tasks.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 import math, os, tqdm
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 from mygpt import BracketedSequence
  16
  17 try:
  18     from graph import save_attention_image
  19 except ImportError:
  20     save_attention_image = None
  21
  22 ######################################################################
  23
  24
  25 def masked_inplace_autoregression(
  26     model,
  27     batch_size,
  28     input,
  29     ar_mask,
  30     deterministic_synthesis,
  31     forbidden_tokens=None,
  32     progress_bar_desc="autoregression",
  33     device=torch.device("cpu"),
  34 ):
  35     assert input.size() == ar_mask.size()
  36
  37     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
  38
  39     if progress_bar_desc is not None:
  40         batches = tqdm.tqdm(
  41             batches,
  42             dynamic_ncols=True,
  43             desc=progress_bar_desc,
  44             total=(input.size(0) + batch_size - 1) // batch_size,
  45         )
  46
  47     with torch.autograd.no_grad():
  48         t = model.training
  49         model.eval()
  50
  51         for input, ar_mask in batches:
  52             model.masked_inplace_autoregression(
  53                 input, ar_mask, forbidden_tokens, deterministic_synthesis
  54             )
  55
  56         model.train(t)
  57
  58
  59 ######################################################################
  60
  61
  62 class Task:
  63     def batches(self, split="train"):
  64         pass
  65
  66     def vocabulary_size(self):
  67         pass
  68
  69     def produce_results(
  70         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  71     ):
  72         pass
  73
  74
  75 ####################
  76
  77 import problems
  78
  79 class SandBox(Task):
  80     def __init__(
  81         self,
  82         problem,
  83         nb_train_samples,
  84         nb_test_samples,
  85         batch_size,
  86         logger=None,
  87         device=torch.device("cpu"),
  88         max_nb_codes=1024,
  89     ):
  90         super().__init__()
  91
  92         self.batch_size = batch_size
  93         self.device = device
  94         self.problem = problem
  95
  96         self.train_input, self.train_ar_mask = self.problem.generate_sequences(
  97             nb_train_samples
  98         )
  99         self.test_input, self.test_ar_mask = self.problem.generate_sequences(
 100             nb_test_samples
 101         )
 102
 103         self.train_input, self.train_ar_mask = self.train_input.to(
 104             device
 105         ), self.train_ar_mask.to(device)
 106         self.test_input, self.test_ar_mask = self.test_input.to(
 107             device
 108         ), self.test_ar_mask.to(device)
 109
 110         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 111
 112         # A bit of paranoia never hurts
 113         assert (
 114             self.nb_codes <= max_nb_codes
 115             and self.train_input.min() >= 0
 116             and self.test_input.min() >= 0
 117             and tuple(self.train_ar_mask.unique()) == (0, 1)
 118             and tuple(self.test_ar_mask.unique()) == (0, 1)
 119         )
 120
 121     def batches(self, split="train", nb_to_use=-1, desc=None):
 122         assert split in {"train", "test"}
 123         input = self.train_input if split == "train" else self.test_input
 124         if nb_to_use > 0:
 125             input = input[:nb_to_use]
 126         if desc is None:
 127             desc = f"epoch-{split}"
 128         for batch in tqdm.tqdm(
 129             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 130         ):
 131             yield batch
 132
 133     def vocabulary_size(self):
 134         return self.nb_codes
 135
 136     def produce_results(
 137         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 138     ):
 139         def compute_accuracy(input, ar_mask, logger=None):
 140             input, ar_mask = input[:nmax], ar_mask[:nmax]
 141             result = input.clone() * (1 - ar_mask)
 142
 143             masked_inplace_autoregression(
 144                 model,
 145                 self.batch_size,
 146                 result,
 147                 ar_mask,
 148                 deterministic_synthesis,
 149                 progress_bar_desc=None,
 150                 device=self.device,
 151             )
 152
 153             if logger is not None:
 154                 for sp, st in zip(result[:10], input[:10]):
 155                     logger(
 156                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 157                     )
 158                     logger(
 159                         f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 160                     )
 161
 162             nb_total = ar_mask.sum().item()
 163             nb_correct = ((result == input).long() * ar_mask).sum().item()
 164
 165             return nb_total, nb_correct
 166
 167         train_nb_total, train_nb_correct = compute_accuracy(
 168             self.train_input, self.train_ar_mask
 169         )
 170
 171         logger(
 172             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 173         )
 174
 175         test_nb_total, test_nb_correct = compute_accuracy(
 176             self.test_input, self.test_ar_mask, logger
 177         )
 178
 179         logger(
 180             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 181         )
 182
 183
 184 ######################################################################
