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 random_order(result, fixed_len):
 135     if args.random_regression_order:
 136         order = torch.rand(result.size(), device=result.device)
 137         order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=order.device)
 138         return order.sort(1).indices
 139     else:
 140         return torch.arange(result.size(1)).unsqueeze(0).expand(result.size(0), -1)
 141
 142
 143 def shuffle(x, order, reorder=False):
 144     if x.dim() == 3:
 145         order = order.unsqueeze(-1).expand(-1, -1, x.size(-1))
 146     if reorder:
 147         y = x.new(x.size())
 148         y.scatter_(1, order, x)
 149         return y
 150     else:
 151         return x.gather(1, order)
 152
 153
 154 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 155 # tokens that should be generated
 156
 157
 158 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
 159     for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
 160         i = (ar_mask.sum(0) > 0).nonzero()
 161         if i.min() > 0:
 162             # Needed to initialize the model's cache
 163             model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
 164         for s in range(i.min(), i.max() + 1):
 165             output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
 166             logits = output[:, s]
 167             if args.deterministic_synthesis:
 168                 t_next = logits.argmax(1)
 169             else:
 170                 dist = torch.distributions.categorical.Categorical(logits=logits)
 171                 t_next = dist.sample()
 172             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 173
 174
 175 ######################################################################
 176
 177
 178 def compute_perplexity(model, split="train"):
 179     with torch.autograd.no_grad():
 180         t = model.training
 181         model.eval()
 182
 183         nb_samples, acc_loss = 0, 0.0
 184
 185         for input in task.batches(split=split):
 186             input = input.to(device)
 187             order = random_order(input, task.height * task.width)
 188             input = shuffle(input, order)
 189             output = model(mygpt.BracketedSequence(input), order=order).x
 190             loss = F.cross_entropy(output.transpose(1, 2), input)
 191             acc_loss += loss.item() * input.size(0)
 192             nb_samples += input.size(0)
 193
 194         model.train(t)
 195
 196         return math.exp(min(100, acc_loss / nb_samples))
 197
 198
 199 ######################################################################
 200
 201
 202 def oneshot_policy_loss(mazes, output, policies, height, width):
 203     masks = (mazes == maze.v_empty).unsqueeze(-1)
 204     targets = policies.permute(0, 2, 1) * masks
 205     output = output * masks
 206     return -(output.log_softmax(-1) * targets).sum() / masks.sum()
 207
 208
 209 def oneshot_trace_loss(mazes, output, policies, height, width):
 210     masks = mazes == maze.v_empty
 211     targets = maze.stationary_densities(
 212         mazes.view(-1, height, width), policies.view(-1, 4, height, width)
 213     ).flatten(-2)
 214     targets = targets * masks
 215     output = output.squeeze(-1) * masks
 216     return (output - targets).abs().sum() / masks.sum()
 217
 218
 219 def oneshot(gpt, task):
 220     t = gpt.training
 221     gpt.eval()
 222
 223     if args.oneshot_input == "head":
 224         dim_in = args.dim_model
 225     elif args.oneshot_input == "deep":
 226         dim_in = args.dim_model * args.nb_blocks * 2
 227     else:
 228         raise ValueError(f"{args.oneshot_input=}")
 229
 230     if args.oneshot_output == "policy":
 231         dim_out = 4
 232         compute_loss = oneshot_policy_loss
 233     elif args.oneshot_output == "trace":
 234         dim_out = 1
 235         compute_loss = oneshot_trace_loss
 236     else:
 237         raise ValueError(f"{args.oneshot_output=}")
 238
 239     model = nn.Sequential(
 240         nn.Linear(dim_in, args.dim_model),
 241         nn.ReLU(),
 242         nn.Linear(args.dim_model, args.dim_model),
 243         nn.ReLU(),
 244         nn.Linear(args.dim_model, dim_out),
 245     ).to(device)
 246
 247     for n_epoch in range(args.nb_epochs):
 248         learning_rate = learning_rate_schedule[n_epoch]
 249         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 250
 251         acc_train_loss, nb_train_samples = 0, 0
 252         for mazes, policies in task.policy_batches(split="train"):
 253             order = random_order(mazes, task.height * task.width)
 254             x = shuffle(mazes, order)
 255             x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 256             output_gpt = shuffle(x, order, reorder=True)
