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         return order.sort(1).indices
 139     else:
 140         return (
 141             torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
 142         )
 143
 144
 145 def shuffle(x, order, reorder=False):
 146     if x.dim() == 3:
 147         order = order.unsqueeze(-1).expand(-1, -1, x.size(-1))
 148     if reorder:
 149         y = x.new(x.size())
 150         y.scatter_(1, order, x)
 151         return y
 152     else:
 153         return x.gather(1, order)
 154
 155
 156 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 157 # tokens that should be generated
 158
 159
 160 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
 161     for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
 162         i = (ar_mask.sum(0) > 0).nonzero()
 163         if i.min() > 0:
 164             # Needed to initialize the model's cache
 165             model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
 166         for s in range(i.min(), i.max() + 1):
 167             output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
 168             logits = output[:, s]
 169             if args.deterministic_synthesis:
 170                 t_next = logits.argmax(1)
 171             else:
 172                 dist = torch.distributions.categorical.Categorical(logits=logits)
 173                 t_next = dist.sample()
 174             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 175
 176
 177 ######################################################################
 178
 179
 180 def compute_perplexity(model, split="train"):
 181     with torch.autograd.no_grad():
 182         t = model.training
 183         model.eval()
 184
 185         nb_samples, acc_loss = 0, 0.0
 186
 187         for input in task.batches(split=split):
 188             input = input.to(device)
 189             order = generation_order(input, task.height * task.width)
 190             input = shuffle(input, order)
 191             output = model(mygpt.BracketedSequence(input), order=order).x
 192             loss = F.cross_entropy(output.transpose(1, 2), input)
 193             acc_loss += loss.item() * input.size(0)
 194             nb_samples += input.size(0)
 195
 196         model.train(t)
 197
 198         return math.exp(min(100, acc_loss / nb_samples))
 199
 200
 201 ######################################################################
 202
 203
 204 def oneshot_policy_loss(mazes, output, policies, height, width):
 205     masks = (mazes == maze.v_empty).unsqueeze(-1)
 206     targets = policies.permute(0, 2, 1) * masks
 207     output = output * masks
 208     return -(output.log_softmax(-1) * targets).sum() / masks.sum()
 209
 210
 211 def oneshot_trace_loss(mazes, output, policies, height, width):
 212     masks = mazes == maze.v_empty
 213     targets = maze.stationary_densities(
 214         mazes.view(-1, height, width), policies.view(-1, 4, height, width)
 215     ).flatten(-2)
 216     targets = targets * masks
 217     output = output.squeeze(-1) * masks
 218     return (output - targets).abs().sum() / masks.sum()
 219
 220
 221 def oneshot(gpt, task):
 222     t = gpt.training
 223     gpt.eval()
 224
 225     if args.oneshot_input == "head":
 226         dim_in = args.dim_model
 227     elif args.oneshot_input == "deep":
 228         dim_in = args.dim_model * args.nb_blocks * 2
 229     else:
 230         raise ValueError(f"{args.oneshot_input=}")
 231
 232     if args.oneshot_output == "policy":
 233         dim_out = 4
 234         compute_loss = oneshot_policy_loss
 235     elif args.oneshot_output == "trace":
 236         dim_out = 1
 237         compute_loss = oneshot_trace_loss
 238     else:
 239         raise ValueError(f"{args.oneshot_output=}")
 240
 241     model = nn.Sequential(
 242         nn.Linear(dim_in, args.dim_model),
 243         nn.ReLU(),
 244         nn.Linear(args.dim_model, args.dim_model),
 245         nn.ReLU(),
 246         nn.Linear(args.dim_model, dim_out),
 247     ).to(device)
 248
 249     for n_epoch in range(args.nb_epochs):
 250         learning_rate = learning_rate_schedule[n_epoch]
 251         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 252
 253         acc_train_loss, nb_train_samples = 0, 0
 254         for mazes, policies in task.policy_batches(split="train"):
 255             order = generation_order(mazes, task.height * task.width)
 256             x = shuffle(mazes, order)
 257             x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 258             output_gpt = shuffle(x, order, reorder=True)
 259             output = model(output_gpt)
 260
 261             loss = compute_loss(mazes, output, policies, task.height, task.width)
