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