 185
 186 import picoclvr
 187
 188
 189 class PicoCLVR(Task):
 190     # Make a tensor from a list of strings
 191     def tensorize(self, descr):
 192         token_descr = [s.strip().split(" ") for s in descr]
 193         l = max([len(s) for s in token_descr])
 194         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 195         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 196         return torch.tensor(id_descr, device=self.device)
 197
 198     # Make a list of strings from a tensor
 199     def detensorize(self, x):
 200         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 201
 202     # trim all the tensors in the tuple z to remove as much token from
 203     # left and right in the first tensor. If z is a tuple, all its
 204     # elements are trimed according to the triming for the first
 205     def trim(self, z, token="<nul>"):
 206         n = self.token2id[token]
 207         if type(z) == tuple:
 208             x = z[0]
 209             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 210             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 211             return tuple([t[:, a:b] for t in z])
 212         else:
 213             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 214             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 215             return z[:, a:b]
 216
 217     ######################
 218
 219     def __init__(
 220         self,
 221         nb_train_samples,
 222         nb_test_samples,
 223         batch_size,
 224         height,
 225         width,
 226         nb_colors=5,
 227         logger=None,
 228         device=torch.device("cpu"),
 229         pruner_train=None,
 230         pruner_eval=None,
 231     ):
 232         super().__init__()
 233
 234         def generate_descr(nb, cache_suffix, pruner):
 235             return picoclvr.generate(
 236                 nb,
 237                 height=self.height,
 238                 width=self.width,
 239                 nb_colors=nb_colors,
 240                 pruner=pruner,
 241             )
 242
 243         self.height = height
 244         self.width = width
 245         self.batch_size = batch_size
 246         self.device = device
 247         self.pruner_train = pruner_train
 248         self.pruner_eval = pruner_eval
 249
 250         if logger is not None:
 251             logger(
 252                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 253             )
 254
 255         self.train_descr = generate_descr(
 256             nb_train_samples, "train", pruner=self.pruner_train
 257         )
 258         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 259
 260         # Build the tokenizer
 261         tokens = {"<nul>", "<img>"}
 262         for d in [self.train_descr, self.test_descr]:
 263             for s in d:
 264                 for t in s.strip().split(" "):
 265                     tokens.add(t)
 266         # make this set a sorted list to get the same tensors given
 267         # the same descr
 268         tokens = list(tokens)
 269         tokens.sort()
 270         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 271         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 272         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 273
 274         # Tokenize the train and test sets
 275         self.train_input = self.tensorize(self.train_descr)
 276         self.test_input = self.tensorize(self.test_descr)
 277
 278     def batches(self, split="train"):
 279         assert split in {"train", "test"}
 280         input = self.train_input if split == "train" else self.test_input
 281         for batch in tqdm.tqdm(
 282             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 283         ):
 284             yield self.trim(batch)
 285
 286     def vocabulary_size(self):
 287         return len(self.token2id)
 288
 289     def compute_missing_properties(
 290         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 291     ):
 292         acc_nb_requested_properties = []
 293         acc_nb_missing_properties = []
 294         acc_nb_results = 0
 295
 296         for input in tqdm.tqdm(
 297             self.test_input.split(self.batch_size),
 298             dynamic_ncols=True,
 299             desc=f"test-properties",
 300         ):
 301             result = input.clone()
 302             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 303             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 304             masked_inplace_autoregression(
 305                 model,
 306                 self.batch_size,
 307                 result,
 308                 ar_mask,
 309                 deterministic_synthesis,
 310                 progress_bar_desc=None,
 311                 device=self.device,
 312             )
 313
 314             result_descr = self.detensorize(result)
 315             np = picoclvr.nb_properties(
 316                 result_descr,
 317                 height=self.height,
 318                 width=self.width,
 319                 pruner=pruner,
 320             )
 321             nb_requested_properties, _, nb_missing_properties = zip(*np)
 322             acc_nb_requested_properties += nb_requested_properties
 323             acc_nb_missing_properties += nb_missing_properties
 324             acc_nb_results += len(result_descr)
 325
 326         nb_requested_properties = sum(acc_nb_requested_properties)
 327         nb_missing_properties = sum(acc_nb_missing_properties)
 328
 329         prefix = "" if pruner is None else "pruned_"
