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 # torch.backends.cuda.matmul.allow_tf23
   9 # torch.autocast(torch.bfloat16)
  10
  11 import math, sys, argparse, time, tqdm, os
  12
  13 import torch, torchvision
  14 from torch import nn
  15 from torch.nn import functional as F
  16
  17 import ffutils
  18 import mygpt, tasks
  19
  20 ######################################################################
  21
  22 if torch.cuda.is_available():
  23     device = torch.device("cuda")
  24     torch.backends.cuda.matmul.allow_tf32 = True
  25 else:
  26     device = torch.device("cpu")
  27
  28 ######################################################################
  29
  30 parser = argparse.ArgumentParser(
  31     description="An implementation of GPT with cache.",
  32     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  33 )
  34
  35 parser.add_argument(
  36     "--task",
  37     type=str,
  38     default="sandbox",
  39     help="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world",
  40 )
  41
  42 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  43
  44 parser.add_argument("--result_dir", type=str, default=None)
  45
  46 parser.add_argument("--seed", type=int, default=0)
  47
  48 parser.add_argument("--nb_epochs", type=int, default=None)
  49
  50 parser.add_argument("--batch_size", type=int, default=None)
  51
  52 parser.add_argument("--nb_train_samples", type=int, default=None)
  53
  54 parser.add_argument("--nb_test_samples", type=int, default=None)
  55
  56 parser.add_argument("--optim", type=str, default="adam")
  57
  58 parser.add_argument("--learning_rate", type=float, default=1e-4)
  59
  60 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  61
  62 parser.add_argument("--model", type=str, default="37M")
  63
  64 parser.add_argument("--dim_model", type=int, default=None)
  65
  66 parser.add_argument("--dim_keys", type=int, default=None)
  67
  68 parser.add_argument("--dim_hidden", type=int, default=None)
  69
  70 parser.add_argument("--nb_heads", type=int, default=None)
  71
  72 parser.add_argument("--nb_blocks", type=int, default=None)
  73
  74 parser.add_argument("--dropout", type=float, default=0.1)
  75
  76 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  77
  78 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  79
  80 parser.add_argument("--overwrite_results", action="store_true", default=False)
  81
  82 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  83
  84 ##############################
  85 # picoclvr options
  86
  87 parser.add_argument("--sandbox_level", type=int, default=0)
  88
  89 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
  90
  91 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
  92
  93 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
  94
  95 ##############################
  96 # picoclvr options
  97
  98 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
  99
 100 parser.add_argument("--picoclvr_height", type=int, default=12)
 101
 102 parser.add_argument("--picoclvr_width", type=int, default=16)
 103
 104 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 105
 106 ##############################
 107 # Maze options
 108
 109 parser.add_argument("--maze_height", type=int, default=23)
 110
 111 parser.add_argument("--maze_width", type=int, default=39)
 112
 113 parser.add_argument("--maze_nb_walls", type=int, default=45)
 114
 115 ##############################
 116 # Snake options
 117
 118 parser.add_argument("--snake_height", type=int, default=6)
 119
 120 parser.add_argument("--snake_width", type=int, default=8)
 121
 122 parser.add_argument("--snake_nb_colors", type=int, default=5)
 123
 124 parser.add_argument("--snake_length", type=int, default=200)
 125
 126 ##############################
 127 # Stack options
 128
 129 parser.add_argument("--stack_nb_steps", type=int, default=100)
 130
