beaver.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, itertools, os
  12
  13 import torch, torchvision
  14 from torch import nn
  15 from torch.nn import functional as F
  16
  17 import mygpt, 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(description="A maze shortest path solving with a GPT.")
  30
  31 parser.add_argument("--log_filename", type=str, default="train.log")
  32
  33 parser.add_argument("--result_dir", type=str, default="results_default")
  34
  35 parser.add_argument("--seed", type=int, default=0)
  36
  37 parser.add_argument("--nb_epochs", type=int, default=25)
  38
  39 parser.add_argument("--nb_train_samples", type=int, default=200000)
  40
  41 parser.add_argument("--nb_test_samples", type=int, default=50000)
  42
  43 parser.add_argument("--batch_size", type=int, default=25)
  44
  45 parser.add_argument("--optim", type=str, default="adam")
  46
  47 parser.add_argument("--learning_rate", type=float, default=1e-3)
  48
  49 parser.add_argument(
  50     "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
  51 )
  52
  53 parser.add_argument("--dim_model", type=int, default=512)
  54
  55 parser.add_argument("--dim_keys", type=int, default=64)
  56
  57 parser.add_argument("--dim_hidden", type=int, default=2048)
  58
  59 parser.add_argument("--nb_heads", type=int, default=8)
  60
  61 parser.add_argument("--nb_blocks", type=int, default=12)
  62
  63 parser.add_argument("--dropout", type=float, default=0.1)
  64
  65 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  66
  67 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  68
  69 parser.add_argument("--overwrite_results", action="store_true", default=False)
  70
  71 parser.add_argument("--one_shot", action="store_true", default=False)
  72
  73 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  74
  75 ##############################
  76 # maze options
  77
  78 parser.add_argument("--maze_height", type=int, default=13)
  79
  80 parser.add_argument("--maze_width", type=int, default=21)
  81
  82 parser.add_argument("--maze_nb_walls", type=int, default=15)
  83
  84 ######################################################################
  85
  86 args = parser.parse_args()
  87
  88 try:
  89     os.mkdir(args.result_dir)
  90 except FileExistsError:
  91     if not args.overwrite_results:
  92         print(f"result directory {args.result_dir} already exists")
  93         exit(1)
  94
  95 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
  96
  97 if args.seed >= 0:
  98     # torch.backends.cudnn.deterministic = True
  99     # torch.backends.cudnn.benchmark = False
 100     # torch.use_deterministic_algorithms(True)
 101     torch.manual_seed(args.seed)
 102     if torch.cuda.is_available():
 103         torch.cuda.manual_seed_all(args.seed)
 104
 105 ######################################################################
 106
 107
 108 def log_string(s):
 109     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 110
 111     if log_file is not None:
 112         log_file.write(t + s + "\n")
 113         log_file.flush()
 114
 115     print(t + s)
 116     sys.stdout.flush()
 117
 118
 119 for n in vars(args):
 120     log_string(f"args.{n} {getattr(args, n)}")
 121
 122 ######################################################################
 123
 124
 125 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 126 # tokens that should be generated
 127
 128
 129 def masked_inplace_autoregression(model, batch_size, input, ar_mask):
 130     for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
 131         i = (ar_mask.sum(0) > 0).nonzero()
 132         if i.min() > 0:
 133             # Needed to initialize the model's cache
 134             model(mygpt.BracketedSequence(input, 0, i.min()))
 135         for s in range(i.min(), i.max() + 1):
 136             output = model(mygpt.BracketedSequence(input, s, 1)).x
 137             logits = output[:, s]
 138             if args.deterministic_synthesis:
 139                 t_next = logits.argmax(1)
 140             else:
 141                 dist = torch.distributions.categorical.Categorical(logits=logits)
 142                 t_next = dist.sample()
 143             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 144
 145
 146 ######################################################################
 147
 148
 149 def compute_perplexity(model, split="train"):
 150     with torch.autograd.no_grad():
 151         t = model.training
 152         model.eval()
 153
 154         nb_samples, acc_loss = 0, 0.0
 155
 156         for input in task.batches(split=split):
 157             input = input.to(device)
 158
 159             output = model(mygpt.BracketedSequence(input)).x
 160             loss = F.cross_entropy(output.transpose(1, 2), input)
 161             acc_loss += loss.item() * input.size(0)
 162             nb_samples += input.size(0)
 163
 164         model.train(t)
 165
 166         return math.exp(min(100, acc_loss / nb_samples))
 167
 168
 169 ######################################################################
 170
 171
 172 def one_shot(gpt, task):
 173     pass
 174
 175
 176 ######################################################################
 177
 178
 179 class Task:
 180     def batches(self, split="train"):
 181         pass
 182
 183     def vocabulary_size(self):
 184         pass
 185
 186     def produce_results(self, n_epoch, model):
 187         pass
 188
 189
 190 ######################################################################
