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="sandbox",
  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=6)
 126
 127 parser.add_argument("--snake_width", type=int, default=8)
 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     "sandbox": {
 175         "nb_epochs": 50,
 176         "batch_size": 25,
 177         "nb_train_samples": 100000,
 178         "nb_test_samples": 10000,
 179     },
 180     "picoclvr": {
 181         "nb_epochs": 25,
 182         "batch_size": 25,
 183         "nb_train_samples": 250000,
 184         "nb_test_samples": 10000,
 185     },
 186     "mnist": {
 187         "nb_epochs": 25,
 188         "batch_size": 10,
 189         "nb_train_samples": 250000,
 190         "nb_test_samples": 10000,
 191     },
 192     "maze": {
 193         "nb_epochs": 25,
 194         "batch_size": 5,
 195         "nb_train_samples": 250000,
 196         "nb_test_samples": 10000,
 197     },
 198     "snake": {
 199         "nb_epochs": 5,
 200         "batch_size": 25,
 201         "nb_train_samples": 250000,
 202         "nb_test_samples": 10000,
 203     },
 204     "stack": {
 205         "nb_epochs": 5,
 206         "batch_size": 25,
 207         "nb_train_samples": 100000,
 208         "nb_test_samples": 1000,
 209     },
 210     "expr": {
 211         "nb_epochs": 40,
 212         "batch_size": 25,
 213         "nb_train_samples": 1000000,
 214         "nb_test_samples": 10000,
 215     },
 216     "rpl": {
 217         "nb_epochs": 40,
 218         "batch_size": 25,
 219         "nb_train_samples": 100000,
 220         "nb_test_samples": 10000,
 221     },
 222     "world": {
 223         "nb_epochs": 10,
 224         "batch_size": 25,
 225         "nb_train_samples": 25000,
 226         "nb_test_samples": 1000,
 227     },
 228 }
 229
 230 if args.task in default_task_args:
 231     for k, v in default_task_args[args.task].items():
 232         if getattr(args, k) is None:
 233             setattr(args, k, v)
 234
 235 ######################################################################
 236
 237 default_model_args = {
 238     "17K": {
 239         "dim_model": 32,
 240         "dim_keys": 32,
 241         "dim_hidden": 32,
 242         "nb_heads": 2,
 243         "nb_blocks": 2,
 244     },
 245     "37M": {
 246         "dim_model": 512,
 247         "dim_keys": 64,
 248         "dim_hidden": 2048,
 249         "nb_heads": 8,
 250         "nb_blocks": 12,
 251     },
 252     "122M": {
 253         "dim_model": 768,
 254         "dim_keys": 64,
 255         "dim_hidden": 2048,
 256         "nb_heads": 8,
 257         "nb_blocks": 24,
 258     },
 259     "352M": {
 260         "dim_model": 1024,
 261         "dim_keys": 64,
 262         "dim_hidden": 2048,
 263         "nb_heads": 8,
 264         "nb_blocks": 48,
 265     },
 266 }
 267
 268 if args.model in default_model_args:
 269     for k, v in default_model_args[args.model].items():
 270         if getattr(args, k) is None:
 271             setattr(args, k, v)
 272 else:
 273     raise ValueError(f"Unknown model {args.model}")
 274
 275 ######################################################################
 276
 277 try:
 278     os.mkdir(args.result_dir)
 279 except FileExistsError:
 280     if not args.overwrite_results:
 281         print(f"result directory {args.result_dir} already exists")
 282         exit(1)
 283
 284 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 285
 286 if args.seed >= 0:
 287     # torch.backends.cudnn.deterministic = True
 288     # torch.backends.cudnn.benchmark = False
 289     # torch.use_deterministic_algorithms(True)
 290     torch.manual_seed(args.seed)
 291     if torch.cuda.is_available():
 292         torch.cuda.manual_seed_all(args.seed)
 293
 294 ######################################################################
 295
 296
 297 def log_string(s):