 330         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 331         logger(
 332             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 333         )
 334         logger(
 335             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 336         )
 337
 338     ######################################################################
 339
 340     def produce_results(
 341         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 342     ):
 343         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 344
 345         if self.pruner_eval is not None:
 346             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 347
 348         nb_tokens_to_generate = self.height * self.width + 3
 349         result_descr = []
 350         nb_per_primer = 8
 351         primer = []
 352
 353         for primer_descr in [
 354             "red above green <sep> green top <sep> blue right of red",
 355             "there is red <sep> there is yellow <sep> there is blue",
 356             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 357             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 358         ]:
 359             primer += [primer_descr + " <img>"] * nb_per_primer
 360
 361         result = self.tensorize(primer)
 362         fill = result.new_full(
 363             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 364         )
 365         result = torch.cat((result, fill), 1)
 366         ar_mask = (result == self.t_nul).long()
 367         masked_inplace_autoregression(
 368             model,
 369             self.batch_size,
 370             result,
 371             ar_mask,
 372             deterministic_synthesis,
 373             device=self.device,
 374         )
 375         result_descr = self.detensorize(result)
 376
 377         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 378
 379         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 380         acc_nb_results = len(result_descr)
 381
 382         nb_requested_properties = sum(acc_nb_requested_properties)
 383         nb_missing_properties = sum(acc_nb_missing_properties)
 384
 385         prefix = "demo_"
 386         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 387         logger(
 388             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 389         )
 390         logger(
 391             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 392         )
 393
 394         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 395
 396         if img.dim() == 5:
 397             if img.size(1) == 1:
 398                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 399             else:
 400                 img = torch.cat(
 401                     [
 402                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 403                         for x in img
 404                     ],
 405                     0,
 406                 )
 407
 408         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 409         torchvision.utils.save_image(
 410             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 411         )
 412         logger(f"wrote {image_name}")
 413
 414
 415 ######################################################################
 416
 417
 418 class MNIST(Task):
 419     def __init__(
 420         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 421     ):
 422         super().__init__()
 423
 424         self.nb_train_samples = (nb_train_samples,)
 425         self.nb_test_samples = (nb_test_samples,)
 426         self.batch_size = batch_size
 427         self.device = device
 428         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 429         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 430         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 431         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 432
 433     def batches(self, split="train", nb_to_use=-1, desc=None):
 434         assert split in {"train", "test"}
 435         input = self.train_input if split == "train" else self.test_input
 436         if nb_to_use > 0:
 437             input = input[:nb_to_use]
 438         if desc is None:
 439             desc = f"epoch-{split}"
 440         for batch in tqdm.tqdm(
 441             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 442         ):
 443             yield batch
 444
 445     def vocabulary_size(self):
 446         return 256
 447
 448     def produce_results(
 449         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 450     ):
 451         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 452         ar_mask = torch.full_like(results, 1)
 453         masked_inplace_autoregression(
 454             model,
 455             self.batch_size,
 456             results,
 457             ar_mask,
 458             deterministic_synthesis,
 459             device=self.device,
 460         )
 461         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 462         torchvision.utils.save_image(
 463             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 464             image_name,
 465             nrow=16,
 466             pad_value=0.8,
 467         )
 468         logger(f"wrote {image_name}")
 469
 470
 471 ######################################################################
 472
 473 import maze
 474
 475
 476 class Maze(Task):
 477     def map2seq(self, *m):
 478         return torch.cat([x.flatten(1) for x in m], 1)
 479
 480     def seq2map(self, s):
