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=None)
  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=3)
  93
  94 parser.add_argument("--rpl_max_input", type=int, default=9)
  95
  96 parser.add_argument("--rpl_prog_len", type=int, default=8)
  97
  98 parser.add_argument("--rpl_nb_runs", type=int, default=5)
  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=13)
 117
 118 parser.add_argument("--maze_width", type=int, default=21)
 119
 120 parser.add_argument("--maze_nb_walls", type=int, default=15)
 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         "model": "37M",
 176         "nb_epochs": 2,
 177         "batch_size": 25,
 178         "nb_train_samples": 50000,
 179         "nb_test_samples": 10000,
 180     },
 181     "learnop": {
 182         "model": "37M",
 183         "nb_epochs": 15,
 184         "batch_size": 25,
 185         "nb_train_samples": 50000,
 186         "nb_test_samples": 10000,
 187     },
 188     "guessop": {
 189         "model": "352M",
 190         "nb_epochs": 5,
 191         "batch_size": 25,
 192         "nb_train_samples": 1000000,
 193         "nb_test_samples": 10000,
 194     },
 195     "twotargets": {
 196         "model": "37M",
 197         "nb_epochs": 10,
 198         "batch_size": 25,
 199         "nb_train_samples": 50000,
 200         "nb_test_samples": 10000,
 201     },
 202     "addition": {
 203         "model": "352M",
 204         "nb_epochs": 50,
 205         "batch_size": 25,
 206         "nb_train_samples": 250000,
 207         "nb_test_samples": 10000,
 208     },
 209     "picoclvr": {
 210         "model": "37M",
 211         "nb_epochs": 25,
 212         "batch_size": 25,
 213         "nb_train_samples": 250000,
 214         "nb_test_samples": 10000,
 215     },
 216     "mnist": {
 217         "model": "37M",
 218         "nb_epochs": 25,
 219         "batch_size": 10,
 220         "nb_train_samples": 60000,
 221         "nb_test_samples": 10000,
 222     },
 223     "maze": {
 224         "model": "37M",
 225         "nb_epochs": 25,
 226         "batch_size": 5,
 227         "nb_train_samples": 100000,
 228         "nb_test_samples": 10000,
 229     },
 230     "snake": {
 231         "model": "37M",
 232         "nb_epochs": 5,
 233         "batch_size": 25,
 234         "nb_train_samples": 250000,
 235         "nb_test_samples": 10000,
 236     },
 237     "stack": {
 238         "model": "37M",
 239         "nb_epochs": 15,
 240         "batch_size": 25,
 241         "nb_train_samples": 100000,
 242         "nb_test_samples": 1000,
 243     },
 244     "expr": {
 245         "model": "352M",
 246         "nb_epochs": 25,
 247         "batch_size": 25,
 248         "nb_train_samples": 2500000,
 249         "nb_test_samples": 10000,
 250     },
 251     "rpl": {
 252         "model": "122M",
 253         "nb_epochs": 50,
 254         "batch_size": 5,
 255         "nb_train_samples": 1000000,
 256         "nb_test_samples": 10000,
 257     },
 258     "world": {
 259         "model": "37M",
 260         "nb_epochs": 10,
 261         "batch_size": 25,
 262         "nb_train_samples": 25000,
 263         "nb_test_samples": 1000,
 264     },
 265 }
 266
 267 if args.task in default_task_args:
 268     for k, v in default_task_args[args.task].items():
 269         if getattr(args, k) is None:
 270             setattr(args, k, v)
 271
 272 ######################################################################
 273
 274 default_model_args = {
 275     "17K": {
 276         "dim_model": 32,
 277         "dim_keys": 32,
 278         "dim_hidden": 32,
 279         "nb_heads": 2,
 280         "nb_blocks": 2,
 281     },
 282     "37M": {
 283         "dim_model": 512,
 284         "dim_keys": 64,
 285         "dim_hidden": 2048,
 286         "nb_heads": 8,
 287         "nb_blocks": 12,
 288     },
 289     "122M": {
 290         "dim_model": 768,
 291         "dim_keys": 64,
 292         "dim_hidden": 2048,
 293         "nb_heads": 8,
 294         "nb_blocks": 24,
 295     },
 296     "352M": {
 297         "dim_model": 1024,
 298         "dim_keys": 64,
 299         "dim_hidden": 2048,
 300         "nb_heads": 8,
 301         "nb_blocks": 48,
 302     },
 303 }
 304
 305 if args.model in default_model_args:
 306     for k, v in default_model_args[args.model].items():
