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
  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("--nb_epochs", type=int, default=None)
  46
  47 parser.add_argument("--batch_size", type=int, default=None)
  48
  49 parser.add_argument("--nb_train_samples", type=int, default=None)
  50
  51 parser.add_argument("--nb_test_samples", type=int, default=None)
  52
  53 parser.add_argument("--optim", type=str, default="adam")
  54
  55 parser.add_argument("--learning_rate", type=float, default=1e-4)
  56
  57 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  58
  59 parser.add_argument("--model", type=str, default="37M")
  60
  61 parser.add_argument("--dim_model", type=int, default=None)
  62
  63 parser.add_argument("--dim_keys", type=int, default=None)
  64
  65 parser.add_argument("--dim_hidden", type=int, default=None)
  66
  67 parser.add_argument("--nb_heads", type=int, default=None)
  68
  69 parser.add_argument("--nb_blocks", type=int, default=None)
  70
  71 parser.add_argument("--dropout", type=float, default=0.1)
  72
  73 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  74
  75 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  76
  77 parser.add_argument("--overwrite_results", action="store_true", default=False)
  78
  79 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  80
  81 ##############################
  82 # rpl options
  83
  84 parser.add_argument("--rpl_nb_starting_values", type=int, default=5)
  85
  86 parser.add_argument("--rpl_max_input", type=int, default=9)
  87
  88 parser.add_argument("--rpl_prog_len", type=int, default=10)
  89
  90 parser.add_argument("--rpl_nb_runs", type=int, default=8)
  91
  92 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
  93
  94 ##############################
  95 # sandbox options
  96
  97 parser.add_argument("--sandbox_level", type=int, default=0)
  98
  99 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
 100
 101 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
 102
 103 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
 104
 105 ##############################
 106 # picoclvr options
 107
 108 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 109
 110 parser.add_argument("--picoclvr_height", type=int, default=12)
 111
 112 parser.add_argument("--picoclvr_width", type=int, default=16)
 113
 114 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 115
 116 ##############################
 117 # Maze options
 118
 119 parser.add_argument("--maze_height", type=int, default=23)
 120
 121 parser.add_argument("--maze_width", type=int, default=39)
 122
 123 parser.add_argument("--maze_nb_walls", type=int, default=45)
 124
 125 ##############################
 126 # Snake options
 127
 128 parser.add_argument("--snake_height", type=int, default=6)
 129
 130 parser.add_argument("--snake_width", type=int, default=8)
 131
 132 parser.add_argument("--snake_nb_colors", type=int, default=5)
 133
 134 parser.add_argument("--snake_length", type=int, default=200)
 135
 136 ##############################
 137 # Stack options
 138
 139 parser.add_argument("--stack_nb_steps", type=int, default=100)
 140
 141 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 142
 143 parser.add_argument("--stack_nb_digits", type=int, default=3)
 144
 145 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 146
 147 ##############################
 148 # Expr options
 149
 150 parser.add_argument("--expr_nb_variables", type=int, default=5)
 151
 152 parser.add_argument("--expr_sequence_length", type=int, default=40)
 153
 154 parser.add_argument("--expr_operand_max", type=int, default=9)
 155
 156 parser.add_argument("--expr_result_max", type=int, default=99)
 157
 158 parser.add_argument("--expr_input_file", type=str, default=None)
 159
