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     t = gpt.training
 174     gpt.eval()
 175     for input, targets in task.policy_batches():
 176         output = gpt(mygpt.BracketedSequence(input), with_readout = False).x
 177     gpt.train(t)
 178
 179 ######################################################################
 180
 181
 182 class Task:
 183     def batches(self, split="train"):
 184         pass
 185
 186     def vocabulary_size(self):
 187         pass
 188
 189     def produce_results(self, n_epoch, model):
 190         pass
 191
 192
 193 ######################################################################
 194
 195 import maze
 196
 197
 198 class TaskMaze(Task):
 199     def map2seq(self, *m):
 200         return torch.cat([x.flatten(1) for x in m], 1)
 201
 202     def seq2map(self, s):
 203         s = s.reshape(s.size(0), -1, self.height, self.width)
 204         return (s[:, k] for k in range(s.size(1)))
 205
 206     def __init__(
 207         self,
 208         nb_train_samples,
 209         nb_test_samples,
 210         batch_size,
 211         height,
 212         width,
 213         nb_walls,
 214         device=torch.device("cpu"),
 215     ):
 216         self.batch_size = batch_size
 217         self.height = height
 218         self.width = width
 219         self.device = device
 220
 221         train_mazes, train_paths, train_policies = maze.create_maze_data(
 222             nb_train_samples,
 223             height=height,
 224             width=width,
 225             nb_walls=nb_walls,
 226             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 227         )
 228         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 229         self.train_policies = train_policies.to(device)
 230
 231         test_mazes, test_paths, test_policies = maze.create_maze_data(
 232             nb_test_samples,
 233             height=height,
 234             width=width,
 235             nb_walls=nb_walls,
 236             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 237         )
 238         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 239         self.test_policies = test_policies.to(device)
 240
 241         self.nb_codes = self.train_input.max() + 1
 242
 243     def batches(self, split="train", nb_to_use=-1):
 244         assert split in {"train", "test"}
 245         input = self.train_input if split == "train" else self.test_input
 246         if nb_to_use > 0:
 247             input = input[:nb_to_use]
 248         for batch in tqdm.tqdm(
 249             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 250         ):
 251             yield batch
 252
 253     def policy_batches(self, split="train", nb_to_use=-1):
 254         assert split in {"train", "test"}
 255         input = self.train_input if split == "train" else self.test_input
 256         targets = self.train_policies if split == "train" else self.test_policies
 257         input = input[:, : self.height * self.width]
 258         targets = targets.flatten(-2) * (input != maze.v_wall)[:,None]
 259
 260         if nb_to_use > 0:
 261             input = input[:nb_to_use]
 262             targets = targets[:nb_to_use]
 263
 264         for batch in tqdm.tqdm(
 265             zip(input.split(self.batch_size), targets.split(self.batch_size)),
 266             dynamic_ncols=True,
 267             desc=f"epoch-{split}",
 268         ):
 269             yield batch
 270
 271     def vocabulary_size(self):
 272         return self.nb_codes
 273
 274     def compute_error(self, model, split="train", nb_to_use=-1):
 275         nb_total, nb_correct = 0, 0
 276         for input in task.batches(split, nb_to_use):
 277             result = input.clone()
 278             ar_mask = result.new_zeros(result.size())
 279             ar_mask[:, self.height * self.width :] = 1
 280             result *= 1 - ar_mask
 281             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 282             mazes, paths = self.seq2map(result)
 283             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 284             nb_total += mazes.size(0)
 285
 286         return nb_total, nb_correct
 287
 288     def produce_results(self, n_epoch, model):
 289         with torch.autograd.no_grad():
 290             t = model.training
 291             model.eval()
 292
 293             train_nb_total, train_nb_correct = self.compute_error(
 294                 model, "train", nb_to_use=1000
 295             )
 296             log_string(
 297                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 298             )
 299
 300             test_nb_total, test_nb_correct = self.compute_error(
 301                 model, "test", nb_to_use=1000
 302             )
 303             log_string(
 304                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 305             )
 306
 307             input = self.test_input[:32]
 308             result = input.clone()
 309             ar_mask = result.new_zeros(result.size())
 310             ar_mask[:, self.height * self.width :] = 1
 311             result *= 1 - ar_mask
 312             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 313
 314             mazes, paths = self.seq2map(input)
 315             _, predicted_paths = self.seq2map(result)
 316             maze.save_image(
 317                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 318                 mazes,
 319                 paths,
 320                 predicted_paths,
 321                 maze.path_correctness(mazes, predicted_paths),
 322             )
 323
 324             model.train(t)
 325
 326
 327 ######################################################################
 328
 329 log_string(f"device {device}")
 330
 331
 332 task = TaskMaze(
 333     nb_train_samples=args.nb_train_samples,
 334     nb_test_samples=args.nb_test_samples,
 335     batch_size=args.batch_size,
 336     height=args.maze_height,
 337     width=args.maze_width,
 338     nb_walls=args.maze_nb_walls,
 339     device=device,
 340 )
 341
 342
 343 vocabulary_size = task.vocabulary_size()
 344
 345 log_string(f"vocabulary_size {vocabulary_size}")
 346
 347 ##############################
 348
 349 model = mygpt.MyGPT(
 350     vocabulary_size=vocabulary_size,
 351     dim_model=args.dim_model,
 352     dim_keys=args.dim_keys,
 353     dim_hidden=args.dim_hidden,
 354     nb_heads=args.nb_heads,
 355     nb_blocks=args.nb_blocks,
 356     causal=True,
 357     dropout=args.dropout,
 358 )
 359
 360 model.to(device)
 361
 362 nb_parameters = sum(p.numel() for p in model.parameters())
 363 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 364
 365 ######################################################################
 366
 367 nb_epochs_finished = 0
 368
 369 if args.no_checkpoint:
 370     log_string(f"not trying to load checkpoint.")