 307         if getattr(args, k) is None:
 308             setattr(args, k, v)
 309 else:
 310     raise ValueError(f"Unknown model {args.model}")
 311
 312 ######################################################################
 313
 314 try:
 315     os.mkdir(args.result_dir)
 316 except FileExistsError:
 317     if not args.overwrite_results:
 318         print(f"result directory {args.result_dir} already exists")
 319         exit(1)
 320
 321 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 322
 323 if args.seed >= 0:
 324     # torch.backends.cudnn.deterministic = True
 325     # torch.backends.cudnn.benchmark = False
 326     # torch.use_deterministic_algorithms(True)
 327     torch.manual_seed(args.seed)
 328     if torch.cuda.is_available():
 329         torch.cuda.manual_seed_all(args.seed)
 330
 331 ######################################################################
 332
 333
 334 def log_string(s):
 335     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 336
 337     if log_file is not None:
 338         log_file.write(t + s + "\n")
 339         log_file.flush()
 340
 341     print(t + s)
 342     sys.stdout.flush()
 343
 344
 345 for n in vars(args):
 346     log_string(f"args.{n} {getattr(args, n)}")
 347
 348
 349 ######################################################################
 350
 351
 352 def picoclvr_pruner_horizontal_green(p):
 353     return not ("green" in p and ("left" in p or "right" in p))
 354
 355
 356 picoclvr_pruner_train = (
 357     picoclvr_pruner_horizontal_green
 358     if args.picocvlr_prune_properties in {"train+eval"}
 359     else None
 360 )
 361
 362 picoclvr_pruner_eval = (
 363     (lambda p: not picoclvr_pruner_horizontal_green(p))
 364     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 365     else None
 366 )
 367
 368 ######################################################################
 369
 370 if args.task == "byheart":
 371     task = tasks.SandBox(
 372         problem=problems.ProblemByHeart(),
 373         nb_train_samples=args.nb_train_samples,
 374         nb_test_samples=args.nb_test_samples,
 375         batch_size=args.batch_size,
 376         logger=log_string,
 377         device=device,
 378     )
 379     args.max_percents_of_test_in_train = -1
 380
 381 elif args.task == "learnop":
 382     task = tasks.SandBox(
 383         problem=problems.ProblemLearnOperator(),
 384         nb_train_samples=args.nb_train_samples,
 385         nb_test_samples=args.nb_test_samples,
 386         batch_size=args.batch_size,
 387         logger=log_string,
 388         device=device,
 389     )
 390
 391
 392 elif args.task == "guessop":
 393     task = tasks.SandBox(
 394         problem=problems.ProblemGuessOperator(),
 395         nb_train_samples=args.nb_train_samples,
 396         nb_test_samples=args.nb_test_samples,
 397         batch_size=args.batch_size,
 398         logger=log_string,
 399         device=device,
 400     )
 401
 402
 403 elif args.task == "twotargets":
 404     task = tasks.SandBox(
 405         problem=problems.ProblemTwoTargets(),
 406         nb_train_samples=args.nb_train_samples,
 407         nb_test_samples=args.nb_test_samples,
 408         batch_size=args.batch_size,
 409         logger=log_string,
 410         device=device,
 411     )
 412
 413 elif args.task == "addition":
 414     task = tasks.SandBox(
 415         problem=problems.ProblemAddition(),
 416         nb_train_samples=args.nb_train_samples,
 417         nb_test_samples=args.nb_test_samples,
 418         batch_size=args.batch_size,
 419         logger=log_string,
 420         device=device,
 421     )
 422
 423 elif args.task == "picoclvr":
 424     task = tasks.PicoCLVR(
 425         nb_train_samples=args.nb_train_samples,
 426         nb_test_samples=args.nb_test_samples,
 427         batch_size=args.batch_size,
 428         height=args.picoclvr_height,
 429         width=args.picoclvr_width,
 430         nb_colors=args.picoclvr_nb_colors,
 431         logger=log_string,
 432         device=device,
 433         pruner_train=picoclvr_pruner_train,
 434         pruner_eval=picoclvr_pruner_eval,
 435     )
 436
 437 elif args.task == "mnist":
 438     task = tasks.MNIST(
 439         nb_train_samples=args.nb_train_samples,
 440         nb_test_samples=args.nb_test_samples,
 441         batch_size=args.batch_size,
