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, 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
 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     "world": {
 185         "nb_epochs": 5,
 186         "batch_size": 25,
 187         "nb_train_samples": 10000,
 188         "nb_test_samples": 1000,
 189     },
 190 }
 191
 192 if args.task in default_args:
 193     for k, v in default_args[args.task].items():
 194         if getattr(args, k) is None:
 195             setattr(args, k, v)
 196
 197 ######################################################################
 198
 199 try:
 200     os.mkdir(args.result_dir)
 201 except FileExistsError:
 202     if not args.overwrite_results:
 203         print(f"result directory {args.result_dir} already exists")
 204         exit(1)
 205
 206 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 207
 208 if args.seed >= 0:
 209     # torch.backends.cudnn.deterministic = True
 210     # torch.backends.cudnn.benchmark = False
 211     # torch.use_deterministic_algorithms(True)
 212     torch.manual_seed(args.seed)
 213     if torch.cuda.is_available():
 214         torch.cuda.manual_seed_all(args.seed)
 215
 216 ######################################################################
 217
 218
 219 def log_string(s):
 220     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 221
 222     if log_file is not None:
 223         log_file.write(t + s + "\n")
 224         log_file.flush()
 225
 226     print(t + s)
 227     sys.stdout.flush()
 228
 229
 230 for n in vars(args):
 231     log_string(f"args.{n} {getattr(args, n)}")
 232
 233
 234 ######################################################################
 235
 236
 237 def picoclvr_pruner_horizontal_green(p):
 238     return not ("green" in p and ("left" in p or "right" in p))
 239
 240
 241 picoclvr_pruner_train = (
 242     picoclvr_pruner_horizontal_green
 243     if args.picocvlr_prune_properties in {"train+eval"}
 244     else None
 245 )
 246
 247 picoclvr_pruner_eval = (
 248     (lambda p: not picoclvr_pruner_horizontal_green(p))
 249     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 250     else None
 251 )
 252
 253 ######################################################################
 254
 255 if args.task == "picoclvr":
 256     task = tasks.PicoCLVR(
 257         nb_train_samples=args.nb_train_samples,
 258         nb_test_samples=args.nb_test_samples,
 259         batch_size=args.batch_size,
 260         height=args.picoclvr_height,
 261         width=args.picoclvr_width,
 262         nb_colors=args.picoclvr_nb_colors,
 263         logger=log_string,
 264         device=device,
 265         pruner_train=picoclvr_pruner_train,
 266         pruner_eval=picoclvr_pruner_eval,
 267     )
 268
 269 elif args.task == "mnist":
 270     task = tasks.MNIST(
 271         nb_train_samples=args.nb_train_samples,
 272         nb_test_samples=args.nb_test_samples,
 273         batch_size=args.batch_size,
 274         device=device,
 275     )
 276
 277 elif args.task == "maze":
 278     task = tasks.Maze(
 279         nb_train_samples=args.nb_train_samples,
 280         nb_test_samples=args.nb_test_samples,
 281         batch_size=args.batch_size,
 282         height=args.maze_height,
 283         width=args.maze_width,
 284         nb_walls=args.maze_nb_walls,
 285         device=device,
 286     )
 287
 288 elif args.task == "snake":
 289     task = tasks.Snake(
 290         nb_train_samples=args.nb_train_samples,
 291         nb_test_samples=args.nb_test_samples,
 292         batch_size=args.batch_size,
 293         height=args.snake_height,
 294         width=args.snake_width,
 295         nb_colors=args.snake_nb_colors,
 296         length=args.snake_length,
 297         prompt_length=args.snake_length // 2,
 298         device=device,
 299     )
 300
 301 elif args.task == "stack":
 302     task = tasks.Stack(
 303         nb_train_samples=args.nb_train_samples,
 304         nb_test_samples=args.nb_test_samples,
 305         batch_size=args.batch_size,
 306         logger=log_string,
 307         nb_steps=args.stack_nb_steps,
 308         nb_stacks=args.stack_nb_stacks,
 309         nb_digits=args.stack_nb_digits,
 310         fraction_values_for_train=args.stack_fraction_values_for_train,
 311         device=device,
 312     )
 313
 314 elif args.task == "expr":
 315     task = tasks.Expr(
 316         nb_train_samples=args.nb_train_samples,
 317         nb_test_samples=args.nb_test_samples,
 318         nb_variables=args.expr_nb_variables,
 319         sequence_length=args.expr_sequence_length,
 320         operand_max=args.expr_operand_max,
 321         result_max=args.expr_result_max,
 322         batch_size=args.batch_size,
 323         device=device,
 324     )
 325
 326 elif args.task == "world":
 327     task = tasks.World(
 328         nb_train_samples=args.nb_train_samples,
 329         nb_test_samples=args.nb_test_samples,
 330         batch_size=args.batch_size,
 331         device=device,
 332     )
 333
 334 else:
 335     raise ValueError(f"Unknown task {args.task}")
 336
 337 ######################################################################
 338
 339 log_string(f"device {device}")
 340
 341 vocabulary_size = task.vocabulary_size()
 342
 343 log_string(f"vocabulary_size {vocabulary_size}")
 344
 345 ##############################
 346
 347 model = mygpt.MyGPT(
 348     vocabulary_size=vocabulary_size,
 349     dim_model=args.dim_model,
 350     dim_keys=args.dim_keys,
 351     dim_hidden=args.dim_hidden,
 352     nb_heads=args.nb_heads,
 353     nb_blocks=args.nb_blocks,
 354     causal=True,
 355     dropout=args.dropout,
 356 )
 357
 358 model.to(device)
 359
 360 nb_parameters = sum(p.numel() for p in model.parameters())
 361 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 362
 363 ######################################################################
 364
 365 nb_epochs_finished = 0
 366
 367 if args.no_checkpoint:
 368     log_string(f"not trying to load checkpoint.")