 131 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 132
 133 parser.add_argument("--stack_nb_digits", type=int, default=3)
 134
 135 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 136
 137 ##############################
 138 # Expr options
 139
 140 parser.add_argument("--expr_nb_variables", type=int, default=5)
 141
 142 parser.add_argument("--expr_sequence_length", type=int, default=40)
 143
 144 parser.add_argument("--expr_operand_max", type=int, default=9)
 145
 146 parser.add_argument("--expr_result_max", type=int, default=99)
 147
 148 parser.add_argument("--expr_input_file", type=str, default=None)
 149
 150 ##############################
 151 # World options
 152
 153 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 154
 155 ######################################################################
 156
 157 args = parser.parse_args()
 158
 159 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 160
 161 if args.result_dir is None:
 162     args.result_dir = f"results_{args.task}"
 163
 164 ######################################################################
 165
 166 default_task_args = {
 167     "sandbox": {
 168         "nb_epochs": 50,
 169         "batch_size": 25,
 170         "nb_train_samples": 100000,
 171         "nb_test_samples": 10000,
 172     },
 173     "picoclvr": {
 174         "nb_epochs": 25,
 175         "batch_size": 25,
 176         "nb_train_samples": 250000,
 177         "nb_test_samples": 10000,
 178     },
 179     "mnist": {
 180         "nb_epochs": 25,
 181         "batch_size": 10,
 182         "nb_train_samples": 250000,
 183         "nb_test_samples": 10000,
 184     },
 185     "maze": {
 186         "nb_epochs": 25,
 187         "batch_size": 5,
 188         "nb_train_samples": 250000,
 189         "nb_test_samples": 10000,
 190     },
 191     "snake": {
 192         "nb_epochs": 5,
 193         "batch_size": 25,
 194         "nb_train_samples": 250000,
 195         "nb_test_samples": 10000,
 196     },
 197     "stack": {
 198         "nb_epochs": 5,
 199         "batch_size": 25,
 200         "nb_train_samples": 100000,
 201         "nb_test_samples": 1000,
 202     },
 203     "expr": {
 204         "nb_epochs": 40,
 205         "batch_size": 25,
 206         "nb_train_samples": 1000000,
 207         "nb_test_samples": 10000,
 208     },
 209     "rpl": {
 210         "nb_epochs": 40,
 211         "batch_size": 25,
 212         "nb_train_samples": 100000,
 213         "nb_test_samples": 10000,
 214     },
 215     "world": {
 216         "nb_epochs": 10,
 217         "batch_size": 25,
 218         "nb_train_samples": 25000,
 219         "nb_test_samples": 1000,
 220     },
 221 }
 222
 223 if args.task in default_task_args:
 224     for k, v in default_task_args[args.task].items():
 225         if getattr(args, k) is None:
 226             setattr(args, k, v)
 227
 228 ######################################################################
 229
 230 default_model_args = {
 231     "17K": {
 232         "dim_model": 32,
 233         "dim_keys": 32,
 234         "dim_hidden": 32,
 235         "nb_heads": 2,
 236         "nb_blocks": 2,
 237     },
 238     "37M": {
 239         "dim_model": 512,
 240         "dim_keys": 64,
 241         "dim_hidden": 2048,
 242         "nb_heads": 8,
 243         "nb_blocks": 12,
 244     },
 245     "122M": {
 246         "dim_model": 768,
 247         "dim_keys": 64,
 248         "dim_hidden": 2048,
 249         "nb_heads": 8,
 250         "nb_blocks": 24,
 251     },
 252     "352M": {
 253         "dim_model": 1024,
 254         "dim_keys": 64,
 255         "dim_hidden": 2048,
 256         "nb_heads": 8,
 257         "nb_blocks": 48,
 258     },
 259 }
 260
 261 if args.model in default_model_args:
 262     for k, v in default_model_args[args.model].items():
 263         if getattr(args, k) is None:
 264             setattr(args, k, v)
 265 else:
