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