 191
 192 import maze
 193
 194
 195 class TaskMaze(Task):
 196     def map2seq(self, *m):
 197         return torch.cat([x.flatten(1) for x in m], 1)
 198
 199     def seq2map(self, s):
 200         s = s.reshape(s.size(0), -1, self.height, self.width)
 201         return (s[:, k] for k in range(s.size(1)))
 202
 203     def __init__(
 204         self,
 205         nb_train_samples,
 206         nb_test_samples,
 207         batch_size,
 208         height,
 209         width,
 210         nb_walls,
 211         device=torch.device("cpu"),
 212     ):
 213         self.batch_size = batch_size
 214         self.height = height
 215         self.width = width
 216         self.device = device
 217
 218         mazes_train, paths_train = maze.create_maze_data(
 219             nb_train_samples,
 220             height=height,
 221             width=width,
 222             nb_walls=nb_walls,
 223             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 224         )
 225         mazes_train, paths_train = mazes_train.to(device), paths_train.to(device)
 226         self.train_input = self.map2seq(mazes_train, paths_train)
 227
 228         mazes_test, paths_test = maze.create_maze_data(
 229             nb_test_samples,
 230             height=height,
 231             width=width,
 232             nb_walls=nb_walls,
 233             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 234         )
 235         mazes_test, paths_test = mazes_test.to(device), paths_test.to(device)
 236         self.test_input = self.map2seq(mazes_test, paths_test)
 237
 238         self.nb_codes = self.train_input.max() + 1
 239
 240     def batches(self, split="train", nb_to_use=-1):
 241         assert split in {"train", "test"}
 242         input = self.train_input if split == "train" else self.test_input
 243         if nb_to_use > 0:
 244             input = input[:nb_to_use]
 245         for batch in tqdm.tqdm(
 246             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 247         ):
 248             yield batch
 249
 250     def vocabulary_size(self):
 251         return self.nb_codes
 252
 253     def compute_error(self, model, split="train", nb_to_use=-1):
 254         nb_total, nb_correct = 0, 0
 255         for input in task.batches(split, nb_to_use):
 256             result = input.clone()
 257             ar_mask = result.new_zeros(result.size())
 258             ar_mask[:, self.height * self.width :] = 1
 259             result *= 1 - ar_mask
 260             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 261             mazes, paths = self.seq2map(result)
 262             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 263             nb_total += mazes.size(0)
 264
 265         return nb_total, nb_correct
 266
 267     def produce_results(self, n_epoch, model):
 268         with torch.autograd.no_grad():
 269             t = model.training
 270             model.eval()
 271
 272             train_nb_total, train_nb_correct = self.compute_error(
 273                 model, "train", nb_to_use=1000
 274             )
 275             log_string(
 276                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 277             )
 278
 279             test_nb_total, test_nb_correct = self.compute_error(
 280                 model, "test", nb_to_use=1000
 281             )
 282             log_string(
 283                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 284             )
 285
 286             input = self.test_input[:32]
 287             result = input.clone()
 288             ar_mask = result.new_zeros(result.size())
 289             ar_mask[:, self.height * self.width :] = 1
 290             result *= 1 - ar_mask
 291             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 292
 293             mazes, paths = self.seq2map(input)
 294             _, predicted_paths = self.seq2map(result)
 295             maze.save_image(
 296                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 297                 mazes,
 298                 paths,
 299                 predicted_paths,
 300                 maze.path_correctness(mazes, predicted_paths),
 301             )
 302
 303             model.train(t)
 304
 305
 306 ######################################################################
 307
 308 log_string(f"device {device}")
 309
 310
 311 task = TaskMaze(
 312     nb_train_samples=args.nb_train_samples,
 313     nb_test_samples=args.nb_test_samples,
 314     batch_size=args.batch_size,
 315     height=args.maze_height,
 316     width=args.maze_width,
 317     nb_walls=args.maze_nb_walls,
 318     device=device,
 319 )
 320
 321
 322 vocabulary_size = task.vocabulary_size()
 323
 324 log_string(f"vocabulary_size {vocabulary_size}")
 325
 326 ##############################
 327
 328 model = mygpt.MyGPT(
 329     vocabulary_size=vocabulary_size,
 330     dim_model=args.dim_model,
 331     dim_keys=args.dim_keys,
 332     dim_hidden=args.dim_hidden,
 333     nb_heads=args.nb_heads,
 334     nb_blocks=args.nb_blocks,
 335     causal=True,
 336     dropout=args.dropout,
 337 )
 338
 339 model.to(device)
 340
 341 nb_parameters = sum(p.numel() for p in model.parameters())
 342 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 343
 344 ######################################################################
 345
 346 nb_epochs_finished = 0
 347
 348 if args.no_checkpoint:
 349     log_string(f"not trying to load checkpoint.")
