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