 160 ##############################
 161 # World options
 162
 163 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 164
 165 ######################################################################
 166
 167 args = parser.parse_args()
 168
 169 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 170
 171 if args.result_dir is None:
 172     args.result_dir = f"results_{args.task}"
 173
 174 ######################################################################
 175
 176 default_task_args = {
 177     "sandbox": {
 178         "nb_epochs": 50,
 179         "batch_size": 25,
 180         "nb_train_samples": 100000,
 181         "nb_test_samples": 10000,
 182     },
 183     "picoclvr": {
 184         "nb_epochs": 25,
 185         "batch_size": 25,
 186         "nb_train_samples": 250000,
 187         "nb_test_samples": 10000,
 188     },
 189     "mnist": {
 190         "nb_epochs": 25,
 191         "batch_size": 10,
 192         "nb_train_samples": 250000,
 193         "nb_test_samples": 10000,
 194     },
 195     "maze": {
 196         "nb_epochs": 25,
 197         "batch_size": 5,
 198         "nb_train_samples": 250000,
 199         "nb_test_samples": 10000,
 200     },
 201     "snake": {
 202         "nb_epochs": 5,
 203         "batch_size": 25,
 204         "nb_train_samples": 250000,
 205         "nb_test_samples": 10000,
 206     },
 207     "stack": {
 208         "nb_epochs": 5,
 209         "batch_size": 25,
 210         "nb_train_samples": 100000,
 211         "nb_test_samples": 1000,
 212     },
 213     "expr": {
 214         "nb_epochs": 40,
 215         "batch_size": 25,
 216         "nb_train_samples": 1000000,
 217         "nb_test_samples": 10000,
 218     },
 219     "rpl": {
 220         "nb_epochs": 40,
 221         "batch_size": 25,
 222         "nb_train_samples": 100000,
 223         "nb_test_samples": 10000,
 224     },
 225     "world": {
 226         "nb_epochs": 10,
 227         "batch_size": 25,
 228         "nb_train_samples": 25000,
 229         "nb_test_samples": 1000,
 230     },
 231 }
 232
 233 if args.task in default_task_args:
 234     for k, v in default_task_args[args.task].items():
 235         if getattr(args, k) is None:
 236             setattr(args, k, v)
 237
 238 ######################################################################
 239
 240 default_model_args = {
 241     "17K": {
 242         "dim_model": 32,
 243         "dim_keys": 32,
 244         "dim_hidden": 32,
 245         "nb_heads": 2,
 246         "nb_blocks": 2,
 247     },
 248     "37M": {
 249         "dim_model": 512,
 250         "dim_keys": 64,
 251         "dim_hidden": 2048,
 252         "nb_heads": 8,
 253         "nb_blocks": 12,
 254     },
 255     "122M": {
 256         "dim_model": 768,
 257         "dim_keys": 64,
 258         "dim_hidden": 2048,
 259         "nb_heads": 8,
 260         "nb_blocks": 24,
 261     },
 262     "352M": {
 263         "dim_model": 1024,
 264         "dim_keys": 64,
 265         "dim_hidden": 2048,
 266         "nb_heads": 8,
 267         "nb_blocks": 48,
 268     },
 269 }
 270
 271 if args.model in default_model_args:
 272     for k, v in default_model_args[args.model].items():
 273         if getattr(args, k) is None:
 274             setattr(args, k, v)
 275 else:
 276     raise ValueError(f"Unknown model {args.model}")
 277
 278 ######################################################################
 279
 280 try:
 281     os.mkdir(args.result_dir)
 282 except FileExistsError:
 283     if not args.overwrite_results:
 284         print(f"result directory {args.result_dir} already exists")
 285         exit(1)
 286
 287 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 288
 289 if args.seed >= 0:
 290     # torch.backends.cudnn.deterministic = True
