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("--random_regression_order", 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 # one-shot prediction
  86
  87 parser.add_argument("--oneshot", action="store_true", default=False)
  88
  89 parser.add_argument("--oneshot_input", type=str, default="head")
  90
  91 parser.add_argument("--oneshot_output", type=str, default="trace")
  92
  93 ######################################################################
  94
  95 args = parser.parse_args()
  96
  97 try:
  98     os.mkdir(args.result_dir)
  99 except FileExistsError:
 100     if not args.overwrite_results:
 101         print(f"result directory {args.result_dir} already exists")
 102         exit(1)
 103
 104 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 105
 106 if args.seed >= 0:
 107     # torch.backends.cudnn.deterministic = True
 108     # torch.backends.cudnn.benchmark = False
 109     # torch.use_deterministic_algorithms(True)
 110     torch.manual_seed(args.seed)
 111     if torch.cuda.is_available():
 112         torch.cuda.manual_seed_all(args.seed)
 113
 114 ######################################################################
 115
 116
 117 def log_string(s):
 118     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 119
 120     if log_file is not None:
 121         log_file.write(t + s + "\n")
 122         log_file.flush()
 123
 124     print(t + s)
 125     sys.stdout.flush()
 126
 127
 128 for n in vars(args):
 129     log_string(f"args.{n} {getattr(args, n)}")
 130
 131 ######################################################################
 132
 133
 134 def generation_order(x, fixed_len):
 135     if args.random_regression_order:
 136         order = torch.rand(x.size(), device=x.device)
 137         order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=x.device)
 138         order = order.sort(1).indices
 139     else:
 140         order = (
 141             torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
 142         )
 143     return order
 144
 145
 146 def reorder(x, order, back=False):  # x is NxTxD1x...xDk, order is NxT'
 147     u = x.reshape(x.size()[:2] + (-1,))
 148     order = order.unsqueeze(-1).expand(-1, -1, u.size(-1))
 149     if back:
 150         v = u.new(u.size())
 151         v.scatter_(1, order, u)
 152     else:
 153         v = u.gather(1, order)
 154     v = v.reshape(v.size()[:2] + x.size()[2:])
 155     return v
 156
 157
 158 def shuffle(x, fixed_len):
 159     order = generation_order(x, fixed_len)
 160     return reorder(x, order), order
 161
 162
 163 ######################################################################
 164
 165 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 166 # tokens that should be generated
 167
 168
 169 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
 170     for input, ar_mask, order in zip(
 171         input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size)
 172     ):
 173         i = (ar_mask.sum(0) > 0).nonzero()
 174         if i.min() > 0:
 175             # Needed to initialize the model's cache
 176             model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
 177         for s in range(i.min(), i.max() + 1):
 178             output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
 179             logits = output[:, s]
 180             if args.deterministic_synthesis:
 181                 t_next = logits.argmax(1)
 182             else:
 183                 dist = torch.distributions.categorical.Categorical(logits=logits)
 184                 t_next = dist.sample()
 185             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 186
 187
 188 ######################################################################
 189
 190
 191 def compute_perplexity(model, task, fixed_len, split="train"):
 192     with torch.autograd.no_grad():
 193         t = model.training
 194         model.eval()
 195
 196         nb_samples, acc_loss = 0, 0.0
 197
 198         for input in task.batches(split=split):
 199             input = input.to(device)
 200             x, order = shuffle(input, fixed_len)
 201             x = model(mygpt.BracketedSequence(x), order=order).x
 202             output = reorder(x, order, back=True)
 203             loss = F.cross_entropy(output.transpose(1, 2), input)
 204             acc_loss += loss.item() * input.size(0)
 205             nb_samples += input.size(0)
 206
 207         model.train(t)
 208
 209         return math.exp(min(100, acc_loss / nb_samples))
 210
 211
 212 ######################################################################
 213
 214
 215 def oneshot_policy_loss(mazes, output, policies, height, width):
 216     masks = (mazes == maze.v_empty).unsqueeze(-1)
 217     targets = policies.permute(0, 2, 1) * masks
 218     output = output * masks
 219     return -(output.log_softmax(-1) * targets).sum() / masks.sum()
 220
 221
 222 def oneshot_trace_loss(mazes, output, policies, height, width):
 223     masks = mazes == maze.v_empty
 224     targets = maze.stationary_densities(
 225         mazes.view(-1, height, width), policies.view(-1, 4, height, width)
 226     ).flatten(-2)
 227     targets = targets * masks
 228     output = output.squeeze(-1) * masks
 229     return (output - targets).abs().sum() / masks.sum()
 230
 231
 232 def oneshot(gpt, task):
 233     t = gpt.training
 234     gpt.eval()
 235
 236     if args.oneshot_input == "head":
 237         dim_in = args.dim_model
 238     elif args.oneshot_input == "deep":
 239         dim_in = args.dim_model * args.nb_blocks * 2
