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
 176     model = nn.Sequential(
 177         nn.Linear(args.dim_model, args.dim_model),
 178         nn.ReLU(),
 179         nn.Linear(args.dim_model, args.dim_model),
 180         nn.ReLU(),
 181         nn.Linear(args.dim_model, 4),
 182     ).to(device)
 183
 184     for n_epoch in range(args.nb_epochs):
 185         learning_rate = learning_rate_schedule[n_epoch]
 186         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 187
 188         acc_train_loss, nb_train_samples = 0, 0
 189         for input, targets in task.policy_batches(split="train"):
 190             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 191             output = model(output_gpt)
 192             targets = targets * (input.unsqueeze(-1) == maze.v_empty)
 193             output = output * (input.unsqueeze(-1) == maze.v_empty)
 194             # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
 195             loss = (
 196                 -(output.log_softmax(-1) * targets).sum()
 197                 / (input == maze.v_empty).sum()
 198             )
 199             acc_train_loss += loss.item() * input.size(0)
 200             nb_train_samples += input.size(0)
 201
 202             optimizer.zero_grad()
 203             loss.backward()
 204             optimizer.step()
 205
 206         acc_test_loss, nb_test_samples = 0, 0
 207         for input, targets in task.policy_batches(split="test"):
 208             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 209             output = model(output_gpt)
 210             targets = targets * (input.unsqueeze(-1) == maze.v_empty)
 211             output = output * (input.unsqueeze(-1) == maze.v_empty)
 212             # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
 213             loss = (
 214                 -(output.log_softmax(-1) * targets).sum()
 215                 / (input == maze.v_empty).sum()
 216             )
 217             acc_test_loss += loss.item() * input.size(0)
 218             nb_test_samples += input.size(0)
 219
 220         log_string(
 221             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 222         )
 223
 224         # -------------------
 225         input = task.test_input[:32, : task.height * task.width]
 226         targets = task.test_policies[:32]
 227         output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 228         output = model(output_gpt)
 229         # losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
 230         # losses = losses * (input == maze.v_empty)
 231         # losses = losses / losses.max()
 232         # losses = (output.softmax(-1) - targets).abs().max(-1).values
 233         # losses = (losses >= 0.05).float()
 234         losses = (
 235             (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 236         ).float()
 237         losses = losses.reshape(-1, args.maze_height, args.maze_width)
 238         input = input.reshape(-1, args.maze_height, args.maze_width)
 239         maze.save_image(
 240             os.path.join(args.result_dir, f"oneshot_{n_epoch:04d}.png"),
 241             mazes=input,
 242             score_paths=losses,
 243         )
 244         # -------------------
 245
 246     gpt.train(t)
 247
 248
 249 ######################################################################
 250
 251
 252 class Task:
 253     def batches(self, split="train"):
 254         pass
 255
 256     def vocabulary_size(self):
 257         pass
 258
 259     def produce_results(self, n_epoch, model):
 260         pass
 261
 262
 263 ######################################################################
 264
 265 import maze
 266
 267
 268 class TaskMaze(Task):
 269     def map2seq(self, *m):
 270         return torch.cat([x.flatten(1) for x in m], 1)
 271
 272     def seq2map(self, s):
 273         s = s.reshape(s.size(0), -1, self.height, self.width)
 274         return (s[:, k] for k in range(s.size(1)))
 275
 276     def __init__(
 277         self,
 278         nb_train_samples,
 279         nb_test_samples,
 280         batch_size,
 281         height,
 282         width,
 283         nb_walls,
 284         device=torch.device("cpu"),
 285     ):
 286         self.batch_size = batch_size
 287         self.height = height
 288         self.width = width
 289         self.device = device
 290
 291         train_mazes, train_paths, train_policies = maze.create_maze_data(
 292             nb_train_samples,
 293             height=height,
 294             width=width,
 295             nb_walls=nb_walls,
 296             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 297         )
 298         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 299         self.train_policies = train_policies.flatten(-2).permute(0, 2, 1).to(device)
 300
 301         test_mazes, test_paths, test_policies = maze.create_maze_data(
 302             nb_test_samples,
 303             height=height,
 304             width=width,