 371
 372 else:
 373     try:
 374         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 375         checkpoint = torch.load(checkpoint_name)
 376         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 377         model.load_state_dict(checkpoint["model_state"])
 378         torch.set_rng_state(checkpoint["rng_state"])
 379         if torch.cuda.is_available():
 380             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 381
 382         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 383
 384     except FileNotFoundError:
 385         log_string("starting from scratch.")
 386
 387     except:
 388         log_string("error when loading the checkpoint.")
 389         exit(1)
 390
 391 ######################################################################
 392
 393 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 394
 395 token_count = 0
 396 for input in task.batches(split="train"):
 397     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 398 token_probas = token_count / token_count.sum()
 399 entropy = -torch.xlogy(token_probas, token_probas).sum()
 400 train_set_perplexity = math.exp(entropy)
 401
 402 ##############################
 403
 404 if args.learning_rate_schedule == "cos":
 405     learning_rate_schedule = {}
 406     for n_epoch in range(args.nb_epochs):
 407         u = n_epoch / args.nb_epochs * math.pi
 408         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 409 else:
 410     u = {
 411         int(k): float(v)
 412         for k, v in [
 413             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 414         ]
 415     }
 416
 417     learning_rate_schedule = {}
 418     learning_rate = args.learning_rate
 419     for n_epoch in range(args.nb_epochs):
 420         if n_epoch in u:
 421             learning_rate = u[n_epoch]
 422         learning_rate_schedule[n_epoch] = learning_rate
 423
 424 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 425
 426 ##############################
 427
 428 if args.one_shot:
 429     one_shot(model, task)
 430     exit(0)
 431
 432 ##############################
 433
 434 if nb_epochs_finished >= nb_epochs:
 435     n_epoch = nb_epochs_finished
 436     train_perplexity = compute_perplexity(model, split="train")
 437     test_perplexity = compute_perplexity(model, split="test")
 438
 439     log_string(
 440         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 441     )
 442
 443     task.produce_results(n_epoch, model)
 444
 445     exit(0)
 446
 447 ##############################
 448
 449 for n_epoch in range(nb_epochs_finished, nb_epochs):
 450     learning_rate = learning_rate_schedule[n_epoch]
 451
 452     log_string(f"learning_rate {learning_rate}")
 453
 454     if args.optim == "sgd":
 455         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 456     elif args.optim == "adam":
 457         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 458     elif args.optim == "adamw":
 459         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 460     else:
 461         raise ValueError(f"Unknown optimizer {args.optim}.")
 462
 463     model.train()
 464
 465     nb_train_samples, acc_train_loss = 0, 0.0
 466
 467     for input in task.batches(split="train"):
 468         input = input.to(device)
 469         output = model(mygpt.BracketedSequence(input)).x
 470         loss = F.cross_entropy(output.transpose(1, 2), input)
 471         acc_train_loss += loss.item() * input.size(0)
 472         nb_train_samples += input.size(0)
 473
 474         optimizer.zero_grad()
 475         loss.backward()
 476         optimizer.step()
 477
 478     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 479     test_perplexity = compute_perplexity(model, split="test")
 480
 481     log_string(
 482         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 483     )
 484
 485     task.produce_results(n_epoch, model)
 486
 487     checkpoint = {
 488         "nb_epochs_finished": n_epoch + 1,
 489         "model_state": model.state_dict(),
 490         "rng_state": torch.get_rng_state(),
 491     }
 492
 493     if torch.cuda.is_available():
 494         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 495
 496     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 497     torch.save(checkpoint, checkpoint_name)
 498     log_string(f"saved checkpoint {checkpoint_name}")
 499
 500 ######################################################################