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