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