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, tensorstack
  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=30)
 129
 130 ######################################################################
 131
 132 args = parser.parse_args()
 133
 134 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 135
 136 if args.result_dir is None:
 137     args.result_dir = f"results_{args.task}"
 138
 139 ######################################################################
 140
 141 default_args = {
 142     "picoclvr": {
 143         "nb_epochs": 25,
 144         "batch_size": 25,
 145         "nb_train_samples": 250000,
 146         "nb_test_samples": 10000,
 147     },
 148     "mnist": {
 149         "nb_epochs": 25,
 150         "batch_size": 10,
 151         "nb_train_samples": 250000,
 152         "nb_test_samples": 10000,
 153     },
 154     "maze": {
 155         "nb_epochs": 25,
 156         "batch_size": 5,
 157         "nb_train_samples": 250000,
 158         "nb_test_samples": 10000,
 159     },
 160     "snake": {
 161         "nb_epochs": 5,
 162         "batch_size": 25,
 163         "nb_train_samples": 250000,
 164         "nb_test_samples": 10000,
 165     },
 166     "stack": {
 167         "nb_epochs": 5,
 168         "batch_size": 25,
 169         "nb_train_samples": 100000,
 170         "nb_test_samples": 1000,
 171     },
 172     "expr": {
 173         "nb_epochs": 50,
 174         "batch_size": 25,
 175         "nb_train_samples": 250000,
 176         "nb_test_samples": 10000,
 177     },
 178 }
 179
 180 if args.task in default_args:
 181     for k, v in default_args[args.task].items():
 182         if getattr(args, k) is None:
 183             setattr(args, k, v)
 184
 185 ######################################################################
 186
 187 try:
 188     os.mkdir(args.result_dir)
 189 except FileExistsError:
 190     if not args.overwrite_results:
 191         print(f"result directory {args.result_dir} already exists")
 192         exit(1)
 193
 194 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 195
 196 if args.seed >= 0:
 197     # torch.backends.cudnn.deterministic = True
 198     # torch.backends.cudnn.benchmark = False
 199     # torch.use_deterministic_algorithms(True)
 200     torch.manual_seed(args.seed)
 201     if torch.cuda.is_available():
 202         torch.cuda.manual_seed_all(args.seed)
 203
 204 ######################################################################
 205
 206
 207 def log_string(s):
 208     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 209
 210     if log_file is not None:
 211         log_file.write(t + s + "\n")
 212         log_file.flush()
 213
 214     print(t + s)
 215     sys.stdout.flush()
 216
 217
 218 for n in vars(args):
 219     log_string(f"args.{n} {getattr(args, n)}")
 220
 221
 222 ######################################################################
 223
 224
 225 def picoclvr_pruner_horizontal_green(p):
 226     return not ("green" in p and ("left" in p or "right" in p))
 227
 228
 229 picoclvr_pruner_train = (
 230     picoclvr_pruner_horizontal_green
 231     if args.picocvlr_prune_properties in {"train+eval"}
 232     else None
 233 )
 234
 235 picoclvr_pruner_eval = (
 236     (lambda p: not picoclvr_pruner_horizontal_green(p))
 237     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 238     else None
 239 )
 240
 241 ######################################################################
 242
 243 if args.task == "picoclvr":
 244     task = tasks.PicoCLVR(
 245         nb_train_samples=args.nb_train_samples,
 246         nb_test_samples=args.nb_test_samples,
 247         batch_size=args.batch_size,
 248         height=args.picoclvr_height,
 249         width=args.picoclvr_width,
 250         nb_colors=args.picoclvr_nb_colors,
 251         logger=log_string,
 252         device=device,
 253         pruner_train=picoclvr_pruner_train,
 254         pruner_eval=picoclvr_pruner_eval,
 255     )
 256
 257 elif args.task == "mnist":
 258     task = tasks.MNIST(
 259         nb_train_samples=args.nb_train_samples,
 260         nb_test_samples=args.nb_test_samples,
 261         batch_size=args.batch_size,
 262         device=device,
 263     )
 264
 265 elif args.task == "maze":
 266     task = tasks.Maze(
 267         nb_train_samples=args.nb_train_samples,
 268         nb_test_samples=args.nb_test_samples,
 269         batch_size=args.batch_size,
 270         height=args.maze_height,
 271         width=args.maze_width,
 272         nb_walls=args.maze_nb_walls,
 273         device=device,
 274     )
 275
 276 elif args.task == "snake":
 277     task = tasks.Snake(
 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.snake_height,
 282         width=args.snake_width,
 283         nb_colors=args.snake_nb_colors,
 284         length=args.snake_length,
 285         prompt_length=args.snake_length // 2,
 286         device=device,
 287     )
 288
 289 elif args.task == "stack":
 290     task = tasks.Stack(
 291         nb_train_samples=args.nb_train_samples,
 292         nb_test_samples=args.nb_test_samples,
 293         batch_size=args.batch_size,
 294         logger=log_string,
 295         nb_steps=args.stack_nb_steps,
 296         nb_stacks=args.stack_nb_stacks,
 297         nb_digits=args.stack_nb_digits,
 298         fraction_values_for_train=args.stack_fraction_values_for_train,
 299         device=device,
 300     )
 301
 302 elif args.task == "expr":
 303     task = tasks.Expr(
 304         nb_train_samples=args.nb_train_samples,
 305         nb_test_samples=args.nb_test_samples,
 306         nb_variables=args.expr_nb_variables,
 307         sequence_length=args.expr_sequence_length,
 308         batch_size=args.batch_size,
 309         device=device,
 310     )
 311
 312 else:
 313     raise ValueError(f"Unknown task {args.task}")
 314
 315 ######################################################################
 316
 317 log_string(f"device {device}")
 318
 319 vocabulary_size = task.vocabulary_size()
 320
 321 log_string(f"vocabulary_size {vocabulary_size}")
 322
 323 ##############################
 324
 325 model = mygpt.MyGPT(
 326     vocabulary_size=vocabulary_size,
 327     dim_model=args.dim_model,
 328     dim_keys=args.dim_keys,
 329     dim_hidden=args.dim_hidden,
 330     nb_heads=args.nb_heads,
 331     nb_blocks=args.nb_blocks,
 332     causal=True,
 333     dropout=args.dropout,
 334 )
 335
 336 model.to(device)
 337
 338 nb_parameters = sum(p.numel() for p in model.parameters())
 339 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 340
 341 ######################################################################
 342
 343 nb_epochs_finished = 0
 344
 345 if args.no_checkpoint:
 346     log_string(f"not trying to load checkpoint.")