 291     # torch.backends.cudnn.benchmark = False
 292     # torch.use_deterministic_algorithms(True)
 293     torch.manual_seed(args.seed)
 294     if torch.cuda.is_available():
 295         torch.cuda.manual_seed_all(args.seed)
 296
 297 ######################################################################
 298
 299
 300 def log_string(s):
 301     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 302
 303     if log_file is not None:
 304         log_file.write(t + s + "\n")
 305         log_file.flush()
 306
 307     print(t + s)
 308     sys.stdout.flush()
 309
 310
 311 for n in vars(args):
 312     log_string(f"args.{n} {getattr(args, n)}")
 313
 314
 315 ######################################################################
 316
 317
 318 def picoclvr_pruner_horizontal_green(p):
 319     return not ("green" in p and ("left" in p or "right" in p))
 320
 321
 322 picoclvr_pruner_train = (
 323     picoclvr_pruner_horizontal_green
 324     if args.picocvlr_prune_properties in {"train+eval"}
 325     else None
 326 )
 327
 328 picoclvr_pruner_eval = (
 329     (lambda p: not picoclvr_pruner_horizontal_green(p))
 330     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 331     else None
 332 )
 333
 334 ######################################################################
 335
 336 if args.task == "sandbox":
 337     if args.sandbox_level == 0:
 338         problem = tasks.ProblemLevel0(
 339             nb_sentences=args.sandbox_levels_nb_items,
 340             len_prompt=args.sandbox_levels_len_source,
 341             len_result=args.sandbox_levels_len_result,
 342         )
 343     elif args.sandbox_level == 1:
 344         problem = tasks.ProblemLevel1(
 345             nb_operators=args.sandbox_levels_nb_items,
 346             len_source=args.sandbox_levels_len_source,
 347             len_result=args.sandbox_levels_len_result,
 348         )
 349     elif args.sandbox_level == 2:
 350         problem = tasks.ProblemLevel2(
 351             len_source=args.sandbox_levels_len_source,
 352             len_result=args.sandbox_levels_len_result,
 353         )
 354     else:
 355         raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
 356
 357     task = tasks.SandBox(
 358         problem,
 359         # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
 360         nb_train_samples=args.nb_train_samples,
 361         nb_test_samples=args.nb_test_samples,
 362         batch_size=args.batch_size,
 363         logger=log_string,
 364         device=device,
 365     )
 366
 367 elif args.task == "picoclvr":
 368     task = tasks.PicoCLVR(
 369         nb_train_samples=args.nb_train_samples,
 370         nb_test_samples=args.nb_test_samples,
 371         batch_size=args.batch_size,
 372         height=args.picoclvr_height,
 373         width=args.picoclvr_width,
 374         nb_colors=args.picoclvr_nb_colors,
 375         logger=log_string,
 376         device=device,
 377         pruner_train=picoclvr_pruner_train,
 378         pruner_eval=picoclvr_pruner_eval,
 379     )
 380
 381 elif args.task == "mnist":
 382     task = tasks.MNIST(
 383         nb_train_samples=args.nb_train_samples,
 384         nb_test_samples=args.nb_test_samples,
 385         batch_size=args.batch_size,
 386         device=device,
 387     )
 388
 389 elif args.task == "maze":
 390     task = tasks.Maze(
 391         nb_train_samples=args.nb_train_samples,
 392         nb_test_samples=args.nb_test_samples,
 393         batch_size=args.batch_size,
 394         height=args.maze_height,
 395         width=args.maze_width,
 396         nb_walls=args.maze_nb_walls,
 397         device=device,
 398     )
 399
 400 elif args.task == "snake":
 401     task = tasks.Snake(
 402         nb_train_samples=args.nb_train_samples,