 481         s = s.reshape(s.size(0), -1, self.height, self.width)
 482         return (s[:, k] for k in range(s.size(1)))
 483
 484     def __init__(
 485         self,
 486         nb_train_samples,
 487         nb_test_samples,
 488         batch_size,
 489         height,
 490         width,
 491         nb_walls,
 492         device=torch.device("cpu"),
 493     ):
 494         super().__init__()
 495
 496         self.batch_size = batch_size
 497         self.height = height
 498         self.width = width
 499         self.device = device
 500
 501         train_mazes, train_paths, _ = maze.create_maze_data(
 502             nb_train_samples,
 503             height=height,
 504             width=width,
 505             nb_walls=nb_walls,
 506             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 507         )
 508         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 509
 510         test_mazes, test_paths, _ = maze.create_maze_data(
 511             nb_test_samples,
 512             height=height,
 513             width=width,
 514             nb_walls=nb_walls,
 515             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 516         )
 517         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 518
 519         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 520
 521     def batches(self, split="train", nb_to_use=-1, desc=None):
 522         assert split in {"train", "test"}
 523         input = self.train_input if split == "train" else self.test_input
 524         if nb_to_use > 0:
 525             input = input[:nb_to_use]
 526         if desc is None:
 527             desc = f"epoch-{split}"
 528         for batch in tqdm.tqdm(
 529             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 530         ):
 531             yield batch
 532
 533     def vocabulary_size(self):
 534         return self.nb_codes
 535
 536     def compute_error(
 537         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 538     ):
 539         nb_total, nb_correct = 0, 0
 540         count = torch.zeros(
 541             self.width * self.height,
 542             self.width * self.height,
 543             device=self.device,
 544             dtype=torch.int64,
 545         )
 546
 547         for input in self.batches(split, nb_to_use):
 548             result = input.clone()
 549             ar_mask = result.new_zeros(result.size())
 550             ar_mask[:, self.height * self.width :] = 1
 551             result *= 1 - ar_mask
 552             masked_inplace_autoregression(
 553                 model,
 554                 self.batch_size,
 555                 result,
 556                 ar_mask,
 557                 deterministic_synthesis,
 558                 progress_bar_desc=None,
 559                 device=self.device,
 560             )
 561             mazes, paths = self.seq2map(result)
 562             path_correctness = maze.path_correctness(mazes, paths)
 563             nb_correct += path_correctness.long().sum()
 564             nb_total += mazes.size(0)
 565
 566             optimal_path_lengths = (
 567                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 568             )
 569             predicted_path_lengths = (
 570                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 571             )
 572             optimal_path_lengths = optimal_path_lengths[path_correctness]
 573             predicted_path_lengths = predicted_path_lengths[path_correctness]
 574             count[optimal_path_lengths, predicted_path_lengths] += 1
 575
 576         if count.max() == 0:
 577             count = None
 578         else:
 579             count = count[
 580                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 581             ]
 582
 583         return nb_total, nb_correct, count
 584
 585     def produce_results(
 586         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 587     ):
 588         train_nb_total, train_nb_correct, count = self.compute_error(
 589             model,
 590             "train",
 591             nb_to_use=1000,
 592             deterministic_synthesis=deterministic_synthesis,
 593         )
 594         logger(
 595             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 596         )
 597
 598         test_nb_total, test_nb_correct, count = self.compute_error(
 599             model,
 600             "test",
 601             nb_to_use=1000,
 602             deterministic_synthesis=deterministic_synthesis,
 603         )
 604         logger(
 605             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 606         )
 607
 608         if count is not None:
 609             proportion_optimal = count.diagonal().sum().float() / count.sum()
 610             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 611             with open(
 612                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 613             ) as f:
 614                 for i in range(count.size(0)):
 615                     for j in range(count.size(1)):
 616                         eol = " " if j < count.size(1) - 1 else "\n"
 617                         f.write(f"{count[i,j]}{eol}")
 618
 619         input = self.test_input[:48]
 620         result = input.clone()
 621         ar_mask = result.new_zeros(result.size())
 622         ar_mask[:, self.height * self.width :] = 1
 623         result *= 1 - ar_mask
 624         masked_inplace_autoregression(
 625             model,
 626             self.batch_size,
 627             result,
 628             ar_mask,