 240     else:
 241         raise ValueError(f"{args.oneshot_input=}")
 242
 243     if args.oneshot_output == "policy":
 244         dim_out = 4
 245         compute_loss = oneshot_policy_loss
 246     elif args.oneshot_output == "trace":
 247         dim_out = 1
 248         compute_loss = oneshot_trace_loss
 249     else:
 250         raise ValueError(f"{args.oneshot_output=}")
 251
 252     model = nn.Sequential(
 253         nn.Linear(dim_in, args.dim_model),
 254         nn.ReLU(),
 255         nn.Linear(args.dim_model, args.dim_model),
 256         nn.ReLU(),
 257         nn.Linear(args.dim_model, dim_out),
 258     ).to(device)
 259
 260     for n_epoch in range(args.nb_epochs):
 261         learning_rate = learning_rate_schedule[n_epoch]
 262         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 263
 264         acc_train_loss, nb_train_samples = 0, 0
 265         for mazes, policies in task.policy_batches(split="train"):
 266             x, order = shuffle(mazes, task.height * task.width)
 267             x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 268             output_gpt = reorder(x, order, back=True)
 269             output = model(output_gpt)
 270
 271             loss = compute_loss(mazes, output, policies, task.height, task.width)
 272             acc_train_loss += loss.item() * mazes.size(0)
 273             nb_train_samples += mazes.size(0)
 274
 275             optimizer.zero_grad()
 276             loss.backward()
 277             optimizer.step()
 278
 279         acc_test_loss, nb_test_samples = 0, 0
 280         for mazes, policies in task.policy_batches(split="test"):
 281             x, order = shuffle(mazes, task.height * task.width)
 282             x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 283             output_gpt = reorder(x, order, back=True)
 284             output = model(output_gpt)
 285             loss = compute_loss(mazes, output, policies, task.height, task.width)
 286             acc_test_loss += loss.item() * mazes.size(0)
 287             nb_test_samples += mazes.size(0)
 288
 289         log_string(
 290             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 291         )
 292
 293         # -------------------
 294         mazes = task.test_input[:32, : task.height * task.width]
 295         policies = task.test_policies[:32]
 296         x, order = shuffle(mazes, task.height * task.width)
 297         x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 298         output_gpt = reorder(x, order, back=True)
 299         output = model(output_gpt)
 300         if args.oneshot_output == "policy":
 301             targets = policies.permute(0, 2, 1)
 302             scores = (
 303                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 304             ).float()
 305         elif args.oneshot_output == "trace":
 306             targets = maze.stationary_densities(
 307                 mazes.view(-1, task.height, task.width),
 308                 policies.view(-1, 4, task.height, task.width),
 309             ).flatten(-2)
 310             scores = output
 311         else:
 312             raise ValueError(f"{args.oneshot_output=}")
 313
 314         scores = scores.reshape(-1, task.height, task.width)
 315         mazes = mazes.reshape(-1, task.height, task.width)
 316         targets = targets.reshape(-1, task.height, task.width)
 317         filename = (
 318             f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
 319         )
 320         maze.save_image(
 321             os.path.join(args.result_dir, filename),
 322             mazes=mazes,
 323             score_paths=scores,
 324             score_truth=targets,
 325         )
 326         log_string(f"wrote {filename}")
 327
 328         # -------------------
 329
 330     gpt.train(t)
 331
 332
 333 ######################################################################
 334
 335
 336 class Task:
 337     def batches(self, split="train", nb_to_use=-1, desc=None):
 338         pass
 339
 340     def vocabulary_size(self):
 341         pass
 342
 343     def produce_results(self, n_epoch, model):
 344         pass
 345
 346
 347 ######################################################################
 348
 349 import maze
 350
 351
 352 class TaskMaze(Task):
 353     def map2seq(self, *m):
 354         return torch.cat([x.flatten(1) for x in m], 1)
 355
 356     def seq2map(self, s):
 357         s = s.reshape(s.size(0), -1, self.height, self.width)
 358         return (s[:, k] for k in range(s.size(1)))
 359
 360     def __init__(
 361         self,
 362         nb_train_samples,
 363         nb_test_samples,
 364         batch_size,
 365         height,
 366         width,
 367         nb_walls,
 368         device=torch.device("cpu"),
 369     ):
 370         self.batch_size = batch_size
 371         self.height = height
 372         self.width = width
 373         self.device = device
 374
 375         train_mazes, train_paths, train_policies = maze.create_maze_data(
 376             nb_train_samples,
 377             height=height,
 378             width=width,
 379             nb_walls=nb_walls,
 380             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 381         )
 382         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 383         self.train_policies = train_policies.flatten(-2).to(device)
 384
 385         test_mazes, test_paths, test_policies = maze.create_maze_data(
 386             nb_test_samples,
 387             height=height,
 388             width=width,
 389             nb_walls=nb_walls,