 305             nb_walls=nb_walls,
 306             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 307         )
 308         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 309         self.test_policies = test_policies.flatten(-2).permute(0, 2, 1).to(device)
 310
 311         self.nb_codes = self.train_input.max() + 1
 312
 313     def batches(self, split="train", nb_to_use=-1):
 314         assert split in {"train", "test"}
 315         input = self.train_input if split == "train" else self.test_input
 316         if nb_to_use > 0:
 317             input = input[:nb_to_use]
 318         for batch in tqdm.tqdm(
 319             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 320         ):
 321             yield batch
 322
 323     def policy_batches(self, split="train", nb_to_use=-1):
 324         assert split in {"train", "test"}
 325         input = self.train_input if split == "train" else self.test_input
 326         targets = self.train_policies if split == "train" else self.test_policies
 327         input = input[:, : self.height * self.width]
 328         targets = targets * (input != maze.v_wall)[:, :, None]
 329
 330         if nb_to_use > 0:
 331             input = input[:nb_to_use]
 332             targets = targets[:nb_to_use]
 333
 334         for batch in tqdm.tqdm(
 335             zip(input.split(self.batch_size), targets.split(self.batch_size)),
 336             dynamic_ncols=True,
 337             desc=f"epoch-{split}",
 338         ):
 339             yield batch
 340
 341     def vocabulary_size(self):
 342         return self.nb_codes
 343
 344     def compute_error(self, model, split="train", nb_to_use=-1):
 345         nb_total, nb_correct = 0, 0
 346         for input in task.batches(split, nb_to_use):
 347             result = input.clone()
 348             ar_mask = result.new_zeros(result.size())
 349             ar_mask[:, self.height * self.width :] = 1
 350             result *= 1 - ar_mask
 351             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 352             mazes, paths = self.seq2map(result)
 353             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 354             nb_total += mazes.size(0)
 355
 356         return nb_total, nb_correct
 357
 358     def produce_results(self, n_epoch, model):
 359         with torch.autograd.no_grad():
 360             t = model.training
 361             model.eval()
 362
 363             train_nb_total, train_nb_correct = self.compute_error(
 364                 model, "train", nb_to_use=1000
 365             )
 366             log_string(
 367                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 368             )
 369
 370             test_nb_total, test_nb_correct = self.compute_error(
 371                 model, "test", nb_to_use=1000
 372             )
 373             log_string(
 374                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 375             )
 376
 377             input = self.test_input[:32]
 378             result = input.clone()
 379             ar_mask = result.new_zeros(result.size())
 380             ar_mask[:, self.height * self.width :] = 1
 381             result *= 1 - ar_mask
 382             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 383
 384             mazes, paths = self.seq2map(input)
 385             _, predicted_paths = self.seq2map(result)
 386             maze.save_image(
 387                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 388                 mazes=mazes,
 389                 target_paths=paths,
 390                 predicted_paths=predicted_paths,
 391                 path_correct=maze.path_correctness(mazes, predicted_paths),
 392             )
 393
 394             model.train(t)
 395
 396
 397 ######################################################################
 398
 399 log_string(f"device {device}")
 400
 401
 402 task = TaskMaze(
 403     nb_train_samples=args.nb_train_samples,
 404     nb_test_samples=args.nb_test_samples,
 405     batch_size=args.batch_size,
 406     height=args.maze_height,
 407     width=args.maze_width,
 408     nb_walls=args.maze_nb_walls,
 409     device=device,
 410 )
 411
 412
 413 vocabulary_size = task.vocabulary_size()
 414
 415 log_string(f"vocabulary_size {vocabulary_size}")
 416
 417 ##############################
 418
 419 model = mygpt.MyGPT(
 420     vocabulary_size=vocabulary_size,
 421     dim_model=args.dim_model,
 422     dim_keys=args.dim_keys,
 423     dim_hidden=args.dim_hidden,
 424     nb_heads=args.nb_heads,
 425     nb_blocks=args.nb_blocks,
 426     causal=True,
 427     dropout=args.dropout,
 428 )
 429
 430 model.to(device)
 431
 432 nb_parameters = sum(p.numel() for p in model.parameters())
 433 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 434
 435 ######################################################################
 436
 437 nb_epochs_finished = 0
 438
 439 if args.no_checkpoint:
 440     log_string(f"not trying to load checkpoint.")
