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, grid, qmlp",
  37 )
  38
  39 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  40
  41 parser.add_argument("--result_dir", type=str, default=None)
  42
  43 parser.add_argument("--seed", type=int, default=0)
  44
  45 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  46
  47 ########################################
  48
  49 parser.add_argument("--nb_epochs", type=int, default=25)
  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 # grid options
 104
 105 parser.add_argument("--grid_size", type=int, default=6)
 106
 107 ##############################
 108 # picoclvr options
 109
 110 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 111
 112 parser.add_argument("--picoclvr_height", type=int, default=12)
 113
 114 parser.add_argument("--picoclvr_width", type=int, default=16)
 115
 116 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 117
 118 ##############################
 119 # Maze options
 120
 121 parser.add_argument("--maze_height", type=int, default=13)
 122
 123 parser.add_argument("--maze_width", type=int, default=21)
 124
 125 parser.add_argument("--maze_nb_walls", type=int, default=15)
 126
 127 ##############################
 128 # Snake options
 129
 130 parser.add_argument("--snake_height", type=int, default=9)
 131
 132 parser.add_argument("--snake_width", type=int, default=12)
 133
 134 parser.add_argument("--snake_nb_colors", type=int, default=5)
 135
 136 parser.add_argument("--snake_length", type=int, default=200)
 137
 138 ##############################
 139 # Stack options
 140
 141 parser.add_argument("--stack_nb_steps", type=int, default=100)
 142
 143 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 144
 145 parser.add_argument("--stack_nb_digits", type=int, default=3)
 146
 147 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 148
 149 ##############################
 150 # Expr options
 151
 152 parser.add_argument("--expr_nb_variables", type=int, default=5)
 153
 154 parser.add_argument("--expr_sequence_length", type=int, default=40)
 155
 156 parser.add_argument("--expr_operand_max", type=int, default=9)
 157
 158 parser.add_argument("--expr_result_max", type=int, default=99)
 159
 160 parser.add_argument("--expr_input_file", type=str, default=None)
 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     "addition": {
 175         "model": "352M",
 176         "batch_size": 25,
 177         "nb_train_samples": 250000,
 178         "nb_test_samples": 10000,
 179     },
 180     "byheart": {
 181         "model": "37M",
 182         "batch_size": 25,
 183         "nb_train_samples": 50000,
 184         "nb_test_samples": 10000,
 185     },
 186     "expr": {
 187         "model": "352M",
 188         "batch_size": 25,
 189         "nb_train_samples": 2500000,
 190         "nb_test_samples": 10000,
 191     },
 192     "grid": {
 193         "model": "37M",
 194         "batch_size": 25,
 195         "nb_train_samples": 250000,
 196         "nb_test_samples": 10000,
 197     },
 198     "qmlp": {
 199         "model": "37M",
 200         "batch_size": 10,
 201         "nb_train_samples": 100000,
 202         "nb_test_samples": 1000,
 203     },
 204     "guessop": {
 205         "model": "352M",
 206         "batch_size": 25,
 207         "nb_train_samples": 1000000,
 208         "nb_test_samples": 10000,
 209     },
 210     "learnop": {
 211         "model": "37M",
 212         "batch_size": 25,
 213         "nb_train_samples": 50000,
 214         "nb_test_samples": 10000,
 215     },
 216     "maze": {
 217         "model": "37M",
 218         "batch_size": 5,
 219         "nb_train_samples": 100000,
 220         "nb_test_samples": 10000,
 221     },
