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="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world",
  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 # sandbox options
 104
 105 parser.add_argument("--sandbox_level", type=int, default=0)
 106
 107 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
 108
 109 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
 110
 111 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
 112
 113 ##############################
 114 # picoclvr options
 115
 116 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 117
 118 parser.add_argument("--picoclvr_height", type=int, default=12)
 119
 120 parser.add_argument("--picoclvr_width", type=int, default=16)
 121
 122 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 123
 124 ##############################
 125 # Maze options
 126
 127 parser.add_argument("--maze_height", type=int, default=23)
 128
 129 parser.add_argument("--maze_width", type=int, default=39)
 130
 131 parser.add_argument("--maze_nb_walls", type=int, default=45)
 132
 133 ##############################
 134 # Snake options
 135
 136 parser.add_argument("--snake_height", type=int, default=6)
 137
 138 parser.add_argument("--snake_width", type=int, default=8)
 139
 140 parser.add_argument("--snake_nb_colors", type=int, default=5)
 141
 142 parser.add_argument("--snake_length", type=int, default=200)
 143
 144 ##############################
 145 # Stack options
 146
 147 parser.add_argument("--stack_nb_steps", type=int, default=100)
 148
 149 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 150
 151 parser.add_argument("--stack_nb_digits", type=int, default=3)
 152
 153 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 154
 155 ##############################
 156 # Expr options
 157
 158 parser.add_argument("--expr_nb_variables", type=int, default=5)
 159
 160 parser.add_argument("--expr_sequence_length", type=int, default=40)
 161
 162 parser.add_argument("--expr_operand_max", type=int, default=9)
 163
 164 parser.add_argument("--expr_result_max", type=int, default=99)
 165
 166 parser.add_argument("--expr_input_file", type=str, default=None)
 167
 168 ##############################
 169 # World options
 170
 171 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 172
 173 ######################################################################
 174
 175 args = parser.parse_args()
 176
 177 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 178
 179 if args.result_dir is None:
 180     args.result_dir = f"results_{args.task}"
 181
 182 ######################################################################
 183
 184 default_task_args = {
 185     "sandbox": {
 186         "nb_epochs": 50,
 187         "batch_size": 25,
 188         "nb_train_samples": 100000,
 189         "nb_test_samples": 10000,
 190     },
 191     "picoclvr": {
 192         "nb_epochs": 25,
 193         "batch_size": 25,
 194         "nb_train_samples": 250000,
 195         "nb_test_samples": 10000,
 196     },
 197     "mnist": {
 198         "nb_epochs": 25,
 199         "batch_size": 10,
 200         "nb_train_samples": 250000,
 201         "nb_test_samples": 10000,
 202     },
 203     "maze": {
 204         "nb_epochs": 25,
 205         "batch_size": 5,
 206         "nb_train_samples": 250000,
 207         "nb_test_samples": 10000,
 208     },
 209     "snake": {
 210         "nb_epochs": 5,
 211         "batch_size": 25,
 212         "nb_train_samples": 250000,
 213         "nb_test_samples": 10000,
 214     },
 215     "stack": {
 216         "nb_epochs": 5,
 217         "batch_size": 25,
 218         "nb_train_samples": 100000,
 219         "nb_test_samples": 1000,
 220     },
 221     "expr": {
 222         "nb_epochs": 40,
 223         "batch_size": 25,
 224         "nb_train_samples": 1000000,
 225         "nb_test_samples": 10000,
 226     },
 227     "rpl": {
 228         "nb_epochs": 40,
 229         "batch_size": 25,
 230         "nb_train_samples": 100000,
 231         "nb_test_samples": 10000,
 232     },
 233     "world": {
 234         "nb_epochs": 10,
 235         "batch_size": 25,
 236         "nb_train_samples": 25000,
 237         "nb_test_samples": 1000,
 238     },
 239 }
 240
 241 if args.task in default_task_args:
 242     for k, v in default_task_args[args.task].items():
 243         if getattr(args, k) is None:
 244             setattr(args, k, v)
 245
 246 ######################################################################
 247
 248 default_model_args = {
 249     "17K": {