 262             acc_train_loss += loss.item() * mazes.size(0)
 263             nb_train_samples += mazes.size(0)
 264
 265             optimizer.zero_grad()
 266             loss.backward()
 267             optimizer.step()
 268
 269         acc_test_loss, nb_test_samples = 0, 0
 270         for mazes, policies in task.policy_batches(split="test"):
 271             order = generation_order(mazes, task.height * task.width)
 272             x = shuffle(mazes, order)
 273             x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 274             output_gpt = shuffle(x, order, reorder=True)
 275             output = model(output_gpt)
 276             loss = compute_loss(mazes, output, policies, task.height, task.width)
 277             acc_test_loss += loss.item() * mazes.size(0)
 278             nb_test_samples += mazes.size(0)
 279
 280         log_string(
 281             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 282         )
 283
 284         # -------------------
 285         mazes = task.test_input[:32, : task.height * task.width]
 286         policies = task.test_policies[:32]
 287         order = generation_order(mazes, task.height * task.width)
 288         x = shuffle(mazes, order)
 289         x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
 290         output_gpt = shuffle(x, order, reorder=True)
 291         output = model(output_gpt)
 292         if args.oneshot_output == "policy":
 293             targets = policies.permute(0, 2, 1)
 294             scores = (
 295                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 296             ).float()
 297         elif args.oneshot_output == "trace":
 298             targets = maze.stationary_densities(
 299                 mazes.view(-1, task.height, task.width),
 300                 policies.view(-1, 4, task.height, task.width),
 301             ).flatten(-2)
 302             scores = output
 303         else:
 304             raise ValueError(f"{args.oneshot_output=}")
 305
 306         scores = scores.reshape(-1, task.height, task.width)
 307         mazes = mazes.reshape(-1, task.height, task.width)
 308         targets = targets.reshape(-1, task.height, task.width)
 309         maze.save_image(
 310             os.path.join(
 311                 args.result_dir,
 312                 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
 313             ),
 314             mazes=mazes,
 315             score_paths=scores,
 316             score_truth=targets,
 317         )
 318         # -------------------
 319
 320     gpt.train(t)
 321
 322
 323 ######################################################################
 324
 325
 326 class Task:
 327     def batches(self, split="train", nb_to_use=-1, desc=None):
 328         pass
 329
 330     def vocabulary_size(self):
 331         pass
 332
 333     def produce_results(self, n_epoch, model):
 334         pass
 335
 336
 337 ######################################################################
 338
 339 import maze
 340
 341
 342 class TaskMaze(Task):
 343     def map2seq(self, *m):
 344         return torch.cat([x.flatten(1) for x in m], 1)
 345
 346     def seq2map(self, s):
 347         s = s.reshape(s.size(0), -1, self.height, self.width)
 348         return (s[:, k] for k in range(s.size(1)))
 349
 350     def __init__(
 351         self,
 352         nb_train_samples,
 353         nb_test_samples,
 354         batch_size,
 355         height,
 356         width,
 357         nb_walls,
 358         device=torch.device("cpu"),
 359     ):
 360         self.batch_size = batch_size
 361         self.height = height
 362         self.width = width
 363         self.device = device
 364
 365         train_mazes, train_paths, train_policies = maze.create_maze_data(
 366             nb_train_samples,
 367             height=height,
 368             width=width,
 369             nb_walls=nb_walls,
 370             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 371         )
 372         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 373         self.train_policies = train_policies.flatten(-2).to(device)
 374
 375         test_mazes, test_paths, test_policies = maze.create_maze_data(
 376             nb_test_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-test"),
 381         )
 382         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 383         self.test_policies = test_policies.flatten(-2).to(device)
 384
 385         self.nb_codes = self.train_input.max() + 1
 386
 387     def batches(self, split="train", nb_to_use=-1, desc=None):
 388         assert split in {"train", "test"}
 389         input = self.train_input if split == "train" else self.test_input
 390         if nb_to_use > 0:
 391             input = input[:nb_to_use]