 222     "picoclvr": {
 223         "model": "37M",
 224         "batch_size": 25,
 225         "nb_train_samples": 250000,
 226         "nb_test_samples": 10000,
 227     },
 228     "rpl": {
 229         "model": "352M",
 230         "batch_size": 5,
 231         "nb_train_samples": 2500000,
 232         "nb_test_samples": 10000,
 233     },
 234     "snake": {
 235         "model": "37M",
 236         "batch_size": 25,
 237         "nb_train_samples": 250000,
 238         "nb_test_samples": 10000,
 239     },
 240     "stack": {
 241         "model": "37M",
 242         "batch_size": 25,
 243         "nb_train_samples": 100000,
 244         "nb_test_samples": 1000,
 245     },
 246     "twotargets": {
 247         "model": "37M",
 248         "batch_size": 25,
 249         "nb_train_samples": 50000,
 250         "nb_test_samples": 10000,
 251     },
 252     "mnist": {
 253         "model": "37M",
 254         "batch_size": 10,
 255         "nb_train_samples": 60000,
 256         "nb_test_samples": 10000,
 257     },
 258 }
 259
 260 if args.task in default_task_args:
 261     for k, v in default_task_args[args.task].items():
 262         if getattr(args, k) is None:
 263             setattr(args, k, v)
 264
 265 ######################################################################
 266
 267 default_model_args = {
 268     "17K": {
 269         "dim_model": 32,
 270         "dim_keys": 32,
 271         "dim_hidden": 32,
 272         "nb_heads": 2,
 273         "nb_blocks": 2,
 274     },
 275     "37M": {
 276         "dim_model": 512,
 277         "dim_keys": 64,
 278         "dim_hidden": 2048,
 279         "nb_heads": 8,
 280         "nb_blocks": 12,
 281     },
 282     "122M": {
 283         "dim_model": 768,
 284         "dim_keys": 64,
 285         "dim_hidden": 2048,
 286         "nb_heads": 8,
 287         "nb_blocks": 24,
 288     },
 289     "352M": {
 290         "dim_model": 1024,
 291         "dim_keys": 64,
 292         "dim_hidden": 2048,
 293         "nb_heads": 8,
 294         "nb_blocks": 48,
 295     },
 296 }
 297
 298 if args.model in default_model_args:
 299     for k, v in default_model_args[args.model].items():
 300         if getattr(args, k) is None:
 301             setattr(args, k, v)
 302 else:
 303     raise ValueError(f"Unknown model {args.model}")
 304
 305 ######################################################################
 306
 307 try:
 308     os.mkdir(args.result_dir)
 309 except FileExistsError:
 310     if not args.overwrite_results:
 311         print(f"result directory {args.result_dir} already exists")
 312         exit(1)
 313
 314 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 315
 316 if args.seed >= 0:
 317     # torch.backends.cudnn.deterministic = True
 318     # torch.backends.cudnn.benchmark = False
 319     # torch.use_deterministic_algorithms(True)
 320     torch.manual_seed(args.seed)
 321     if torch.cuda.is_available():
 322         torch.cuda.manual_seed_all(args.seed)
 323
 324 ######################################################################
 325
 326
 327 def log_string(s):
 328     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 329
 330     if log_file is not None:
 331         log_file.write(t + s + "\n")
 332         log_file.flush()
 333
 334     print(t + s)
 335     sys.stdout.flush()
 336
 337
 338 for n in vars(args):
 339     log_string(f"args.{n} {getattr(args, n)}")
 340
 341
 342 ######################################################################
 343
 344
 345 def picoclvr_pruner_horizontal_green(p):
 346     return not ("green" in p and ("left" in p or "right" in p))
 347
 348
 349 picoclvr_pruner_train = (
 350     picoclvr_pruner_horizontal_green