 250         "dim_model": 32,
 251         "dim_keys": 32,
 252         "dim_hidden": 32,
 253         "nb_heads": 2,
 254         "nb_blocks": 2,
 255     },
 256     "37M": {
 257         "dim_model": 512,
 258         "dim_keys": 64,
 259         "dim_hidden": 2048,
 260         "nb_heads": 8,
 261         "nb_blocks": 12,
 262     },
 263     "122M": {
 264         "dim_model": 768,
 265         "dim_keys": 64,
 266         "dim_hidden": 2048,
 267         "nb_heads": 8,
 268         "nb_blocks": 24,
 269     },
 270     "352M": {
 271         "dim_model": 1024,
 272         "dim_keys": 64,
 273         "dim_hidden": 2048,
 274         "nb_heads": 8,
 275         "nb_blocks": 48,
 276     },
 277 }
 278
 279 if args.model in default_model_args:
 280     for k, v in default_model_args[args.model].items():
 281         if getattr(args, k) is None:
 282             setattr(args, k, v)
 283 else:
 284     raise ValueError(f"Unknown model {args.model}")
 285
 286 ######################################################################
 287
 288 try:
 289     os.mkdir(args.result_dir)
 290 except FileExistsError:
 291     if not args.overwrite_results:
 292         print(f"result directory {args.result_dir} already exists")
 293         exit(1)
 294
 295 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 296
 297 if args.seed >= 0:
 298     # torch.backends.cudnn.deterministic = True
 299     # torch.backends.cudnn.benchmark = False
 300     # torch.use_deterministic_algorithms(True)
 301     torch.manual_seed(args.seed)
 302     if torch.cuda.is_available():
 303         torch.cuda.manual_seed_all(args.seed)
 304
 305 ######################################################################
 306
 307
 308 def log_string(s):
 309     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 310
 311     if log_file is not None:
 312         log_file.write(t + s + "\n")
 313         log_file.flush()
 314
 315     print(t + s)
 316     sys.stdout.flush()
 317
 318
 319 for n in vars(args):
 320     log_string(f"args.{n} {getattr(args, n)}")
 321
 322
 323 ######################################################################
 324
 325
 326 def picoclvr_pruner_horizontal_green(p):
 327     return not ("green" in p and ("left" in p or "right" in p))
 328
 329
 330 picoclvr_pruner_train = (
 331     picoclvr_pruner_horizontal_green
 332     if args.picocvlr_prune_properties in {"train+eval"}
 333     else None
 334 )
 335
 336 picoclvr_pruner_eval = (
 337     (lambda p: not picoclvr_pruner_horizontal_green(p))
 338     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 339     else None
 340 )
 341
 342 ######################################################################
 343
 344 if args.task == "sandbox":
 345     if args.sandbox_level == 0:
 346         problem = problems.ProblemLevel0(
 347             nb_sentences=args.sandbox_levels_nb_items,
 348             len_prompt=args.sandbox_levels_len_source,
 349             len_result=args.sandbox_levels_len_result,
 350         )
 351     elif args.sandbox_level == 1:
 352         problem = problems.ProblemLevel1(
 353             nb_operators=args.sandbox_levels_nb_items,
 354             len_source=args.sandbox_levels_len_source,
 355             len_result=args.sandbox_levels_len_result,
 356         )
 357     elif args.sandbox_level == 2:
 358         problem = problems.ProblemLevel2(
 359             len_source=args.sandbox_levels_len_source,
 360             len_result=args.sandbox_levels_len_result,
 361         )
 362     else:
 363         raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
 364
 365     task = tasks.SandBox(
 366         # problem,
 367         # problems.ProblemAddition(zero_padded=False, inverted_result=False),
 368         # problems.ProblemLenId(len_max=args.sandbox_levels_len_source),
 369         problems.ProblemTwoTargets(len_total=12, len_targets=4),
 370         nb_train_samples=args.nb_train_samples,
 371         nb_test_samples=args.nb_test_samples,
 372         batch_size=args.batch_size,
 373         logger=log_string,
 374         device=device,
 375     )
 376
 377 elif args.task == "picoclvr":
 378     task = tasks.PicoCLVR(
 379         nb_train_samples=args.nb_train_samples,
 380         nb_test_samples=args.nb_test_samples,
 381         batch_size=args.batch_size,
 382         height=args.picoclvr_height,
 383         width=args.picoclvr_width,
 384         nb_colors=args.picoclvr_nb_colors,
 385         logger=log_string,
 386         device=device,