 298     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 299
 300     if log_file is not None:
 301         log_file.write(t + s + "\n")
 302         log_file.flush()
 303
 304     print(t + s)
 305     sys.stdout.flush()
 306
 307
 308 for n in vars(args):
 309     log_string(f"args.{n} {getattr(args, n)}")
 310
 311
 312 ######################################################################
 313
 314
 315 def picoclvr_pruner_horizontal_green(p):
 316     return not ("green" in p and ("left" in p or "right" in p))
 317
 318
 319 picoclvr_pruner_train = (
 320     picoclvr_pruner_horizontal_green
 321     if args.picocvlr_prune_properties in {"train+eval"}
 322     else None
 323 )
 324
 325 picoclvr_pruner_eval = (
 326     (lambda p: not picoclvr_pruner_horizontal_green(p))
 327     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 328     else None
 329 )
 330
 331 ######################################################################
 332
 333 if args.task == "byheart":
 334     task = tasks.SandBox(
 335         problem=problems.ProblemByHeart(),
 336         nb_train_samples=args.nb_train_samples,
 337         nb_test_samples=args.nb_test_samples,
 338         batch_size=args.batch_size,
 339         logger=log_string,
 340         device=device,
 341     )
 342
 343
 344 elif args.task == "learnop":
 345     task = tasks.SandBox(
 346         problem=problems.ProblemLearnOperator(),
 347         nb_train_samples=args.nb_train_samples,
 348         nb_test_samples=args.nb_test_samples,
 349         batch_size=args.batch_size,
 350         logger=log_string,
 351         device=device,
 352     )
 353
 354
 355 elif args.task == "guessop":
 356     task = tasks.SandBox(
 357         problem=problems.ProblemGuessOperator(),
 358         nb_train_samples=args.nb_train_samples,
 359         nb_test_samples=args.nb_test_samples,
 360         batch_size=args.batch_size,
 361         logger=log_string,
 362         device=device,
 363     )
 364
 365
 366 elif args.task == "twotargets":
 367     task = tasks.SandBox(
 368         problem=problems.ProblemTwoTargets(),
 369         nb_train_samples=args.nb_train_samples,
 370         nb_test_samples=args.nb_test_samples,
 371         batch_size=args.batch_size,
 372         logger=log_string,
 373         device=device,
 374     )
 375
 376 elif args.task == "addition":
 377     task = tasks.SandBox(
 378         problem=problems.ProblemAddition(),
 379         nb_train_samples=args.nb_train_samples,
 380         nb_test_samples=args.nb_test_samples,
 381         batch_size=args.batch_size,
 382         logger=log_string,
 383         device=device,
 384     )
 385
 386 elif args.task == "picoclvr":
 387     task = tasks.PicoCLVR(
 388         nb_train_samples=args.nb_train_samples,
 389         nb_test_samples=args.nb_test_samples,
 390         batch_size=args.batch_size,
 391         height=args.picoclvr_height,
 392         width=args.picoclvr_width,
 393         nb_colors=args.picoclvr_nb_colors,
 394         logger=log_string,
 395         device=device,
 396         pruner_train=picoclvr_pruner_train,
 397         pruner_eval=picoclvr_pruner_eval,
 398     )
 399
 400 elif args.task == "mnist":
 401     task = tasks.MNIST(
 402         nb_train_samples=args.nb_train_samples,
 403         nb_test_samples=args.nb_test_samples,
 404         batch_size=args.batch_size,
 405         device=device,
 406     )
 407
 408 elif args.task == "maze":
 409     task = tasks.Maze(
 410         nb_train_samples=args.nb_train_samples,
 411         nb_test_samples=args.nb_test_samples,
 412         batch_size=args.batch_size,
 413         height=args.maze_height,
 414         width=args.maze_width,
 415         nb_walls=args.maze_nb_walls,
 416         device=device,
 417     )
 418
 419 elif args.task == "snake":
 420     task = tasks.Snake(
