 535
 536 nb_epochs_finished = 0
 537
 538 if args.no_checkpoint:
 539     log_string(f"not trying to load checkpoint.")
 540
 541 else:
 542     try:
 543         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 544         checkpoint = torch.load(checkpoint_name)
 545         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 546         model.load_state_dict(checkpoint["model_state"])
 547         torch.set_rng_state(checkpoint["rng_state"])
 548         if torch.cuda.is_available():
 549             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 550
 551         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 552
 553     except FileNotFoundError:
 554         log_string("starting from scratch.")
 555
 556     except:
 557         log_string("error when loading the checkpoint.")
 558         exit(1)
 559
 560 ######################################################################
 561
 562 if args.task == "expr" and args.expr_input_file is not None:
 563     task.produce_results(
 564         n_epoch=nb_epochs_finished,
 565         model=model,
 566         result_dir=args.result_dir,
 567         logger=log_string,
 568         deterministic_synthesis=args.deterministic_synthesis,
 569         input_file=args.expr_input_file,
 570     )
 571
 572     exit(0)
 573
 574 ######################################################################
 575
 576 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 577
 578 # Compute the entropy of the training tokens
 579
 580 token_count = 0
 581 for input in task.batches(split="train"):
 582     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 583 token_probas = token_count / token_count.sum()
 584 entropy = -torch.xlogy(token_probas, token_probas).sum()
 585 train_set_perplexity = math.exp(entropy)
 586
 587 ######################################################################
 588 # A bit of paranoia never hurts
 589
 590 if args.max_percents_of_test_in_train >= 0:
 591
 592     def subsets_as_tuples(batches, cs):
 593         s = set()
 594         for batch in batches:
 595             for x in batch:
 596                 s.add(tuple([v.item() for v in x]))
 597                 if len(s) == cs:
 598                     yield s
 599                     s = set()
 600         yield s
 601
 602     nb_test, nb_in_train = 0, 0
 603     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 604         in_train = set()
 605         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 606             in_train.update(test_subset.intersection(train_subset))
 607         nb_in_train += len(in_train)
 608         nb_test += len(test_subset)
 609
 610     log_string(
 611         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 612     )
 613
 614     assert (
 615         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 616     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 617
 618 ##############################
 619
 620 if args.learning_rate_schedule == "cos":
 621     learning_rate_schedule = {}
 622     for n_epoch in range(args.nb_epochs):
 623         u = n_epoch / args.nb_epochs * math.pi
 624         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 625 else:
 626     u = {
 627         int(k): float(v)
 628         for k, v in [
 629             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 630         ]
 631     }
 632
 633     learning_rate_schedule = {}
 634     learning_rate = args.learning_rate
 635     for n_epoch in range(args.nb_epochs):
 636         if n_epoch in u:
 637             learning_rate = u[n_epoch]
 638         learning_rate_schedule[n_epoch] = learning_rate
 639
 640 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 641
 642 ##############################
 643
 644 nb_samples_seen = 0
 645
 646 if nb_epochs_finished >= nb_epochs:
 647     task.produce_results(
 648         n_epoch=nb_epochs_finished,
 649         model=model,
 650         result_dir=args.result_dir,
 651         logger=log_string,
 652         deterministic_synthesis=args.deterministic_synthesis,
 653     )
 654
 655 for n_epoch in range(nb_epochs_finished, nb_epochs):
 656     learning_rate = learning_rate_schedule[n_epoch]
 657
 658     log_string(f"learning_rate {learning_rate}")
 659
 660     if args.optim == "sgd":
 661         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 662     elif args.optim == "adam":
 663         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 664     elif args.optim == "adamw":
 665         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 666     else:
 667         raise ValueError(f"Unknown optimizer {args.optim}.")
 668
 669     model.train()
 670
 671     nb_train_samples, acc_train_loss = 0, 0.0
 672
 673     for input in task.batches(split="train"):
 674         input = input.to(device)
 675         output = model(mygpt.BracketedSequence(input)).x
 676         loss = F.cross_entropy(output.transpose(1, 2), input)
 677         acc_train_loss += loss.item() * input.size(0)
 678         nb_train_samples += input.size(0)
 679         nb_samples_seen += input.size(0)
 680
 681         optimizer.zero_grad()
 682         loss.backward()
 683         optimizer.step()
 684
 685     with torch.autograd.no_grad():
 686         model.eval()
 687
 688         nb_test_samples, acc_test_loss = 0, 0.0
 689
 690         for input in task.batches(split="test"):
 691             input = input.to(device)
 692
 693             output = model(mygpt.BracketedSequence(input)).x
 694             loss = F.cross_entropy(output.transpose(1, 2), input)
 695             acc_test_loss += loss.item() * input.size(0)
 696             nb_test_samples += input.size(0)
 697
 698         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 699         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 700
 701         log_string(
 702             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 703         )
 704
 705         task.produce_results(
 706             n_epoch=n_epoch,
 707             model=model,
 708             result_dir=args.result_dir,
 709             logger=log_string,
 710             deterministic_synthesis=args.deterministic_synthesis,
 711         )
 712
 713     checkpoint = {
 714         "nb_epochs_finished": n_epoch + 1,
 715         "model_state": model.state_dict(),
 716         "rng_state": torch.get_rng_state(),
 717     }
 718
 719     if torch.cuda.is_available():
 720         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 721
 722     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 723     torch.save(checkpoint, checkpoint_name)
 724     log_string(f"saved checkpoint {checkpoint_name}")
 725
 726 ######################################################################