 387         pruner_train=picoclvr_pruner_train,
 388         pruner_eval=picoclvr_pruner_eval,
 389     )
 390
 391 elif args.task == "mnist":
 392     task = tasks.MNIST(
 393         nb_train_samples=args.nb_train_samples,
 394         nb_test_samples=args.nb_test_samples,
 395         batch_size=args.batch_size,
 396         device=device,
 397     )
 398
 399 elif args.task == "maze":
 400     task = tasks.Maze(
 401         nb_train_samples=args.nb_train_samples,
 402         nb_test_samples=args.nb_test_samples,
 403         batch_size=args.batch_size,
 404         height=args.maze_height,
 405         width=args.maze_width,
 406         nb_walls=args.maze_nb_walls,
 407         device=device,
 408     )
 409
 410 elif args.task == "snake":
 411     task = tasks.Snake(
 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.snake_height,
 416         width=args.snake_width,
 417         nb_colors=args.snake_nb_colors,
 418         length=args.snake_length,
 419         prompt_length=args.snake_length // 2,
 420         device=device,
 421     )
 422
 423 elif args.task == "stack":
 424     task = tasks.Stack(
 425         nb_train_samples=args.nb_train_samples,
 426         nb_test_samples=args.nb_test_samples,
 427         batch_size=args.batch_size,
 428         logger=log_string,
 429         nb_steps=args.stack_nb_steps,
 430         nb_stacks=args.stack_nb_stacks,
 431         nb_digits=args.stack_nb_digits,
 432         fraction_values_for_train=args.stack_fraction_values_for_train,
 433         device=device,
 434     )
 435
 436 elif args.task == "expr":
 437     task = tasks.Expr(
 438         nb_train_samples=args.nb_train_samples,
 439         nb_test_samples=args.nb_test_samples,
 440         nb_variables=args.expr_nb_variables,
 441         sequence_length=args.expr_sequence_length,
 442         operand_max=args.expr_operand_max,
 443         result_max=args.expr_result_max,
 444         batch_size=args.batch_size,
 445         device=device,
 446     )
 447
 448 elif args.task == "rpl":
 449     task = tasks.RPL(
 450         nb_train_samples=args.nb_train_samples,
 451         nb_test_samples=args.nb_test_samples,
 452         batch_size=args.batch_size,
 453         nb_starting_values=args.rpl_nb_starting_values,
 454         max_input=args.rpl_max_input,
 455         prog_len=args.rpl_prog_len,
 456         nb_runs=args.rpl_nb_runs,
 457         no_prog=args.rpl_no_prog,
 458         logger=log_string,
 459         device=device,
 460     )
 461
 462 elif args.task == "world":
 463     task = tasks.World(
 464         nb_train_samples=args.nb_train_samples,
 465         nb_test_samples=args.nb_test_samples,
 466         batch_size=args.batch_size,
 467         vqae_nb_epochs=args.world_vqae_nb_epochs,
 468         logger=log_string,
 469         device=device,
 470     )
 471
 472 else:
 473     raise ValueError(f"Unknown task {args.task}")
 474
 475 ######################################################################
 476
 477 log_string(f"device {device}")
 478
 479 vocabulary_size = task.vocabulary_size()
 480
 481 log_string(f"vocabulary_size {vocabulary_size}")
 482
 483 ##############################
 484
 485 model = mygpt.MyGPT(
 486     vocabulary_size=vocabulary_size,
 487     dim_model=args.dim_model,
 488     dim_keys=args.dim_keys,
 489     dim_hidden=args.dim_hidden,
 490     nb_heads=args.nb_heads,
 491     nb_blocks=args.nb_blocks,
 492     causal=True,
 493     dropout=args.dropout,
 494 )
 495
 496 model.to(device)
 497
 498 nb_parameters = sum(p.numel() for p in model.parameters())
 499 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 500
 501 ######################################################################
 502
 503 nb_epochs_finished = 0
 504
 505 if args.no_checkpoint:
 506     log_string(f"not trying to load checkpoint.")
 507
 508 else:
 509     try:
 510         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 511         checkpoint = torch.load(checkpoint_name)
 512         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 513         model.load_state_dict(checkpoint["model_state"])
 514         torch.set_rng_state(checkpoint["rng_state"])
 515         if torch.cuda.is_available():
 516             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 517
 518         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 519
 520     except FileNotFoundError:
 521         log_string("starting from scratch.")
 522
 523     except:
 524         log_string("error when loading the checkpoint.")