 257             output = model(output_gpt)
 258
 259             loss = compute_loss(mazes, output, policies, task.height, task.width)
 260             acc_train_loss += loss.item() * mazes.size(0)
 261             nb_train_samples += mazes.size(0)
 262
 263             optimizer.zero_grad()
 264             loss.backward()
 265             optimizer.step()
 266
 267         acc_test_loss, nb_test_samples = 0, 0
 268         for mazes, policies in task.policy_batches(split="test"):
 269             order = random_order(mazes, task.height * task.width)
 270             x = shuffle(mazes, order)
 271             x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 272             output_gpt = shuffle(x, order, reorder=True)
 273             output = model(output_gpt)
 274             loss = compute_loss(mazes, output, policies, task.height, task.width)
 275             acc_test_loss += loss.item() * mazes.size(0)
 276             nb_test_samples += mazes.size(0)
 277
 278         log_string(
 279             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 280         )
 281
 282         # -------------------
 283         mazes = task.test_input[:32, : task.height * task.width]
 284         policies = task.test_policies[:32]
 285         order = random_order(mazes, task.height * task.width)
 286         x = shuffle(mazes, order)
 287         x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 288         output_gpt = shuffle(x, order, reorder=True)
 289         output = model(output_gpt)
 290         if args.oneshot_output == "policy":
 291             targets = policies.permute(0, 2, 1)
 292             scores = (
 293                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 294             ).float()
 295         elif args.oneshot_output == "trace":
 296             targets = maze.stationary_densities(
 297                 mazes.view(-1, task.height, task.width),
 298                 policies.view(-1, 4, task.height, task.width),
 299             ).flatten(-2)
 300             scores = output
 301         else:
 302             raise ValueError(f"{args.oneshot_output=}")
 303
 304         scores = scores.reshape(-1, task.height, task.width)
 305         mazes = mazes.reshape(-1, task.height, task.width)
 306         targets = targets.reshape(-1, task.height, task.width)
 307         maze.save_image(
 308             os.path.join(
 309                 args.result_dir,
 310                 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
 311             ),
 312             mazes=mazes,
 313             score_paths=scores,
 314             score_truth=targets,
 315         )
 316         # -------------------
 317
 318     gpt.train(t)
 319
 320
 321 ######################################################################
 322
 323
 324 class Task:
 325     def batches(self, split="train", nb_to_use=-1, desc=None):
 326         pass
 327
 328     def vocabulary_size(self):
 329         pass
 330
 331     def produce_results(self, n_epoch, model):
 332         pass
 333
 334
 335 ######################################################################
 336
 337 import maze
 338
 339
 340 class TaskMaze(Task):
 341     def map2seq(self, *m):
 342         return torch.cat([x.flatten(1) for x in m], 1)
 343
 344     def seq2map(self, s):
 345         s = s.reshape(s.size(0), -1, self.height, self.width)
 346         return (s[:, k] for k in range(s.size(1)))
 347
 348     def __init__(
 349         self,
 350         nb_train_samples,
 351         nb_test_samples,
 352         batch_size,
 353         height,
 354         width,
 355         nb_walls,
 356         device=torch.device("cpu"),
 357     ):
 358         self.batch_size = batch_size
 359         self.height = height
 360         self.width = width
 361         self.device = device
 362
 363         train_mazes, train_paths, train_policies = maze.create_maze_data(
 364             nb_train_samples,
 365             height=height,
 366             width=width,
 367             nb_walls=nb_walls,
 368             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 369         )
 370         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 371         self.train_policies = train_policies.flatten(-2).to(device)
 372
 373         test_mazes, test_paths, test_policies = maze.create_maze_data(
 374             nb_test_samples,
 375             height=height,
 376             width=width,
 377             nb_walls=nb_walls,
 378             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 379         )
 380         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 381         self.test_policies = test_policies.flatten(-2).to(device)
 382
 383         self.nb_codes = self.train_input.max() + 1
 384
 385     def batches(self, split="train", nb_to_use=-1, desc=None):