 347
 348 else:
 349     try:
 350         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 351         checkpoint = torch.load(checkpoint_name)
 352         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 353         model.load_state_dict(checkpoint["model_state"])
 354         torch.set_rng_state(checkpoint["rng_state"])
 355         if torch.cuda.is_available():
 356             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 357
 358         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 359
 360     except FileNotFoundError:
 361         log_string("starting from scratch.")
 362
 363     except:
 364         log_string("error when loading the checkpoint.")
 365         exit(1)
 366
 367 ######################################################################
 368
 369 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 370
 371 # Compute the entropy of the training tokens
 372
 373 token_count = 0
 374 for input in task.batches(split="train"):
 375     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 376 token_probas = token_count / token_count.sum()
 377 entropy = -torch.xlogy(token_probas, token_probas).sum()
 378 train_set_perplexity = math.exp(entropy)
 379
 380 ##############################
 381
 382 # A bit of paranoia never hurts
 383
 384 train_examples = {}
 385
 386 for input in task.batches(split="train"):
 387     assert input.dim()==2 and input.dtype==torch.int64
 388     for x in input:
 389         train_examples[x.sum().item()]=x
 390
 391 for input in task.batches(split="test"):
 392     assert input.dim()==2 and input.dtype==torch.int64
 393     for x in input:
 394         y = train_examples.get(x.sum().item())
 395         if y is not None:
 396             assert x.size() != y.size() or (x-y).abs().sum() > 0
 397
 398 del train_examples
 399
 400 ##############################
 401
 402 if args.learning_rate_schedule == "cos":
 403     learning_rate_schedule = {}
 404     for n_epoch in range(args.nb_epochs):
 405         u = n_epoch / args.nb_epochs * math.pi
 406         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 407 else:
 408     u = {
 409         int(k): float(v)
 410         for k, v in [
 411             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 412         ]
 413     }
 414
 415     learning_rate_schedule = {}
 416     learning_rate = args.learning_rate
 417     for n_epoch in range(args.nb_epochs):
 418         if n_epoch in u:
 419             learning_rate = u[n_epoch]
 420         learning_rate_schedule[n_epoch] = learning_rate
 421
 422 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 423
 424 ##############################
 425
 426 nb_samples_seen = 0
 427
 428 if nb_epochs_finished >= nb_epochs:
 429     task.produce_results(
 430         nb_epochs_finished,
 431         model,
 432         args.result_dir,
 433         log_string,
 434         args.deterministic_synthesis,
 435     )
 436
 437 for n_epoch in range(nb_epochs_finished, nb_epochs):
 438     learning_rate = learning_rate_schedule[n_epoch]
 439
 440     log_string(f"learning_rate {learning_rate}")
 441
 442     if args.optim == "sgd":
 443         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 444     elif args.optim == "adam":
 445         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 446     elif args.optim == "adamw":
 447         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 448     else:
 449         raise ValueError(f"Unknown optimizer {args.optim}.")
 450
 451     model.train()
 452
 453     nb_train_samples, acc_train_loss = 0, 0.0
 454
 455     for input in task.batches(split="train"):
 456         input = input.to(device)
 457         output = model(mygpt.BracketedSequence(input)).x
 458         loss = F.cross_entropy(output.transpose(1, 2), input)
 459         acc_train_loss += loss.item() * input.size(0)
 460         nb_train_samples += input.size(0)
 461         nb_samples_seen += input.size(0)
 462
 463         optimizer.zero_grad()
 464         loss.backward()
 465         optimizer.step()
 466
 467     with torch.autograd.no_grad():
 468         model.eval()
 469
 470         nb_test_samples, acc_test_loss = 0, 0.0
 471
 472         for input in task.batches(split="test"):
 473             input = input.to(device)
 474
 475             output = model(mygpt.BracketedSequence(input)).x
 476             loss = F.cross_entropy(output.transpose(1, 2), input)
 477             acc_test_loss += loss.item() * input.size(0)
 478             nb_test_samples += input.size(0)
 479
 480         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 481         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 482
 483         log_string(
 484             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 485         )
 486
 487         task.produce_results(
 488             n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
 489         )
 490
 491     checkpoint = {
 492         "nb_epochs_finished": n_epoch + 1,
 493         "model_state": model.state_dict(),
 494         "rng_state": torch.get_rng_state(),
 495     }
 496
 497     if torch.cuda.is_available():
 498         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 499
 500     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 501     torch.save(checkpoint, checkpoint_name)
 502     log_string(f"saved checkpoint {checkpoint_name}")
 503
 504 ######################################################################