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