 386         assert split in {"train", "test"}
 387         input = self.train_input if split == "train" else self.test_input
 388         if nb_to_use > 0:
 389             input = input[:nb_to_use]
 390         if desc is None:
 391             desc = f"epoch-{split}"
 392         for batch in tqdm.tqdm(
 393             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 394         ):
 395             yield batch
 396
 397     def policy_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         policies = self.train_policies if split == "train" else self.test_policies
 401         input = input[:, : self.height * self.width]
 402         policies = policies * (input != maze.v_wall)[:, None]
 403
 404         if nb_to_use > 0:
 405             input = input[:nb_to_use]
 406             policies = policies[:nb_to_use]
 407
 408         if desc is None:
 409             desc = f"epoch-{split}"
 410         for batch in tqdm.tqdm(
 411             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 412             dynamic_ncols=True,
 413             desc=desc,
 414         ):
 415             yield batch
 416
 417     def vocabulary_size(self):
 418         return self.nb_codes
 419
 420     def compute_error(self, model, split="train", nb_to_use=-1):
 421         nb_total, nb_correct = 0, 0
 422         for input in task.batches(split, nb_to_use):
 423             result = input.clone()
 424             ar_mask = result.new_zeros(result.size())
 425             ar_mask[:, self.height * self.width :] = 1
 426             result *= 1 - ar_mask
 427             order = random_order(result, self.height * self.width)
 428             masked_inplace_autoregression(
 429                 model, self.batch_size, result, ar_mask, order=order
 430             )
 431             result = shuffle(result, order, reorder=True)
 432             mazes, paths = self.seq2map(result)
 433             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 434             nb_total += mazes.size(0)
 435
 436         return nb_total, nb_correct
 437
 438     def produce_results(self, n_epoch, model):
 439         with torch.autograd.no_grad():
 440             t = model.training
 441             model.eval()
 442
 443             train_nb_total, train_nb_correct = self.compute_error(
 444                 model, "train", nb_to_use=1000
 445             )
 446             log_string(
 447                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 448             )
 449
 450             test_nb_total, test_nb_correct = self.compute_error(
 451                 model, "test", nb_to_use=1000
 452             )
 453             log_string(
 454                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 455             )
 456
 457             input = self.test_input[:32]
 458             result = input.clone()
 459             ar_mask = result.new_zeros(result.size())
 460             ar_mask[:, self.height * self.width :] = 1
 461             result *= 1 - ar_mask
 462             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 463
 464             mazes, paths = self.seq2map(input)
 465             _, predicted_paths = self.seq2map(result)
 466             maze.save_image(
 467                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 468                 mazes=mazes,
 469                 target_paths=paths,
 470                 predicted_paths=predicted_paths,
 471                 path_correct=maze.path_correctness(mazes, predicted_paths),
 472             )
 473
 474             model.train(t)
 475
 476
 477 ######################################################################
 478
 479 log_string(f"device {device}")
 480
 481
 482 task = TaskMaze(
 483     nb_train_samples=args.nb_train_samples,
 484     nb_test_samples=args.nb_test_samples,
 485     batch_size=args.batch_size,
 486     height=args.maze_height,
 487     width=args.maze_width,
 488     nb_walls=args.maze_nb_walls,
 489     device=device,
 490 )
 491
 492
 493 vocabulary_size = task.vocabulary_size()
 494
 495 log_string(f"vocabulary_size {vocabulary_size}")
 496
 497 ##############################
 498
 499 model = mygpt.MyGPT(
 500     vocabulary_size=vocabulary_size,
 501     dim_model=args.dim_model,
 502     dim_keys=args.dim_keys,
 503     dim_hidden=args.dim_hidden,
 504     nb_heads=args.nb_heads,
 505     nb_blocks=args.nb_blocks,
 506     causal=True,
 507     dropout=args.dropout,
 508 )
 509
 510 model.to(device)
 511
 512 nb_parameters = sum(p.numel() for p in model.parameters())
 513 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 514
 515 ######################################################################
 516
 517 nb_epochs_finished = 0
 518
 519 if args.no_checkpoint:
 520     log_string(f"not trying to load checkpoint.")