 421         nb_train_samples=args.nb_train_samples,
 422         nb_test_samples=args.nb_test_samples,
 423         batch_size=args.batch_size,
 424         height=args.snake_height,
 425         width=args.snake_width,
 426         nb_colors=args.snake_nb_colors,
 427         length=args.snake_length,
 428         prompt_length=args.snake_length // 2,
 429         device=device,
 430     )
 431
 432 elif args.task == "stack":
 433     task = tasks.Stack(
 434         nb_train_samples=args.nb_train_samples,
 435         nb_test_samples=args.nb_test_samples,
 436         batch_size=args.batch_size,
 437         logger=log_string,
 438         nb_steps=args.stack_nb_steps,
 439         nb_stacks=args.stack_nb_stacks,
 440         nb_digits=args.stack_nb_digits,
 441         fraction_values_for_train=args.stack_fraction_values_for_train,
 442         device=device,
 443     )
 444
 445 elif args.task == "expr":
 446     task = tasks.Expr(
 447         nb_train_samples=args.nb_train_samples,
 448         nb_test_samples=args.nb_test_samples,
 449         nb_variables=args.expr_nb_variables,
 450         sequence_length=args.expr_sequence_length,
 451         operand_max=args.expr_operand_max,
 452         result_max=args.expr_result_max,
 453         batch_size=args.batch_size,
 454         device=device,
 455     )
 456
 457 elif args.task == "rpl":
 458     task = tasks.RPL(
 459         nb_train_samples=args.nb_train_samples,
 460         nb_test_samples=args.nb_test_samples,
 461         batch_size=args.batch_size,
 462         nb_starting_values=args.rpl_nb_starting_values,
 463         max_input=args.rpl_max_input,
 464         prog_len=args.rpl_prog_len,
 465         nb_runs=args.rpl_nb_runs,
 466         no_prog=args.rpl_no_prog,
 467         logger=log_string,
 468         device=device,
 469     )
 470
 471 elif args.task == "world":
 472     task = tasks.World(
 473         nb_train_samples=args.nb_train_samples,
 474         nb_test_samples=args.nb_test_samples,
 475         batch_size=args.batch_size,
 476         vqae_nb_epochs=args.world_vqae_nb_epochs,
 477         logger=log_string,
 478         device=device,
 479     )
 480
 481 else:
 482     raise ValueError(f"Unknown task {args.task}")
 483
 484 ######################################################################
 485
 486 log_string(f"device {device}")
 487
 488 vocabulary_size = task.vocabulary_size()
 489
 490 log_string(f"vocabulary_size {vocabulary_size}")
 491
 492 ##############################
 493
 494 model = mygpt.MyGPT(
 495     vocabulary_size=vocabulary_size,
 496     dim_model=args.dim_model,
 497     dim_keys=args.dim_keys,
 498     dim_hidden=args.dim_hidden,
 499     nb_heads=args.nb_heads,
 500     nb_blocks=args.nb_blocks,
 501     causal=True,
 502     dropout=args.dropout,
 503 )
 504
 505 model.to(device)
 506
 507 nb_parameters = sum(p.numel() for p in model.parameters())
 508 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 509
 510 ######################################################################
 511
 512 nb_epochs_finished = 0
 513
 514 if args.no_checkpoint:
 515     log_string(f"not trying to load checkpoint.")
 516
 517 else:
 518     try:
 519         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 520         checkpoint = torch.load(checkpoint_name)
 521         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 522         model.load_state_dict(checkpoint["model_state"])
 523         torch.set_rng_state(checkpoint["rng_state"])
 524         if torch.cuda.is_available():
 525             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 526
 527         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 528
 529     except FileNotFoundError:
 530         log_string("starting from scratch.")
 531
 532     except:
 533         log_string("error when loading the checkpoint.")