 525         exit(1)
 526
 527 ######################################################################
 528
 529 if args.task == "expr" and args.expr_input_file is not None:
 530     task.produce_results(
 531         n_epoch=nb_epochs_finished,
 532         model=model,
 533         result_dir=args.result_dir,
 534         logger=log_string,
 535         deterministic_synthesis=args.deterministic_synthesis,
 536         input_file=args.expr_input_file,
 537     )
 538
 539     exit(0)
 540
 541 ######################################################################
 542
 543 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 544
 545 # Compute the entropy of the training tokens
 546
 547 token_count = 0
 548 for input in task.batches(split="train"):
 549     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 550 token_probas = token_count / token_count.sum()
 551 entropy = -torch.xlogy(token_probas, token_probas).sum()
 552 train_set_perplexity = math.exp(entropy)
 553
 554 ######################################################################
 555 # A bit of paranoia never hurts
 556
 557
 558 def subsets_as_tuples(batches, cs):
 559     s = set()
 560     for batch in batches:
 561         for x in batch:
 562             s.add(tuple([v.item() for v in x]))
 563             if len(s) == cs:
 564                 yield s
 565                 s = set()
 566     yield s
 567
 568
 569 nb_test, nb_in_train = 0, 0
 570 for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 571     in_train = set()
 572     for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 573         in_train.update(test_subset.intersection(train_subset))
 574     nb_in_train += len(in_train)
 575     nb_test += len(test_subset)
 576
 577 log_string(
 578     f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 579 )
 580
 581 assert (
 582     nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 583 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 584
 585 ##############################
 586
 587 if args.learning_rate_schedule == "cos":
 588     learning_rate_schedule = {}
 589     for n_epoch in range(args.nb_epochs):
 590         u = n_epoch / args.nb_epochs * math.pi
 591         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 592 else:
 593     u = {
 594         int(k): float(v)
 595         for k, v in [
 596             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 597         ]
 598     }
 599
 600     learning_rate_schedule = {}
 601     learning_rate = args.learning_rate
 602     for n_epoch in range(args.nb_epochs):
 603         if n_epoch in u:
 604             learning_rate = u[n_epoch]
 605         learning_rate_schedule[n_epoch] = learning_rate
 606
 607 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 608
 609 ##############################
 610
 611 nb_samples_seen = 0
 612
 613 if nb_epochs_finished >= nb_epochs:
 614     task.produce_results(
 615         n_epoch=nb_epochs_finished,
 616         model=model,
 617         result_dir=args.result_dir,
 618         logger=log_string,
 619         deterministic_synthesis=args.deterministic_synthesis,
 620     )
 621
 622 for n_epoch in range(nb_epochs_finished, nb_epochs):
 623     learning_rate = learning_rate_schedule[n_epoch]
 624
 625     log_string(f"learning_rate {learning_rate}")
 626
 627     if args.optim == "sgd":
 628         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 629     elif args.optim == "adam":
 630         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 631     elif args.optim == "adamw":
 632         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 633     else:
 634         raise ValueError(f"Unknown optimizer {args.optim}.")
 635
 636     model.train()
 637
 638     nb_train_samples, acc_train_loss = 0, 0.0
 639
 640     for input in task.batches(split="train"):
 641         input = input.to(device)
 642         output = model(mygpt.BracketedSequence(input)).x
 643         loss = F.cross_entropy(output.transpose(1, 2), input)
 644         acc_train_loss += loss.item() * input.size(0)
 645         nb_train_samples += input.size(0)
 646         nb_samples_seen += input.size(0)
 647
 648         optimizer.zero_grad()
 649         loss.backward()
 650         optimizer.step()
 651
 652     with torch.autograd.no_grad():
 653         model.eval()
 654
 655         nb_test_samples, acc_test_loss = 0, 0.0
 656
 657         for input in task.batches(split="test"):
 658             input = input.to(device)
 659
 660             output = model(mygpt.BracketedSequence(input)).x
 661             loss = F.cross_entropy(output.transpose(1, 2), input)
 662             acc_test_loss += loss.item() * input.size(0)
 663             nb_test_samples += input.size(0)
 664
 665         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 666         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 667
 668         log_string(
 669             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 670         )
 671
 672         task.produce_results(
 673             n_epoch=n_epoch,
 674             model=model,
 675             result_dir=args.result_dir,
 676             logger=log_string,
 677             deterministic_synthesis=args.deterministic_synthesis,
 678         )
 679
 680     checkpoint = {
 681         "nb_epochs_finished": n_epoch + 1,
 682         "model_state": model.state_dict(),
 683         "rng_state": torch.get_rng_state(),
 684     }
 685
 686     if torch.cuda.is_available():
 687         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 688
 689     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 690     torch.save(checkpoint, checkpoint_name)
 691     log_string(f"saved checkpoint {checkpoint_name}")
 692
 693 ######################################################################