 392         if desc is None:
 393             desc = f"epoch-{split}"
 394         for batch in tqdm.tqdm(
 395             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 396         ):
 397             yield batch
 398
 399     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 400         assert split in {"train", "test"}
 401         input = self.train_input if split == "train" else self.test_input
 402         policies = self.train_policies if split == "train" else self.test_policies
 403         input = input[:, : self.height * self.width]
 404         policies = policies * (input != maze.v_wall)[:, None]
 405
 406         if nb_to_use > 0:
 407             input = input[:nb_to_use]
 408             policies = policies[:nb_to_use]
 409
 410         if desc is None:
 411             desc = f"epoch-{split}"
 412         for batch in tqdm.tqdm(
 413             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 414             dynamic_ncols=True,
 415             desc=desc,
 416         ):
 417             yield batch
 418
 419     def vocabulary_size(self):
 420         return self.nb_codes
 421
 422     def compute_error(self, model, split="train", nb_to_use=-1):
 423         nb_total, nb_correct = 0, 0
 424         for input in task.batches(split, nb_to_use):
 425             result = input.clone()
 426             ar_mask = result.new_zeros(result.size())
 427             ar_mask[:, self.height * self.width :] = 1
 428             result *= 1 - ar_mask
 429             order = generation_order(result, self.height * self.width)
 430             masked_inplace_autoregression(
 431                 model, self.batch_size, result, ar_mask, order=order
 432             )
 433             result = shuffle(result, order, reorder=True)
 434             mazes, paths = self.seq2map(result)
 435             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 436             nb_total += mazes.size(0)
 437
 438         return nb_total, nb_correct
 439
 440     def produce_results(self, n_epoch, model):
 441         with torch.autograd.no_grad():
 442             t = model.training
 443             model.eval()
 444
 445             train_nb_total, train_nb_correct = self.compute_error(
 446                 model, "train", nb_to_use=1000
 447             )
 448             log_string(
 449                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 450             )
 451
 452             test_nb_total, test_nb_correct = self.compute_error(
 453                 model, "test", nb_to_use=1000
 454             )
 455             log_string(
 456                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 457             )
 458
 459             input = self.test_input[:32]
 460             result = input.clone()
 461             ar_mask = result.new_zeros(result.size())
 462             ar_mask[:, self.height * self.width :] = 1
 463             result *= 1 - ar_mask
 464             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 465
 466             mazes, paths = self.seq2map(input)
 467             _, predicted_paths = self.seq2map(result)
 468             maze.save_image(
 469                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 470                 mazes=mazes,
 471                 target_paths=paths,
 472                 predicted_paths=predicted_paths,
 473                 path_correct=maze.path_correctness(mazes, predicted_paths),
 474             )
 475
 476             model.train(t)
 477
 478
 479 ######################################################################
 480
 481 log_string(f"device {device}")
 482
 483
 484 task = TaskMaze(
 485     nb_train_samples=args.nb_train_samples,
 486     nb_test_samples=args.nb_test_samples,
 487     batch_size=args.batch_size,
 488     height=args.maze_height,
 489     width=args.maze_width,
 490     nb_walls=args.maze_nb_walls,
 491     device=device,
 492 )
 493
 494
 495 vocabulary_size = task.vocabulary_size()
 496
 497 log_string(f"vocabulary_size {vocabulary_size}")
 498
 499 ##############################
 500
 501 model = mygpt.MyGPT(
 502     vocabulary_size=vocabulary_size,
 503     dim_model=args.dim_model,
 504     dim_keys=args.dim_keys,
 505     dim_hidden=args.dim_hidden,
 506     nb_heads=args.nb_heads,
 507     nb_blocks=args.nb_blocks,
 508     causal=True,
 509     dropout=args.dropout,
 510 )
 511
 512 model.to(device)
 513
 514 nb_parameters = sum(p.numel() for p in model.parameters())
 515 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 516
 517 ######################################################################
 518
 519 nb_epochs_finished = 0
 520
 521 if args.no_checkpoint:
 522     log_string(f"not trying to load checkpoint.")