 534         exit(1)
 535
 536 ######################################################################
 537
 538 if args.task == "expr" and args.expr_input_file is not None:
 539     task.produce_results(
 540         n_epoch=nb_epochs_finished,
 541         model=model,
 542         result_dir=args.result_dir,
 543         logger=log_string,
 544         deterministic_synthesis=args.deterministic_synthesis,
 545         input_file=args.expr_input_file,
 546     )
 547
 548     exit(0)
 549
 550 ######################################################################
 551
 552 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 553
 554 # Compute the entropy of the training tokens
 555
 556 token_count = 0
 557 for input in task.batches(split="train"):
 558     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 559 token_probas = token_count / token_count.sum()
 560 entropy = -torch.xlogy(token_probas, token_probas).sum()
 561 train_set_perplexity = math.exp(entropy)
 562
 563 ######################################################################
 564 # A bit of paranoia never hurts
 565
 566
 567 def subsets_as_tuples(batches, cs):
 568     s = set()
 569     for batch in batches:
 570         for x in batch:
 571             s.add(tuple([v.item() for v in x]))
 572             if len(s) == cs:
 573                 yield s
 574                 s = set()
 575     yield s
 576
 577
 578 nb_test, nb_in_train = 0, 0
 579 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 580     in_train = set()
 581     for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 582         in_train.update(test_subset.intersection(train_subset))
 583     nb_in_train += len(in_train)
 584     nb_test += len(test_subset)
 585
 586 log_string(
 587     f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 588 )
 589
 590 assert (
 591     nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 592 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 593
 594 ##############################
 595
 596 if args.learning_rate_schedule == "cos":
 597     learning_rate_schedule = {}
 598     for n_epoch in range(args.nb_epochs):
 599         u = n_epoch / args.nb_epochs * math.pi
 600         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 601 else:
 602     u = {
 603         int(k): float(v)
 604         for k, v in [
 605             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 606         ]
 607     }
 608
 609     learning_rate_schedule = {}
 610     learning_rate = args.learning_rate
 611     for n_epoch in range(args.nb_epochs):
 612         if n_epoch in u:
 613             learning_rate = u[n_epoch]
 614         learning_rate_schedule[n_epoch] = learning_rate
 615
 616 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 617
 618 ##############################
 619
 620 nb_samples_seen = 0
 621
 622 if nb_epochs_finished >= nb_epochs:
 623     task.produce_results(
 624         n_epoch=nb_epochs_finished,
 625         model=model,
 626         result_dir=args.result_dir,
 627         logger=log_string,
 628         deterministic_synthesis=args.deterministic_synthesis,
 629     )
 630
 631 for n_epoch in range(nb_epochs_finished, nb_epochs):
 632     learning_rate = learning_rate_schedule[n_epoch]
 633
 634     log_string(f"learning_rate {learning_rate}")
 635
 636     if args.optim == "sgd":
 637         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 638     elif args.optim == "adam":
 639         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 640     elif args.optim == "adamw":
 641         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 642     else:
 643         raise ValueError(f"Unknown optimizer {args.optim}.")
 644
 645     model.train()
 646
 647     nb_train_samples, acc_train_loss = 0, 0.0
 648
 649     for input in task.batches(split="train"):
 650         input = input.to(device)
 651         output = model(mygpt.BracketedSequence(input)).x
 652         loss = F.cross_entropy(output.transpose(1, 2), input)
 653         acc_train_loss += loss.item() * input.size(0)
 654         nb_train_samples += input.size(0)
 655         nb_samples_seen += input.size(0)
 656
 657         optimizer.zero_grad()
 658         loss.backward()
 659         optimizer.step()
 660
 661     with torch.autograd.no_grad():
 662         model.eval()
 663
 664         nb_test_samples, acc_test_loss = 0, 0.0
 665
 666         for input in task.batches(split="test"):
 667             input = input.to(device)
 668
 669             output = model(mygpt.BracketedSequence(input)).x
 670             loss = F.cross_entropy(output.transpose(1, 2), input)
 671             acc_test_loss += loss.item() * input.size(0)
 672             nb_test_samples += input.size(0)
 673
 674         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 675         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 676
 677         log_string(
 678             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 679         )
 680
 681         task.produce_results(
 682             n_epoch=n_epoch,
 683             model=model,
 684             result_dir=args.result_dir,
 685             logger=log_string,
 686             deterministic_synthesis=args.deterministic_synthesis,
 687         )
 688
 689     checkpoint = {
 690         "nb_epochs_finished": n_epoch + 1,
 691         "model_state": model.state_dict(),
 692         "rng_state": torch.get_rng_state(),
 693     }
 694
 695     if torch.cuda.is_available():
 696         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 697
 698     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 699     torch.save(checkpoint, checkpoint_name)
 700     log_string(f"saved checkpoint {checkpoint_name}")
 701
 702 ######################################################################