 521
 522 else:
 523     try:
 524         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 525         checkpoint = torch.load(checkpoint_name)
 526         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 527         model.load_state_dict(checkpoint["model_state"])
 528         torch.set_rng_state(checkpoint["rng_state"])
 529         if torch.cuda.is_available():
 530             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 531
 532         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 533
 534     except FileNotFoundError:
 535         log_string("starting from scratch.")
 536
 537     except:
 538         log_string("error when loading the checkpoint.")
 539         exit(1)
 540
 541 ######################################################################
 542
 543 token_count = 0
 544 for input in task.batches(split="train"):
 545     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 546 token_probas = token_count / token_count.sum()
 547 entropy = -torch.xlogy(token_probas, token_probas).sum()
 548 train_set_perplexity = math.exp(entropy)
 549
 550 ##############################
 551
 552 if args.learning_rate_schedule == "cos":
 553     learning_rate_schedule = {}
 554     for n_epoch in range(args.nb_epochs):
 555         u = n_epoch / args.nb_epochs * math.pi
 556         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 557 else:
 558     u = {
 559         int(k): float(v)
 560         for k, v in [
 561             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 562         ]
 563     }
 564
 565     learning_rate_schedule = {}
 566     learning_rate = args.learning_rate
 567     for n_epoch in range(args.nb_epochs):
 568         if n_epoch in u:
 569             learning_rate = u[n_epoch]
 570         learning_rate_schedule[n_epoch] = learning_rate
 571
 572 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 573
 574 ##############################
 575
 576 if nb_epochs_finished >= args.nb_epochs:
 577     n_epoch = nb_epochs_finished
 578     train_perplexity = compute_perplexity(model, split="train")
 579     test_perplexity = compute_perplexity(model, split="test")
 580
 581     log_string(
 582         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 583     )
 584
 585     task.produce_results(n_epoch, model)
 586
 587 ##############################
 588
 589 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 590     learning_rate = learning_rate_schedule[n_epoch]
 591
 592     log_string(f"learning_rate {learning_rate}")
 593
 594     if args.optim == "sgd":
 595         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 596     elif args.optim == "adam":
 597         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 598     elif args.optim == "adamw":
 599         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 600     else:
 601         raise ValueError(f"{args.optim=}")
 602
 603     model.train()
 604
 605     nb_train_samples, acc_train_loss = 0, 0.0
 606
 607     for input in task.batches(split="train"):
 608         input = input.to(device)
 609         order = random_order(input, task.height * task.width)
 610         input = shuffle(input, order)
 611         output = model(mygpt.BracketedSequence(input), order=order).x
 612         loss = F.cross_entropy(output.transpose(1, 2), input)
 613         acc_train_loss += loss.item() * input.size(0)
 614         nb_train_samples += input.size(0)
 615
 616         optimizer.zero_grad()
 617         loss.backward()
 618         optimizer.step()
 619
 620     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 621     test_perplexity = compute_perplexity(model, split="test")
 622
 623     log_string(
 624         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 625     )
 626
 627     task.produce_results(n_epoch, model)
 628
 629     checkpoint = {
 630         "nb_epochs_finished": n_epoch + 1,
 631         "model_state": model.state_dict(),
 632         "rng_state": torch.get_rng_state(),
 633     }
 634
 635     if torch.cuda.is_available():
 636         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 637
 638     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 639     torch.save(checkpoint, checkpoint_name)
 640     log_string(f"saved checkpoint {checkpoint_name}")
 641
 642 ######################################################################
 643
 644 if args.oneshot:
 645     oneshot(model, task)
 646
 647 ######################################################################