 266     raise ValueError(f"Unknown model {args.model}")
 267
 268 ######################################################################
 269
 270 try:
 271     os.mkdir(args.result_dir)
 272 except FileExistsError:
 273     if not args.overwrite_results:
 274         print(f"result directory {args.result_dir} already exists")
 275         exit(1)
 276
 277 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 278
 279 if args.seed >= 0:
 280     # torch.backends.cudnn.deterministic = True
 281     # torch.backends.cudnn.benchmark = False
 282     # torch.use_deterministic_algorithms(True)
 283     torch.manual_seed(args.seed)
 284     if torch.cuda.is_available():
 285         torch.cuda.manual_seed_all(args.seed)
 286
 287 ######################################################################
 288
 289
 290 def log_string(s):
 291     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 292
 293     if log_file is not None:
 294         log_file.write(t + s + "\n")
 295         log_file.flush()
 296
 297     print(t + s)
 298     sys.stdout.flush()
 299
 300
 301 for n in vars(args):
 302     log_string(f"args.{n} {getattr(args, n)}")
 303
 304
 305 ######################################################################
 306
 307
 308 def picoclvr_pruner_horizontal_green(p):
 309     return not ("green" in p and ("left" in p or "right" in p))
 310
 311
 312 picoclvr_pruner_train = (
 313     picoclvr_pruner_horizontal_green
 314     if args.picocvlr_prune_properties in {"train+eval"}
 315     else None
 316 )
 317
 318 picoclvr_pruner_eval = (
 319     (lambda p: not picoclvr_pruner_horizontal_green(p))
 320     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 321     else None
 322 )
 323
 324 ######################################################################
 325
 326 if args.task == "sandbox":
 327     if args.sandbox_level == 0:
 328         problem = tasks.ProblemLevel0(
 329             nb_sentences=args.sandbox_levels_nb_items,
 330             len_prompt=args.sandbox_levels_len_source,
 331             len_result=args.sandbox_levels_len_result,
 332         )
 333     elif args.sandbox_level == 1:
 334         problem = tasks.ProblemLevel1(
 335             nb_operators=args.sandbox_levels_nb_items,
 336             len_source=args.sandbox_levels_len_source,
 337             len_result=args.sandbox_levels_len_result,
 338         )
 339     elif args.sandbox_level == 2:
 340         problem = tasks.ProblemLevel2(
 341             len_source=args.sandbox_levels_len_source,
 342             len_result=args.sandbox_levels_len_result,
 343         )
 344     else:
 345         raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
 346
 347     task = tasks.SandBox(
 348         problem,
 349         # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
 350         nb_train_samples=args.nb_train_samples,
 351         nb_test_samples=args.nb_test_samples,
 352         batch_size=args.batch_size,
 353         logger=log_string,
 354         device=device,
 355     )
 356
 357 elif args.task == "picoclvr":
 358     task = tasks.PicoCLVR(
 359         nb_train_samples=args.nb_train_samples,
 360         nb_test_samples=args.nb_test_samples,
 361         batch_size=args.batch_size,
 362         height=args.picoclvr_height,
 363         width=args.picoclvr_width,
 364         nb_colors=args.picoclvr_nb_colors,
 365         logger=log_string,
 366         device=device,
 367         pruner_train=picoclvr_pruner_train,
 368         pruner_eval=picoclvr_pruner_eval,
 369     )
 370
 371 elif args.task == "mnist":
 372     task = tasks.MNIST(
 373         nb_train_samples=args.nb_train_samples,
 374         nb_test_samples=args.nb_test_samples,
 375         batch_size=args.batch_size,
 376         device=device,
 377     )
 378
 379 elif args.task == "maze":
 380     task = tasks.Maze(
 381         nb_train_samples=args.nb_train_samples,