 351     if args.picocvlr_prune_properties in {"train+eval"}
 352     else None
 353 )
 354
 355 picoclvr_pruner_eval = (
 356     (lambda p: not picoclvr_pruner_horizontal_green(p))
 357     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 358     else None
 359 )
 360
 361 ######################################################################
 362
 363 if args.task == "byheart":
 364     task = tasks.SandBox(
 365         problem=problems.ProblemByHeart(),
 366         nb_train_samples=args.nb_train_samples,
 367         nb_test_samples=args.nb_test_samples,
 368         batch_size=args.batch_size,
 369         logger=log_string,
 370         device=device,
 371     )
 372     args.max_percents_of_test_in_train = -1
 373
 374 elif args.task == "learnop":
 375     task = tasks.SandBox(
 376         problem=problems.ProblemLearnOperator(),
 377         nb_train_samples=args.nb_train_samples,
 378         nb_test_samples=args.nb_test_samples,
 379         batch_size=args.batch_size,
 380         logger=log_string,
 381         device=device,
 382     )
 383
 384
 385 elif args.task == "guessop":
 386     task = tasks.SandBox(
 387         problem=problems.ProblemGuessOperator(),
 388         nb_train_samples=args.nb_train_samples,
 389         nb_test_samples=args.nb_test_samples,
 390         batch_size=args.batch_size,
 391         logger=log_string,
 392         device=device,
 393     )
 394
 395
 396 elif args.task == "twotargets":
 397     task = tasks.SandBox(
 398         problem=problems.ProblemTwoTargets(),
 399         nb_train_samples=args.nb_train_samples,
 400         nb_test_samples=args.nb_test_samples,
 401         batch_size=args.batch_size,
 402         logger=log_string,
 403         device=device,
 404     )
 405
 406 elif args.task == "addition":
 407     task = tasks.SandBox(
 408         problem=problems.ProblemAddition(),
 409         nb_train_samples=args.nb_train_samples,
 410         nb_test_samples=args.nb_test_samples,
 411         batch_size=args.batch_size,
 412         logger=log_string,
 413         device=device,
 414     )
 415
 416 elif args.task == "picoclvr":
 417     task = tasks.PicoCLVR(
 418         nb_train_samples=args.nb_train_samples,
 419         nb_test_samples=args.nb_test_samples,
 420         batch_size=args.batch_size,
 421         height=args.picoclvr_height,
 422         width=args.picoclvr_width,
 423         nb_colors=args.picoclvr_nb_colors,
 424         logger=log_string,
 425         device=device,
 426         pruner_train=picoclvr_pruner_train,
 427         pruner_eval=picoclvr_pruner_eval,
 428     )
 429
 430 elif args.task == "mnist":
 431     task = tasks.MNIST(
 432         nb_train_samples=args.nb_train_samples,
 433         nb_test_samples=args.nb_test_samples,
 434         batch_size=args.batch_size,
 435         device=device,
 436     )
 437
 438 elif args.task == "maze":
 439     task = tasks.Maze(
 440         nb_train_samples=args.nb_train_samples,
 441         nb_test_samples=args.nb_test_samples,
 442         batch_size=args.batch_size,
 443         height=args.maze_height,
 444         width=args.maze_width,
 445         nb_walls=args.maze_nb_walls,
 446         device=device,
 447     )
 448
 449 elif args.task == "snake":
 450     task = tasks.Snake(
 451         nb_train_samples=args.nb_train_samples,
 452         nb_test_samples=args.nb_test_samples,
 453         batch_size=args.batch_size,
 454         height=args.snake_height,
 455         width=args.snake_width,
 456         nb_colors=args.snake_nb_colors,
 457         length=args.snake_length,
 458         prompt_length=args.snake_length // 2,
 459         device=device,
 460     )
 461
 462 elif args.task == "stack":