 523
 524 else:
 525     try:
 526         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 527         checkpoint = torch.load(checkpoint_name)
 528         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 529         model.load_state_dict(checkpoint["model_state"])
 530         torch.set_rng_state(checkpoint["rng_state"])
 531         if torch.cuda.is_available():
 532             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 533
 534         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 535
 536     except FileNotFoundError:
 537         log_string("starting from scratch.")
 538
 539     except:
 540         log_string("error when loading the checkpoint.")
 541         exit(1)
 542
 543 ######################################################################
 544
 545 token_count = 0
 546 for input in task.batches(split="train"):
 547     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 548 token_probas = token_count / token_count.sum()
 549 entropy = -torch.xlogy(token_probas, token_probas).sum()
 550 train_set_perplexity = math.exp(entropy)
 551
 552 ##############################
 553
 554 if args.learning_rate_schedule == "cos":
 555     learning_rate_schedule = {}
 556     for n_epoch in range(args.nb_epochs):
 557         u = n_epoch / args.nb_epochs * math.pi
 558         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 559 else:
 560     u = {
 561         int(k): float(v)
 562         for k, v in [
 563             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 564         ]
 565     }
 566
 567     learning_rate_schedule = {}
 568     learning_rate = args.learning_rate
 569     for n_epoch in range(args.nb_epochs):
 570         if n_epoch in u:
 571             learning_rate = u[n_epoch]
 572         learning_rate_schedule[n_epoch] = learning_rate
 573
 574 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 575
 576 ##############################
 577
 578 if nb_epochs_finished >= args.nb_epochs:
 579     n_epoch = nb_epochs_finished
 580     train_perplexity = compute_perplexity(model, split="train")
 581     test_perplexity = compute_perplexity(model, split="test")
 582
 583     log_string(
 584         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 585     )
 586
 587     task.produce_results(n_epoch, model)
 588
 589 ##############################
 590
 591 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 592     learning_rate = learning_rate_schedule[n_epoch]
 593
 594     log_string(f"learning_rate {learning_rate}")
 595
 596     if args.optim == "sgd":
 597         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 598     elif args.optim == "adam":
 599         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 600     elif args.optim == "adamw":
 601         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 602     else:
 603         raise ValueError(f"{args.optim=}")
 604
 605     model.train()
 606
 607     nb_train_samples, acc_train_loss = 0, 0.0
 608
 609     for input in task.batches(split="train"):
 610         input = input.to(device)
 611         order = generation_order(input, task.height * task.width)
 612         input = shuffle(input, order)
 613         output = model(mygpt.BracketedSequence(input), order=order).x
 614         loss = F.cross_entropy(output.transpose(1, 2), input)
 615         acc_train_loss += loss.item() * input.size(0)
 616         nb_train_samples += input.size(0)
 617
 618         optimizer.zero_grad()
 619         loss.backward()
 620         optimizer.step()
 621
 622     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 623     test_perplexity = compute_perplexity(model, split="test")
 624
 625     log_string(
 626         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 627     )
 628
 629     task.produce_results(n_epoch, model)
 630
 631     checkpoint = {
 632         "nb_epochs_finished": n_epoch + 1,
 633         "model_state": model.state_dict(),
 634         "rng_state": torch.get_rng_state(),
 635     }
 636
 637     if torch.cuda.is_available():
 638         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 639
 640     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 641     torch.save(checkpoint, checkpoint_name)
 642     log_string(f"saved checkpoint {checkpoint_name}")
 643
 644 ######################################################################
 645
 646 if args.oneshot:
 647     oneshot(model, task)
 648
 649 ######################################################################