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