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