 463     task = tasks.Stack(
 464         nb_train_samples=args.nb_train_samples,
 465         nb_test_samples=args.nb_test_samples,
 466         batch_size=args.batch_size,
 467         logger=log_string,
 468         nb_steps=args.stack_nb_steps,
 469         nb_stacks=args.stack_nb_stacks,
 470         nb_digits=args.stack_nb_digits,
 471         fraction_values_for_train=args.stack_fraction_values_for_train,
 472         device=device,
 473     )
 474
 475 elif args.task == "expr":
 476     task = tasks.Expr(
 477         nb_train_samples=args.nb_train_samples,
 478         nb_test_samples=args.nb_test_samples,
 479         nb_variables=args.expr_nb_variables,
 480         sequence_length=args.expr_sequence_length,
 481         operand_max=args.expr_operand_max,
 482         result_max=args.expr_result_max,
 483         batch_size=args.batch_size,
 484         device=device,
 485     )
 486
 487 elif args.task == "rpl":
 488     task = tasks.RPL(
 489         nb_train_samples=args.nb_train_samples,
 490         nb_test_samples=args.nb_test_samples,
 491         batch_size=args.batch_size,
 492         nb_starting_values=args.rpl_nb_starting_values,
 493         max_input=args.rpl_max_input,
 494         prog_len=args.rpl_prog_len,
 495         nb_runs=args.rpl_nb_runs,
 496         no_prog=args.rpl_no_prog,
 497         logger=log_string,
 498         device=device,
 499     )
 500
 501 elif args.task == "grid":
 502     task = tasks.Grid(
 503         nb_train_samples=args.nb_train_samples,
 504         nb_test_samples=args.nb_test_samples,
 505         batch_size=args.batch_size,
 506         size=args.grid_size,
 507         logger=log_string,
 508         device=device,
 509     )
 510
 511 elif args.task == "qmlp":
 512     task = tasks.QMLP(
 513         nb_train_samples=args.nb_train_samples,
 514         nb_test_samples=args.nb_test_samples,
 515         batch_size=args.batch_size,
 516         result_dir=args.result_dir,
 517         logger=log_string,
 518         device=device,
 519     )
 520
 521 else:
 522     raise ValueError(f"Unknown task {args.task}")
 523
 524 ######################################################################
 525
 526 log_string(f"device {device}")
 527
 528 vocabulary_size = task.vocabulary_size()
 529
 530 log_string(f"vocabulary_size {vocabulary_size}")
 531
 532 ##############################
 533
 534 model = mygpt.MyGPT(
 535     vocabulary_size=vocabulary_size,
 536     dim_model=args.dim_model,
 537     dim_keys=args.dim_keys,
 538     dim_hidden=args.dim_hidden,
 539     nb_heads=args.nb_heads,
 540     nb_blocks=args.nb_blocks,
 541     causal=True,
 542     dropout=args.dropout,
 543 )
 544
 545 model.to(device)
 546
 547 nb_parameters = sum(p.numel() for p in model.parameters())
 548 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 549
 550 ######################################################################
 551
 552 nb_epochs_finished = 0
 553
 554 if args.no_checkpoint:
 555     log_string(f"not trying to load checkpoint.")
 556
 557 else:
 558     try:
 559         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 560         checkpoint = torch.load(checkpoint_name)
 561         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 562         model.load_state_dict(checkpoint["model_state"])
 563         torch.set_rng_state(checkpoint["rng_state"])
 564         if torch.cuda.is_available():
 565             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 566
 567         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 568
 569     except FileNotFoundError:
 570         log_string("starting from scratch.")
 571
 572     except:
 573         log_string("error when loading the checkpoint.")