 442         device=device,
 443     )
 444
 445 elif args.task == "maze":
 446     task = tasks.Maze(
 447         nb_train_samples=args.nb_train_samples,
 448         nb_test_samples=args.nb_test_samples,
 449         batch_size=args.batch_size,
 450         height=args.maze_height,
 451         width=args.maze_width,
 452         nb_walls=args.maze_nb_walls,
 453         device=device,
 454     )
 455
 456 elif args.task == "snake":
 457     task = tasks.Snake(
 458         nb_train_samples=args.nb_train_samples,
 459         nb_test_samples=args.nb_test_samples,
 460         batch_size=args.batch_size,
 461         height=args.snake_height,
 462         width=args.snake_width,
 463         nb_colors=args.snake_nb_colors,
 464         length=args.snake_length,
 465         prompt_length=args.snake_length // 2,
 466         device=device,
 467     )
 468
 469 elif args.task == "stack":
 470     task = tasks.Stack(
 471         nb_train_samples=args.nb_train_samples,
 472         nb_test_samples=args.nb_test_samples,
 473         batch_size=args.batch_size,
 474         logger=log_string,
 475         nb_steps=args.stack_nb_steps,
 476         nb_stacks=args.stack_nb_stacks,
 477         nb_digits=args.stack_nb_digits,
 478         fraction_values_for_train=args.stack_fraction_values_for_train,
 479         device=device,
 480     )
 481
 482 elif args.task == "expr":
 483     task = tasks.Expr(
 484         nb_train_samples=args.nb_train_samples,
 485         nb_test_samples=args.nb_test_samples,
 486         nb_variables=args.expr_nb_variables,
 487         sequence_length=args.expr_sequence_length,
 488         operand_max=args.expr_operand_max,
 489         result_max=args.expr_result_max,
 490         batch_size=args.batch_size,
 491         device=device,
 492     )
 493
 494 elif args.task == "rpl":
 495     task = tasks.RPL(
 496         nb_train_samples=args.nb_train_samples,
 497         nb_test_samples=args.nb_test_samples,
 498         batch_size=args.batch_size,
 499         nb_starting_values=args.rpl_nb_starting_values,
 500         max_input=args.rpl_max_input,
 501         prog_len=args.rpl_prog_len,
 502         nb_runs=args.rpl_nb_runs,
 503         no_prog=args.rpl_no_prog,
 504         logger=log_string,
 505         device=device,
 506     )
 507
 508 elif args.task == "world":
 509     task = tasks.World(
 510         nb_train_samples=args.nb_train_samples,
 511         nb_test_samples=args.nb_test_samples,
 512         batch_size=args.batch_size,
 513         vqae_nb_epochs=args.world_vqae_nb_epochs,
 514         logger=log_string,
 515         device=device,
 516     )
 517
 518 else:
 519     raise ValueError(f"Unknown task {args.task}")
 520
 521 ######################################################################
 522
 523 log_string(f"device {device}")
 524
 525 vocabulary_size = task.vocabulary_size()
 526
 527 log_string(f"vocabulary_size {vocabulary_size}")
 528
 529 ##############################
 530
 531 model = mygpt.MyGPT(
 532     vocabulary_size=vocabulary_size,
 533     dim_model=args.dim_model,
 534     dim_keys=args.dim_keys,
 535     dim_hidden=args.dim_hidden,
 536     nb_heads=args.nb_heads,
 537     nb_blocks=args.nb_blocks,
 538     causal=True,
 539     dropout=args.dropout,
 540 )
 541
 542 model.to(device)
 543
 544 nb_parameters = sum(p.numel() for p in model.parameters())
 545 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 546
 547 ######################################################################
 548
 549 nb_epochs_finished = 0
 550
 551 if args.no_checkpoint:
 552     log_string(f"not trying to load checkpoint.")
 553
 554 else:
 555     try:
 556         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 557         checkpoint = torch.load(checkpoint_name)
 558         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 559         model.load_state_dict(checkpoint["model_state"])
 560         torch.set_rng_state(checkpoint["rng_state"])
 561         if torch.cuda.is_available():
 562             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 563
 564         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 565
 566     except FileNotFoundError:
 567         log_string("starting from scratch.")
 568
 569     except:
 570         log_string("error when loading the checkpoint.")