 441
 442 else:
 443     try:
 444         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 445         checkpoint = torch.load(checkpoint_name)
 446         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 447         model.load_state_dict(checkpoint["model_state"])
 448         torch.set_rng_state(checkpoint["rng_state"])
 449         if torch.cuda.is_available():
 450             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 451
 452         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 453
 454     except FileNotFoundError:
 455         log_string("starting from scratch.")
 456
 457     except:
 458         log_string("error when loading the checkpoint.")
 459         exit(1)
 460
 461 ######################################################################
 462
 463 token_count = 0
 464 for input in task.batches(split="train"):
 465     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 466 token_probas = token_count / token_count.sum()
 467 entropy = -torch.xlogy(token_probas, token_probas).sum()
 468 train_set_perplexity = math.exp(entropy)
 469
 470 ##############################
 471
 472 if args.learning_rate_schedule == "cos":
 473     learning_rate_schedule = {}
 474     for n_epoch in range(args.nb_epochs):
 475         u = n_epoch / args.nb_epochs * math.pi
 476         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 477 else:
 478     u = {
 479         int(k): float(v)
 480         for k, v in [
 481             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 482         ]
 483     }
 484
 485     learning_rate_schedule = {}
 486     learning_rate = args.learning_rate
 487     for n_epoch in range(args.nb_epochs):
 488         if n_epoch in u:
 489             learning_rate = u[n_epoch]
 490         learning_rate_schedule[n_epoch] = learning_rate
 491
 492 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 493
 494 ##############################
 495
 496 if args.one_shot:
 497     one_shot(model, task)
 498     exit(0)
 499
 500 ##############################
 501
 502 if nb_epochs_finished >= args.nb_epochs:
 503     n_epoch = nb_epochs_finished
 504     train_perplexity = compute_perplexity(model, split="train")
 505     test_perplexity = compute_perplexity(model, split="test")
 506
 507     log_string(
 508         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 509     )
 510
 511     task.produce_results(n_epoch, model)
 512
 513     exit(0)
 514
 515 ##############################
 516
 517 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 518     learning_rate = learning_rate_schedule[n_epoch]
 519
 520     log_string(f"learning_rate {learning_rate}")
 521
 522     if args.optim == "sgd":
 523         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 524     elif args.optim == "adam":
 525         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 526     elif args.optim == "adamw":
 527         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 528     else:
 529         raise ValueError(f"Unknown optimizer {args.optim}.")
 530
 531     model.train()
 532
 533     nb_train_samples, acc_train_loss = 0, 0.0
 534
 535     for input in task.batches(split="train"):
 536         input = input.to(device)
 537         output = model(mygpt.BracketedSequence(input)).x
 538         loss = F.cross_entropy(output.transpose(1, 2), input)
 539         acc_train_loss += loss.item() * input.size(0)
 540         nb_train_samples += input.size(0)
 541
 542         optimizer.zero_grad()
 543         loss.backward()
 544         optimizer.step()
 545
 546     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 547     test_perplexity = compute_perplexity(model, split="test")
 548
 549     log_string(
 550         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 551     )
 552
 553     task.produce_results(n_epoch, model)
 554
 555     checkpoint = {
 556         "nb_epochs_finished": n_epoch + 1,
 557         "model_state": model.state_dict(),
 558         "rng_state": torch.get_rng_state(),
 559     }
 560
 561     if torch.cuda.is_available():
 562         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 563
 564     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 565     torch.save(checkpoint, checkpoint_name)
 566     log_string(f"saved checkpoint {checkpoint_name}")
 567
 568 ######################################################################