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