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