 571         exit(1)
 572
 573 ######################################################################
 574
 575 if args.task == "expr" and args.expr_input_file is not None:
 576     task.produce_results(
 577         n_epoch=nb_epochs_finished,
 578         model=model,
 579         result_dir=args.result_dir,
 580         logger=log_string,
 581         deterministic_synthesis=args.deterministic_synthesis,
 582         input_file=args.expr_input_file,
 583     )
 584
 585     exit(0)
 586
 587 ######################################################################
 588
 589 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 590
 591 # Compute the entropy of the training tokens
 592
 593 token_count = 0
 594 for input in task.batches(split="train"):
 595     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 596 token_probas = token_count / token_count.sum()
 597 entropy = -torch.xlogy(token_probas, token_probas).sum()
 598 train_set_perplexity = math.exp(entropy)
 599
 600 ######################################################################
 601 # A bit of paranoia never hurts
 602
 603 if args.max_percents_of_test_in_train >= 0:
 604
 605     def subsets_as_tuples(batches, cs):
 606         s = set()
 607         for batch in batches:
 608             for x in batch:
 609                 s.add(tuple([v.item() for v in x]))
 610                 if len(s) == cs:
 611                     yield s
 612                     s = set()
 613         yield s
 614
 615     nb_test, nb_in_train = 0, 0
 616     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 617         in_train = set()
 618         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 619             in_train.update(test_subset.intersection(train_subset))
 620         nb_in_train += len(in_train)
 621         nb_test += len(test_subset)
 622
 623     log_string(
 624         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 625     )
 626
 627     assert (
 628         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 629     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 630
 631 ##############################
 632
 633 if args.learning_rate_schedule == "cos":
 634     learning_rate_schedule = {}
 635     for n_epoch in range(args.nb_epochs):
 636         u = n_epoch / args.nb_epochs * math.pi
 637         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 638 else:
 639     u = {
 640         int(k): float(v)
 641         for k, v in [
 642             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 643         ]
 644     }
 645
 646     learning_rate_schedule = {}
 647     learning_rate = args.learning_rate
 648     for n_epoch in range(args.nb_epochs):
 649         if n_epoch in u:
 650             learning_rate = u[n_epoch]
 651         learning_rate_schedule[n_epoch] = learning_rate
 652
 653 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 654
 655 ##############################
 656
 657 nb_samples_seen = 0
 658
 659 if nb_epochs_finished >= nb_epochs:
 660     task.produce_results(
 661         n_epoch=nb_epochs_finished,
 662         model=model,
 663         result_dir=args.result_dir,
 664         logger=log_string,
 665         deterministic_synthesis=args.deterministic_synthesis,
 666     )
 667
 668 for n_epoch in range(nb_epochs_finished, nb_epochs):
 669     learning_rate = learning_rate_schedule[n_epoch]
 670
 671     log_string(f"learning_rate {learning_rate}")
 672
 673     if args.optim == "sgd":
 674         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 675     elif args.optim == "adam":
 676         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 677     elif args.optim == "adamw":
 678         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 679     else:
 680         raise ValueError(f"Unknown optimizer {args.optim}.")
 681
 682     model.train()
 683
 684     nb_train_samples, acc_train_loss = 0, 0.0
 685
 686     for input in task.batches(split="train"):
 687         input = input.to(device)
 688         output = model(mygpt.BracketedSequence(input)).x
 689         loss = F.cross_entropy(output.transpose(1, 2), input)
 690         acc_train_loss += loss.item() * input.size(0)
 691         nb_train_samples += input.size(0)
 692         nb_samples_seen += input.size(0)
 693
 694         optimizer.zero_grad()
 695         loss.backward()
 696         optimizer.step()
 697
 698     with torch.autograd.no_grad():
 699         model.eval()
 700
 701         nb_test_samples, acc_test_loss = 0, 0.0
 702
 703         for input in task.batches(split="test"):
 704             input = input.to(device)
 705
 706             output = model(mygpt.BracketedSequence(input)).x
 707             loss = F.cross_entropy(output.transpose(1, 2), input)
 708             acc_test_loss += loss.item() * input.size(0)
 709             nb_test_samples += input.size(0)
 710
 711         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 712         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 713
 714         log_string(
 715             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 716         )
 717
 718         task.produce_results(
 719             n_epoch=n_epoch,
 720             model=model,
 721             result_dir=args.result_dir,
 722             logger=log_string,
 723             deterministic_synthesis=args.deterministic_synthesis,
 724         )
 725
 726     checkpoint = {
 727         "nb_epochs_finished": n_epoch + 1,
 728         "model_state": model.state_dict(),
 729         "rng_state": torch.get_rng_state(),
 730     }
 731
 732     if torch.cuda.is_available():
 733         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 734
 735     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 736     torch.save(checkpoint, checkpoint_name)
 737     log_string(f"saved checkpoint {checkpoint_name}")
 738
 739 ######################################################################