 574         exit(1)
 575
 576 ######################################################################
 577
 578 if args.task == "expr" and args.expr_input_file is not None:
 579     task.produce_results(
 580         n_epoch=nb_epochs_finished,
 581         model=model,
 582         result_dir=args.result_dir,
 583         logger=log_string,
 584         deterministic_synthesis=args.deterministic_synthesis,
 585         input_file=args.expr_input_file,
 586     )
 587
 588     exit(0)
 589
 590 ######################################################################
 591
 592 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 593
 594 # Compute the entropy of the training tokens
 595
 596 token_count = 0
 597 for input in task.batches(split="train"):
 598     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 599 token_probas = token_count / token_count.sum()
 600 entropy = -torch.xlogy(token_probas, token_probas).sum()
 601 train_set_perplexity = math.exp(entropy)
 602
 603 ######################################################################
 604 # A bit of paranoia never hurts
 605
 606 if args.max_percents_of_test_in_train >= 0:
 607
 608     def subsets_as_tuples(batches, cs):
 609         s = set()
 610         for batch in batches:
 611             for x in batch:
 612                 s.add(tuple([v.item() for v in x]))
 613                 if len(s) == cs:
 614                     yield s
 615                     s = set()
 616         yield s
 617
 618     nb_test, nb_in_train = 0, 0
 619     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 620         in_train = set()
 621         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 622             in_train.update(test_subset.intersection(train_subset))
 623         nb_in_train += len(in_train)
 624         nb_test += len(test_subset)
 625
 626     log_string(
 627         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 628     )
 629
 630     assert (
 631         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 632     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 633
 634 ##############################
 635
 636 if args.learning_rate_schedule == "cos":
 637     learning_rate_schedule = {}
 638     for n_epoch in range(args.nb_epochs):
 639         u = n_epoch / args.nb_epochs * math.pi
 640         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 641 else:
 642     u = {
 643         int(k): float(v)
 644         for k, v in [
 645             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 646         ]
 647     }
 648
 649     learning_rate_schedule = {}
 650     learning_rate = args.learning_rate
 651     for n_epoch in range(args.nb_epochs):
 652         if n_epoch in u:
 653             learning_rate = u[n_epoch]
 654         learning_rate_schedule[n_epoch] = learning_rate
 655
 656 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 657
 658 ##############################
 659
 660 nb_samples_seen = 0
 661
 662 if nb_epochs_finished >= nb_epochs:
 663     task.produce_results(
 664         n_epoch=nb_epochs_finished,
 665         model=model,
 666         result_dir=args.result_dir,
 667         logger=log_string,
 668         deterministic_synthesis=args.deterministic_synthesis,
 669     )
 670
 671 for n_epoch in range(nb_epochs_finished, nb_epochs):
 672     learning_rate = learning_rate_schedule[n_epoch]
 673
 674     log_string(f"learning_rate {learning_rate}")
 675
 676     if args.optim == "sgd":
 677         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 678     elif args.optim == "adam":
 679         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 680     elif args.optim == "adamw":
 681         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 682     else:
 683         raise ValueError(f"Unknown optimizer {args.optim}.")
 684
 685     model.train()
 686
 687     nb_train_samples, acc_train_loss = 0, 0.0
 688
 689     for input in task.batches(split="train"):
 690         input = input.to(device)
 691         output = model(mygpt.BracketedSequence(input)).x
 692         loss = F.cross_entropy(output.transpose(1, 2), input)
 693         acc_train_loss += loss.item() * input.size(0)
 694         nb_train_samples += input.size(0)
 695         nb_samples_seen += input.size(0)
 696
 697         optimizer.zero_grad()
 698         loss.backward()
 699         optimizer.step()
 700
 701     with torch.autograd.no_grad():
 702         model.eval()
 703
 704         nb_test_samples, acc_test_loss = 0, 0.0
 705
 706         for input in task.batches(split="test"):
 707             input = input.to(device)
 708
 709             output = model(mygpt.BracketedSequence(input)).x
 710             loss = F.cross_entropy(output.transpose(1, 2), input)
 711             acc_test_loss += loss.item() * input.size(0)
 712             nb_test_samples += input.size(0)
 713
 714         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 715         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 716
 717         log_string(
 718             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 719         )
 720
 721         task.produce_results(
 722             n_epoch=n_epoch,
 723             model=model,
 724             result_dir=args.result_dir,
 725             logger=log_string,
 726             deterministic_synthesis=args.deterministic_synthesis,
 727         )
 728
 729     checkpoint = {
 730         "nb_epochs_finished": n_epoch + 1,
 731         "model_state": model.state_dict(),
 732         "rng_state": torch.get_rng_state(),
 733     }
 734
 735     if torch.cuda.is_available():
 736         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 737
 738     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 739     torch.save(checkpoint, checkpoint_name)
 740     log_string(f"saved checkpoint {checkpoint_name}")
 741
 742 ######################################################################