 403         nb_test_samples=args.nb_test_samples,
 404         batch_size=args.batch_size,
 405         height=args.snake_height,
 406         width=args.snake_width,
 407         nb_colors=args.snake_nb_colors,
 408         length=args.snake_length,
 409         prompt_length=args.snake_length // 2,
 410         device=device,
 411     )
 412
 413 elif args.task == "stack":
 414     task = tasks.Stack(
 415         nb_train_samples=args.nb_train_samples,
 416         nb_test_samples=args.nb_test_samples,
 417         batch_size=args.batch_size,
 418         logger=log_string,
 419         nb_steps=args.stack_nb_steps,
 420         nb_stacks=args.stack_nb_stacks,
 421         nb_digits=args.stack_nb_digits,
 422         fraction_values_for_train=args.stack_fraction_values_for_train,
 423         device=device,
 424     )
 425
 426 elif args.task == "expr":
 427     task = tasks.Expr(
 428         nb_train_samples=args.nb_train_samples,
 429         nb_test_samples=args.nb_test_samples,
 430         nb_variables=args.expr_nb_variables,
 431         sequence_length=args.expr_sequence_length,
 432         operand_max=args.expr_operand_max,
 433         result_max=args.expr_result_max,
 434         batch_size=args.batch_size,
 435         device=device,
 436     )
 437
 438 elif args.task == "rpl":
 439     task = tasks.RPL(
 440         nb_train_samples=args.nb_train_samples,
 441         nb_test_samples=args.nb_test_samples,
 442         batch_size=args.batch_size,
 443         nb_starting_values=args.rpl_nb_starting_values,
 444         max_input=args.rpl_max_input,
 445         prog_len=args.rpl_prog_len,
 446         nb_runs=args.rpl_nb_runs,
 447         no_prog=args.rpl_no_prog,
 448         logger=log_string,
 449         device=device,
 450     )
 451
 452 elif args.task == "world":
 453     task = tasks.World(
 454         nb_train_samples=args.nb_train_samples,
 455         nb_test_samples=args.nb_test_samples,
 456         batch_size=args.batch_size,
 457         vqae_nb_epochs=args.world_vqae_nb_epochs,
 458         logger=log_string,
 459         device=device,
 460     )
 461
 462 else:
 463     raise ValueError(f"Unknown task {args.task}")
 464
 465 ######################################################################
 466
 467 log_string(f"device {device}")
 468
 469 vocabulary_size = task.vocabulary_size()
 470
 471 log_string(f"vocabulary_size {vocabulary_size}")
 472
 473 ##############################
 474
 475 model = mygpt.MyGPT(
 476     vocabulary_size=vocabulary_size,
 477     dim_model=args.dim_model,
 478     dim_keys=args.dim_keys,
 479     dim_hidden=args.dim_hidden,
 480     nb_heads=args.nb_heads,
 481     nb_blocks=args.nb_blocks,
 482     causal=True,
 483     dropout=args.dropout,
 484 )
 485
 486 model.to(device)
 487
 488 nb_parameters = sum(p.numel() for p in model.parameters())
 489 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 490
 491 ######################################################################
 492
 493 nb_epochs_finished = 0
 494
 495 if args.no_checkpoint:
 496     log_string(f"not trying to load checkpoint.")
 497
 498 else:
 499     try:
 500         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 501         checkpoint = torch.load(checkpoint_name)
 502         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 503         model.load_state_dict(checkpoint["model_state"])
 504         torch.set_rng_state(checkpoint["rng_state"])
 505         if torch.cuda.is_available():
 506             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 507
 508         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 509
 510     except FileNotFoundError:
 511         log_string("starting from scratch.")
 512
 513     except:
 514         log_string("error when loading the checkpoint.")