 382         nb_test_samples=args.nb_test_samples,
 383         batch_size=args.batch_size,
 384         height=args.maze_height,
 385         width=args.maze_width,
 386         nb_walls=args.maze_nb_walls,
 387         device=device,
 388     )
 389
 390 elif args.task == "snake":
 391     task = tasks.Snake(
 392         nb_train_samples=args.nb_train_samples,
 393         nb_test_samples=args.nb_test_samples,
 394         batch_size=args.batch_size,
 395         height=args.snake_height,
 396         width=args.snake_width,
 397         nb_colors=args.snake_nb_colors,
 398         length=args.snake_length,
 399         prompt_length=args.snake_length // 2,
 400         device=device,
 401     )
 402
 403 elif args.task == "stack":
 404     task = tasks.Stack(
 405         nb_train_samples=args.nb_train_samples,
 406         nb_test_samples=args.nb_test_samples,
 407         batch_size=args.batch_size,
 408         logger=log_string,
 409         nb_steps=args.stack_nb_steps,
 410         nb_stacks=args.stack_nb_stacks,
 411         nb_digits=args.stack_nb_digits,
 412         fraction_values_for_train=args.stack_fraction_values_for_train,
 413         device=device,
 414     )
 415
 416 elif args.task == "expr":
 417     task = tasks.Expr(
 418         nb_train_samples=args.nb_train_samples,
 419         nb_test_samples=args.nb_test_samples,
 420         nb_variables=args.expr_nb_variables,
 421         sequence_length=args.expr_sequence_length,
 422         operand_max=args.expr_operand_max,
 423         result_max=args.expr_result_max,
 424         batch_size=args.batch_size,
 425         device=device,
 426     )
 427
 428 elif args.task == "rpl":
 429     task = tasks.RPL(
 430         nb_train_samples=args.nb_train_samples,
 431         nb_test_samples=args.nb_test_samples,
 432         batch_size=args.batch_size,
 433         device=device,
 434     )
 435
 436 elif args.task == "world":
 437     task = tasks.World(
 438         nb_train_samples=args.nb_train_samples,
 439         nb_test_samples=args.nb_test_samples,
 440         batch_size=args.batch_size,
 441         vqae_nb_epochs=args.world_vqae_nb_epochs,
 442         logger=log_string,
 443         device=device,
 444     )
 445
 446 else:
 447     raise ValueError(f"Unknown task {args.task}")
 448
 449 ######################################################################
 450
 451 log_string(f"device {device}")
 452
 453 vocabulary_size = task.vocabulary_size()
 454
 455 log_string(f"vocabulary_size {vocabulary_size}")
 456
 457 ##############################
 458
 459 model = mygpt.MyGPT(
 460     vocabulary_size=vocabulary_size,
 461     dim_model=args.dim_model,
 462     dim_keys=args.dim_keys,
 463     dim_hidden=args.dim_hidden,
 464     nb_heads=args.nb_heads,
 465     nb_blocks=args.nb_blocks,
 466     causal=True,
 467     dropout=args.dropout,
 468 )
 469
 470 model.to(device)
 471
 472 nb_parameters = sum(p.numel() for p in model.parameters())
 473 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 474
 475 ######################################################################
 476
 477 nb_epochs_finished = 0
 478
 479 if args.no_checkpoint:
 480     log_string(f"not trying to load checkpoint.")
 481
 482 else:
 483     try:
 484         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 485         checkpoint = torch.load(checkpoint_name)
 486         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 487         model.load_state_dict(checkpoint["model_state"])
 488         torch.set_rng_state(checkpoint["rng_state"])
 489         if torch.cuda.is_available():
 490             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 491
 492         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 493
 494     except FileNotFoundError:
 495         log_string("starting from scratch.")
 496
 497     except:
 498         log_string("error when loading the checkpoint.")