 390             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 391         )
 392         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 393         self.test_policies = test_policies.flatten(-2).to(device)
 394
 395         self.nb_codes = self.train_input.max() + 1
 396
 397     def batches(self, split="train", nb_to_use=-1, desc=None):
 398         assert split in {"train", "test"}
 399         input = self.train_input if split == "train" else self.test_input
 400         if nb_to_use > 0:
 401             input = input[:nb_to_use]
 402         if desc is None:
 403             desc = f"epoch-{split}"
 404         for batch in tqdm.tqdm(
 405             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 406         ):
 407             yield batch
 408
 409     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 410         assert split in {"train", "test"}
 411         input = self.train_input if split == "train" else self.test_input
 412         policies = self.train_policies if split == "train" else self.test_policies
 413         input = input[:, : self.height * self.width]
 414         policies = policies * (input != maze.v_wall)[:, None]
 415
 416         if nb_to_use > 0:
 417             input = input[:nb_to_use]
 418             policies = policies[:nb_to_use]
 419
 420         if desc is None:
 421             desc = f"epoch-{split}"
 422         for batch in tqdm.tqdm(
 423             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 424             dynamic_ncols=True,
 425             desc=desc,
 426         ):
 427             yield batch
 428
 429     def vocabulary_size(self):
 430         return self.nb_codes
 431
 432     def compute_error(self, model, split="train", nb_to_use=-1):
 433         nb_total, nb_correct = 0, 0
 434         for input in task.batches(split, nb_to_use):
 435             result = input.clone()
 436             ar_mask = result.new_zeros(result.size())
 437             ar_mask[:, self.height * self.width :] = 1
 438             result *= 1 - ar_mask
 439             x, order = shuffle(result, self.height * self.width)
 440             masked_inplace_autoregression(
 441                 model, self.batch_size, x, ar_mask, order=order
 442             )
 443             result = reorder(x, order, back=True)
 444             mazes, paths = self.seq2map(result)
 445             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 446             nb_total += mazes.size(0)
 447
 448         return nb_total, nb_correct
 449
 450     def produce_results(self, n_epoch, model):
 451         with torch.autograd.no_grad():
 452             t = model.training
 453             model.eval()
 454
 455             train_nb_total, train_nb_correct = self.compute_error(
 456                 model, "train", nb_to_use=1000
 457             )
 458             log_string(
 459                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 460             )
 461
 462             test_nb_total, test_nb_correct = self.compute_error(
 463                 model, "test", nb_to_use=1000
 464             )
 465             log_string(
 466                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 467             )
 468
 469             input = self.test_input[:32]
 470             result = input.clone()
 471             ar_mask = result.new_zeros(result.size())
 472             ar_mask[:, self.height * self.width :] = 1
 473             result *= 1 - ar_mask
 474             x, order = shuffle(result, self.height * self.width)
 475             masked_inplace_autoregression(
 476                 model, self.batch_size, x, ar_mask, order=order
 477             )
 478             result = reorder(x, order, back=True)
 479
 480             mazes, paths = self.seq2map(input)
 481             _, predicted_paths = self.seq2map(result)
 482             filename = f"result_{n_epoch:04d}.png"
 483             maze.save_image(
 484                 os.path.join(args.result_dir, filename),
 485                 mazes=mazes,
 486                 target_paths=paths,
 487                 predicted_paths=predicted_paths,
 488                 path_correct=maze.path_correctness(mazes, predicted_paths),
 489             )
 490             log_string(f"wrote {filename}")
 491
 492             model.train(t)
 493
 494
 495 ######################################################################
 496
 497 log_string(f"device {device}")
 498
 499
 500 task = TaskMaze(
 501     nb_train_samples=args.nb_train_samples,
 502     nb_test_samples=args.nb_test_samples,
 503     batch_size=args.batch_size,
 504     height=args.maze_height,
 505     width=args.maze_width,
 506     nb_walls=args.maze_nb_walls,
 507     device=device,
 508 )
 509
 510
 511 vocabulary_size = task.vocabulary_size()
 512
 513 log_string(f"vocabulary_size {vocabulary_size}")
 514
 515 ##############################
 516
 517 model = mygpt.MyGPT(
 518     vocabulary_size=vocabulary_size,
 519     dim_model=args.dim_model,
 520     dim_keys=args.dim_keys,
 521     dim_hidden=args.dim_hidden,
 522     nb_heads=args.nb_heads,
 523     nb_blocks=args.nb_blocks,
 524     causal=True,
 525     dropout=args.dropout,
 526 )
 527
 528 model.to(device)
 529
 530 nb_parameters = sum(p.numel() for p in model.parameters())
 531 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 532
 533 ######################################################################
 534
 535 nb_epochs_finished = 0
 536
 537 if args.no_checkpoint:
 538     log_string(f"not trying to load checkpoint.")