 369
 370 else:
 371     try:
 372         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 373         checkpoint = torch.load(checkpoint_name)
 374         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 375         model.load_state_dict(checkpoint["model_state"])
 376         torch.set_rng_state(checkpoint["rng_state"])
 377         if torch.cuda.is_available():
 378             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 379
 380         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 381
 382     except FileNotFoundError:
 383         log_string("starting from scratch.")
 384
 385     except:
 386         log_string("error when loading the checkpoint.")
 387         exit(1)
 388
 389 ######################################################################
 390
 391 if args.task == "expr" and args.expr_input_file is not None:
 392     task.produce_results(
 393         nb_epochs_finished,
 394         model,
 395         args.result_dir,
 396         log_string,
 397         args.deterministic_synthesis,
 398         args.expr_input_file,
 399     )
 400
 401     exit(0)
 402
 403 ######################################################################
 404
 405 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 406
 407 # Compute the entropy of the training tokens
 408
 409 token_count = 0
 410 for input in task.batches(split="train"):
 411     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 412 token_probas = token_count / token_count.sum()
 413 entropy = -torch.xlogy(token_probas, token_probas).sum()
 414 train_set_perplexity = math.exp(entropy)
 415
 416 ##############################
 417
 418 # A bit of paranoia never hurts
 419
 420 train_examples = {}
 421
 422
 423 for input in task.batches(split="train"):
 424     assert input.dim() == 2 and input.dtype == torch.int64
 425     for x in input:
 426         train_examples[x.sum().item()] = x
 427
 428 nb_total, nb_collisions = 0, 0
 429 for input in task.batches(split="test"):
 430     assert input.dim() == 2 and input.dtype == torch.int64
 431     for x in input:
 432         nb_total += 1
 433         y = train_examples.get(x.sum().item())
 434         if y is not None:
 435             if x.size() == y.size() and (x - y).abs().sum() == 0:
 436                 nb_collisions += 1
 437
 438 del train_examples
 439
 440 log_string(
 441     f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
 442 )
 443
 444 ##############################
 445
 446 if args.learning_rate_schedule == "cos":
 447     learning_rate_schedule = {}
 448     for n_epoch in range(args.nb_epochs):
 449         u = n_epoch / args.nb_epochs * math.pi
 450         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 451 else:
 452     u = {
 453         int(k): float(v)
 454         for k, v in [
 455             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 456         ]
 457     }
 458
 459     learning_rate_schedule = {}
 460     learning_rate = args.learning_rate
 461     for n_epoch in range(args.nb_epochs):
 462         if n_epoch in u:
 463             learning_rate = u[n_epoch]
 464         learning_rate_schedule[n_epoch] = learning_rate
 465
 466 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 467
 468 ##############################
 469
 470 nb_samples_seen = 0
 471
 472 if nb_epochs_finished >= nb_epochs:
 473     task.produce_results(
 474         nb_epochs_finished,
 475         model,
 476         args.result_dir,
 477         log_string,
 478         args.deterministic_synthesis,
 479     )
 480
 481 for n_epoch in range(nb_epochs_finished, nb_epochs):
 482     learning_rate = learning_rate_schedule[n_epoch]
 483
 484     log_string(f"learning_rate {learning_rate}")
 485
 486     if args.optim == "sgd":
 487         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 488     elif args.optim == "adam":
 489         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 490     elif args.optim == "adamw":
 491         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 492     else:
 493         raise ValueError(f"Unknown optimizer {args.optim}.")
 494
 495     model.train()
 496
 497     nb_train_samples, acc_train_loss = 0, 0.0
 498
 499     for input in task.batches(split="train"):
 500         input = input.to(device)
 501         output = model(mygpt.BracketedSequence(input)).x
 502         loss = F.cross_entropy(output.transpose(1, 2), input)
 503         acc_train_loss += loss.item() * input.size(0)
 504         nb_train_samples += input.size(0)
 505         nb_samples_seen += input.size(0)
 506
 507         optimizer.zero_grad()
 508         loss.backward()
 509         optimizer.step()
 510
 511     with torch.autograd.no_grad():
 512         model.eval()
 513
 514         nb_test_samples, acc_test_loss = 0, 0.0
 515
 516         for input in task.batches(split="test"):
 517             input = input.to(device)
 518
 519             output = model(mygpt.BracketedSequence(input)).x
 520             loss = F.cross_entropy(output.transpose(1, 2), input)
 521             acc_test_loss += loss.item() * input.size(0)
 522             nb_test_samples += input.size(0)
 523
 524         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 525         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 526
 527         log_string(
 528             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 529         )
 530
 531         task.produce_results(
 532             n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
 533         )
 534
 535     checkpoint = {
 536         "nb_epochs_finished": n_epoch + 1,
 537         "model_state": model.state_dict(),
 538         "rng_state": torch.get_rng_state(),
 539     }
 540
 541     if torch.cuda.is_available():
 542         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 543
 544     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 545     torch.save(checkpoint, checkpoint_name)
 546     log_string(f"saved checkpoint {checkpoint_name}")
 547
 548 ######################################################################