 499         exit(1)
 500
 501 ######################################################################
 502
 503 if args.task == "expr" and args.expr_input_file is not None:
 504     task.produce_results(
 505         nb_epochs_finished,
 506         model,
 507         args.result_dir,
 508         log_string,
 509         args.deterministic_synthesis,
 510         args.expr_input_file,
 511     )
 512
 513     exit(0)
 514
 515 ######################################################################
 516
 517 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 518
 519 # Compute the entropy of the training tokens
 520
 521 token_count = 0
 522 for input in task.batches(split="train"):
 523     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 524 token_probas = token_count / token_count.sum()
 525 entropy = -torch.xlogy(token_probas, token_probas).sum()
 526 train_set_perplexity = math.exp(entropy)
 527
 528 ##############################
 529
 530 # A bit of paranoia never hurts
 531
 532 train_examples = {}
 533
 534
 535 for input in task.batches(split="train"):
 536     assert input.dim() == 2 and input.dtype == torch.int64
 537     for x in input:
 538         train_examples[x.sum().item()] = x
 539
 540 nb_total, nb_collisions = 0, 0
 541 for input in task.batches(split="test"):
 542     assert input.dim() == 2 and input.dtype == torch.int64
 543     for x in input:
 544         nb_total += 1
 545         y = train_examples.get(x.sum().item())
 546         if y is not None:
 547             if x.size() == y.size() and (x - y).abs().sum() == 0:
 548                 nb_collisions += 1
 549
 550 del train_examples
 551
 552 log_string(
 553     f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
 554 )
 555
 556 ##############################
 557
 558 if args.learning_rate_schedule == "cos":
 559     learning_rate_schedule = {}
 560     for n_epoch in range(args.nb_epochs):
 561         u = n_epoch / args.nb_epochs * math.pi
 562         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 563 else:
 564     u = {
 565         int(k): float(v)
 566         for k, v in [
 567             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 568         ]
 569     }
 570
 571     learning_rate_schedule = {}
 572     learning_rate = args.learning_rate
 573     for n_epoch in range(args.nb_epochs):
 574         if n_epoch in u:
 575             learning_rate = u[n_epoch]
 576         learning_rate_schedule[n_epoch] = learning_rate
 577
 578 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 579
 580 ##############################
 581
 582 nb_samples_seen = 0
 583
 584 if nb_epochs_finished >= nb_epochs:
 585     task.produce_results(
 586         nb_epochs_finished,
 587         model,
 588         args.result_dir,
 589         log_string,
 590         args.deterministic_synthesis,
 591     )
 592
 593 for n_epoch in range(nb_epochs_finished, nb_epochs):
 594     learning_rate = learning_rate_schedule[n_epoch]
 595
 596     log_string(f"learning_rate {learning_rate}")
 597
 598     if args.optim == "sgd":
 599         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 600     elif args.optim == "adam":
 601         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 602     elif args.optim == "adamw":
 603         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 604     else:
 605         raise ValueError(f"Unknown optimizer {args.optim}.")
 606
 607     model.train()
 608
 609     nb_train_samples, acc_train_loss = 0, 0.0
 610
 611     for input in task.batches(split="train"):
 612         input = input.to(device)
 613         output = model(mygpt.BracketedSequence(input)).x
 614         loss = F.cross_entropy(output.transpose(1, 2), input)
 615         acc_train_loss += loss.item() * input.size(0)
 616         nb_train_samples += input.size(0)
 617         nb_samples_seen += input.size(0)
 618
 619         optimizer.zero_grad()
 620         loss.backward()
 621         optimizer.step()
 622
 623     with torch.autograd.no_grad():
 624         model.eval()
 625
 626         nb_test_samples, acc_test_loss = 0, 0.0
 627
 628         for input in task.batches(split="test"):
 629             input = input.to(device)
 630
 631             output = model(mygpt.BracketedSequence(input)).x
 632             loss = F.cross_entropy(output.transpose(1, 2), input)
 633             acc_test_loss += loss.item() * input.size(0)
 634             nb_test_samples += input.size(0)
 635
 636         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 637         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 638
 639         log_string(
 640             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 641         )
 642
 643         task.produce_results(
 644             n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
 645         )
 646
 647     checkpoint = {
 648         "nb_epochs_finished": n_epoch + 1,
 649         "model_state": model.state_dict(),
 650         "rng_state": torch.get_rng_state(),
 651     }
 652
 653     if torch.cuda.is_available():
 654         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 655
 656     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 657     torch.save(checkpoint, checkpoint_name)
 658     log_string(f"saved checkpoint {checkpoint_name}")
 659
 660 ######################################################################