 539
 540 else:
 541     try:
 542         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 543         checkpoint = torch.load(checkpoint_name)
 544         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 545         model.load_state_dict(checkpoint["model_state"])
 546         torch.set_rng_state(checkpoint["rng_state"])
 547         if torch.cuda.is_available():
 548             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 549
 550         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 551
 552     except FileNotFoundError:
 553         log_string("starting from scratch.")
 554
 555     except:
 556         log_string("error when loading the checkpoint.")
 557         exit(1)
 558
 559 ######################################################################
 560
 561 token_count = 0
 562 for input in task.batches(split="train"):
 563     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 564 token_probas = token_count / token_count.sum()
 565 entropy = -torch.xlogy(token_probas, token_probas).sum()
 566 train_set_perplexity = math.exp(entropy)
 567
 568 ##############################
 569
 570 if args.learning_rate_schedule == "cos":
 571     learning_rate_schedule = {}
 572     for n_epoch in range(args.nb_epochs):
 573         u = n_epoch / args.nb_epochs * math.pi
 574         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 575 else:
 576     u = {
 577         int(k): float(v)
 578         for k, v in [
 579             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 580         ]
 581     }
 582
 583     learning_rate_schedule = {}
 584     learning_rate = args.learning_rate
 585     for n_epoch in range(args.nb_epochs):
 586         if n_epoch in u:
 587             learning_rate = u[n_epoch]
 588         learning_rate_schedule[n_epoch] = learning_rate
 589
 590 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 591
 592 ##############################
 593
 594 if nb_epochs_finished >= args.nb_epochs:
 595     n_epoch = nb_epochs_finished
 596     train_perplexity = compute_perplexity(
 597         model, task, fixed_len=task.height * task.width, split="train"
 598     )
 599     test_perplexity = compute_perplexity(
 600         model, task, fixed_len=task.height * task.width, split="test"
 601     )
 602
 603     log_string(
 604         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 605     )
 606
 607     task.produce_results(n_epoch, model)
 608
 609 ##############################
 610
 611 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 612     learning_rate = learning_rate_schedule[n_epoch]
 613
 614     log_string(f"learning_rate {learning_rate}")
 615
 616     if args.optim == "sgd":
 617         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 618     elif args.optim == "adam":
 619         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 620     elif args.optim == "adamw":
 621         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 622     else:
 623         raise ValueError(f"{args.optim=}")
 624
 625     model.train()
 626
 627     nb_train_samples, acc_train_loss = 0, 0.0
 628
 629     for input in task.batches(split="train"):
 630         input = input.to(device)
 631         x, order = shuffle(input, task.height * task.width)
 632         x = model(mygpt.BracketedSequence(x), order=order).x
 633         output = reorder(x, order, back=True)
 634         loss = F.cross_entropy(output.transpose(1, 2), input)
 635         acc_train_loss += loss.item() * input.size(0)
 636         nb_train_samples += input.size(0)
 637
 638         optimizer.zero_grad()
 639         loss.backward()
 640         optimizer.step()
 641
 642     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 643     test_perplexity = compute_perplexity(
 644         model, task, fixed_len=task.height * task.width, split="test"
 645     )
 646
 647     log_string(
 648         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 649     )
 650
 651     task.produce_results(n_epoch, model)
 652
 653     checkpoint = {
 654         "nb_epochs_finished": n_epoch + 1,
 655         "model_state": model.state_dict(),
 656         "rng_state": torch.get_rng_state(),
 657     }
 658
 659     if torch.cuda.is_available():
 660         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 661
 662     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 663     torch.save(checkpoint, checkpoint_name)
 664     log_string(f"saved checkpoint {checkpoint_name}")
 665
 666 ######################################################################
 667
 668 if args.oneshot:
 669     oneshot(model, task)
 670
 671 ######################################################################