 515         exit(1)
 516
 517 ######################################################################
 518
 519 if args.task == "expr" and args.expr_input_file is not None:
 520     task.produce_results(
 521         n_epoch=nb_epochs_finished,
 522         model=model,
 523         result_dir=args.result_dir,
 524         logger=log_string,
 525         deterministic_synthesis=args.deterministic_synthesis,
 526         input_file=args.expr_input_file,
 527     )
 528
 529     exit(0)
 530
 531 ######################################################################
 532
 533 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 534
 535 # Compute the entropy of the training tokens
 536
 537 token_count = 0
 538 for input in task.batches(split="train"):
 539     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 540 token_probas = token_count / token_count.sum()
 541 entropy = -torch.xlogy(token_probas, token_probas).sum()
 542 train_set_perplexity = math.exp(entropy)
 543
 544 ##############################
 545
 546 # A bit of paranoia never hurts
 547
 548 train_examples = {}
 549
 550
 551 for input in task.batches(split="train"):
 552     assert input.dim() == 2 and input.dtype == torch.int64
 553     for x in input:
 554         train_examples[x.sum().item()] = x
 555
 556 nb_total, nb_collisions = 0, 0
 557 for input in task.batches(split="test"):
 558     assert input.dim() == 2 and input.dtype == torch.int64
 559     for x in input:
 560         nb_total += 1
 561         y = train_examples.get(x.sum().item())
 562         if y is not None:
 563             if x.size() == y.size() and (x - y).abs().sum() == 0:
 564                 nb_collisions += 1
 565
 566 del train_examples
 567
 568 log_string(
 569     f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
 570 )
 571
 572 ##############################
 573
 574 if args.learning_rate_schedule == "cos":
 575     learning_rate_schedule = {}
 576     for n_epoch in range(args.nb_epochs):
 577         u = n_epoch / args.nb_epochs * math.pi
 578         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 579 else:
 580     u = {
 581         int(k): float(v)
 582         for k, v in [
 583             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 584         ]
 585     }
 586
 587     learning_rate_schedule = {}
 588     learning_rate = args.learning_rate
 589     for n_epoch in range(args.nb_epochs):
 590         if n_epoch in u:
 591             learning_rate = u[n_epoch]
 592         learning_rate_schedule[n_epoch] = learning_rate
 593
 594 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 595
 596 ##############################
 597
 598 nb_samples_seen = 0
 599
 600 if nb_epochs_finished >= nb_epochs:
 601     task.produce_results(
 602         n_epoch=nb_epochs_finished,
 603         model=model,
 604         result_dir=args.result_dir,
 605         logger=log_string,
 606         deterministic_synthesis=args.deterministic_synthesis,
 607     )
 608
 609 for n_epoch in range(nb_epochs_finished, nb_epochs):
 610     learning_rate = learning_rate_schedule[n_epoch]
 611
 612     log_string(f"learning_rate {learning_rate}")
 613
 614     if args.optim == "sgd":
 615         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 616     elif args.optim == "adam":
 617         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 618     elif args.optim == "adamw":
 619         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 620     else:
 621         raise ValueError(f"Unknown optimizer {args.optim}.")
 622
 623     model.train()
 624
 625     nb_train_samples, acc_train_loss = 0, 0.0
 626
 627     for input in task.batches(split="train"):
 628         input = input.to(device)
 629         output = model(mygpt.BracketedSequence(input)).x
 630         loss = F.cross_entropy(output.transpose(1, 2), input)
 631         acc_train_loss += loss.item() * input.size(0)
 632         nb_train_samples += input.size(0)
 633         nb_samples_seen += input.size(0)
 634
 635         optimizer.zero_grad()
 636         loss.backward()
 637         optimizer.step()
 638
 639     with torch.autograd.no_grad():
 640         model.eval()
 641
 642         nb_test_samples, acc_test_loss = 0, 0.0
 643
 644         for input in task.batches(split="test"):
 645             input = input.to(device)
 646
 647             output = model(mygpt.BracketedSequence(input)).x
 648             loss = F.cross_entropy(output.transpose(1, 2), input)
 649             acc_test_loss += loss.item() * input.size(0)
 650             nb_test_samples += input.size(0)
 651
 652         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 653         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 654
 655         log_string(
 656             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 657         )
 658
 659         task.produce_results(
 660             n_epoch=n_epoch,
 661             model=model,
 662             result_dir=args.result_dir,
 663             logger=log_string,
 664             deterministic_synthesis=args.deterministic_synthesis,
 665         )
 666
 667     checkpoint = {
 668         "nb_epochs_finished": n_epoch + 1,
 669         "model_state": model.state_dict(),
 670         "rng_state": torch.get_rng_state(),
 671     }
 672
 673     if torch.cuda.is_available():
 674         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 675
 676     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 677     torch.save(checkpoint, checkpoint_name)
 678     log_string(f"saved checkpoint {checkpoint_name}")
 679
 680 ######################################################################