 350
 351 else:
 352     try:
 353         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 354         checkpoint = torch.load(checkpoint_name)
 355         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 356         model.load_state_dict(checkpoint["model_state"])
 357         torch.set_rng_state(checkpoint["rng_state"])
 358         if torch.cuda.is_available():
 359             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 360
 361         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 362
 363     except FileNotFoundError:
 364         log_string("starting from scratch.")
 365
 366     except:
 367         log_string("error when loading the checkpoint.")
 368         exit(1)
 369
 370 ######################################################################
 371
 372 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 373
 374 token_count = 0
 375 for input in task.batches(split="train"):
 376     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 377 token_probas = token_count / token_count.sum()
 378 entropy = -torch.xlogy(token_probas, token_probas).sum()
 379 train_set_perplexity = math.exp(entropy)
 380
 381 ##############################
 382
 383 if args.learning_rate_schedule == "cos":
 384     learning_rate_schedule = {}
 385     for n_epoch in range(args.nb_epochs):
 386         u = n_epoch / args.nb_epochs * math.pi
 387         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 388 else:
 389     u = {
 390         int(k): float(v)
 391         for k, v in [
 392             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 393         ]
 394     }
 395
 396     learning_rate_schedule = {}
 397     learning_rate = args.learning_rate
 398     for n_epoch in range(args.nb_epochs):
 399         if n_epoch in u:
 400             learning_rate = u[n_epoch]
 401         learning_rate_schedule[n_epoch] = learning_rate
 402
 403 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 404
 405 ##############################
 406
 407 if args.one_shot:
 408     one_shot(model, task)
 409     exit(0)
 410
 411 ##############################
 412
 413 if nb_epochs_finished >= nb_epochs:
 414     n_epoch = nb_epochs_finished
 415     train_perplexity = compute_perplexity(model, split="train")
 416     test_perplexity = compute_perplexity(model, split="test")
 417
 418     log_string(
 419         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 420     )
 421
 422     task.produce_results(n_epoch, model)
 423
 424     exit(0)
 425
 426 ##############################
 427
 428 for n_epoch in range(nb_epochs_finished, nb_epochs):
 429     learning_rate = learning_rate_schedule[n_epoch]
 430
 431     log_string(f"learning_rate {learning_rate}")
 432
 433     if args.optim == "sgd":
 434         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 435     elif args.optim == "adam":
 436         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 437     elif args.optim == "adamw":
 438         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 439     else:
 440         raise ValueError(f"Unknown optimizer {args.optim}.")
 441
 442     model.train()
 443
 444     nb_train_samples, acc_train_loss = 0, 0.0
 445
 446     for input in task.batches(split="train"):
 447         input = input.to(device)
 448         output = model(mygpt.BracketedSequence(input)).x
 449         loss = F.cross_entropy(output.transpose(1, 2), input)
 450         acc_train_loss += loss.item() * input.size(0)
 451         nb_train_samples += input.size(0)
 452
 453         optimizer.zero_grad()
 454         loss.backward()
 455         optimizer.step()
 456
 457     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 458     test_perplexity = compute_perplexity(model, split="test")
 459
 460     log_string(
 461         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 462     )
 463
 464     task.produce_results(n_epoch, model)
 465
 466     checkpoint = {
 467         "nb_epochs_finished": n_epoch + 1,
 468         "model_state": model.state_dict(),
 469         "rng_state": torch.get_rng_state(),
 470     }
 471
 472     if torch.cuda.is_available():
 473         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 474
 475     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 476     torch.save(checkpoint, checkpoint_name)
 477     log_string(f"saved checkpoint {checkpoint_name}")
 478
 479 ######################################################################