 629             deterministic_synthesis,
 630             device=self.device,
 631         )
 632
 633         mazes, paths = self.seq2map(input)
 634         _, predicted_paths = self.seq2map(result)
 635
 636         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 637         maze.save_image(
 638             filename,
 639             mazes=mazes,
 640             target_paths=paths,
 641             predicted_paths=predicted_paths,
 642             path_correct=maze.path_correctness(mazes, predicted_paths),
 643             path_optimal=maze.path_optimality(paths, predicted_paths),
 644         )
 645         logger(f"wrote {filename}")
 646
 647
 648 ######################################################################
 649
 650
 651 import snake
 652
 653
 654 class Snake(Task):
 655     def __init__(
 656         self,
 657         nb_train_samples,
 658         nb_test_samples,
 659         batch_size,
 660         height,
 661         width,
 662         nb_colors,
 663         length,
 664         prompt_length,
 665         device=torch.device("cpu"),
 666     ):
 667         super().__init__()
 668
 669         self.batch_size = batch_size
 670         self.height = height
 671         self.width = width
 672         self.device = device
 673         self.prompt_length = prompt_length
 674
 675         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 676             nb_train_samples,
 677             height,
 678             width,
 679             nb_colors,
 680             length,
 681             prompt_length,
 682             self.device,
 683         )
 684         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 685             nb_test_samples,
 686             height,
 687             width,
 688             nb_colors,
 689             length,
 690             prompt_length,
 691             self.device,
 692         )
 693
 694         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 695
 696     def batches(self, split="train", nb_to_use=-1, desc=None):
 697         assert split in {"train", "test"}
 698         input = self.train_input if split == "train" else self.test_input
 699         if nb_to_use > 0:
 700             input = input[:nb_to_use]
 701         if desc is None:
 702             desc = f"epoch-{split}"
 703         for batch in tqdm.tqdm(
 704             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 705         ):
 706             yield batch
 707
 708     def vocabulary_size(self):
 709         return self.nb_codes
 710
 711     def produce_results(
 712         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 713     ):
 714         def compute_nb_correct(input, prior_visits):
 715             result = input.clone()
 716             i = torch.arange(result.size(1), device=result.device)[None, :]
 717             ar_mask = (
 718                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 719                 .long()
 720                 .expand_as(result)
 721             )
 722             result *= 1 - ar_mask
 723
 724             masked_inplace_autoregression(
 725                 model,
 726                 self.batch_size,
 727                 result,
 728                 ar_mask,
 729                 deterministic_synthesis,
 730                 device=self.device,
 731             )
 732
 733             nb_total = ((prior_visits > 0) * ar_mask).sum()
 734
 735             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 736
 737             return nb_total, nb_correct
 738
 739         test_nb_total, test_nb_correct = compute_nb_correct(
 740             self.test_input[:1000], self.test_prior_visits[:1000]
 741         )
 742
 743         logger(
 744             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 745         )
 746
 747
 748 ######################################################################
 749
 750
 751 import stack
 752
 753
 754 class Stack(Task):
 755     def __init__(
 756         self,
 757         nb_train_samples,
 758         nb_test_samples,
 759         batch_size,
 760         logger,
 761         nb_steps,
 762         nb_stacks,
 763         nb_digits,
 764         fraction_values_for_train=None,
 765         device=torch.device("cpu"),
 766     ):
 767         super().__init__()
 768
 769         self.batch_size = batch_size
 770         self.nb_steps = nb_steps
 771         self.nb_stacks = nb_stacks
 772         self.nb_digits = nb_digits
 773         self.device = device
 774
 775         if fraction_values_for_train is None:
 776             values_for_train = None
 777             values_for_test = None
 778         else:
 779             all = torch.randperm(10**nb_digits)
 780             nb_for_train = int(all.size(0) * fraction_values_for_train)
 781             values_for_train = all[:nb_for_train]
 782             values_for_test = all[nb_for_train:]
 783
 784         self.train_input, self.train_stack_counts = stack.generate_sequences(
 785             nb_train_samples,
 786             nb_steps,
 787             nb_stacks,
 788             nb_digits,
 789             values_for_train,
 790             self.device,
 791         )
 792
 793         self.test_input, self.test_stack_counts = stack.generate_sequences(
 794             nb_test_samples,
 795             nb_steps,
 796             nb_stacks,
 797             nb_digits,
 798             values_for_test,
 799             self.device,
 800         )
 801
 802         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 803         counts = self.test_stack_counts.flatten()[i.flatten()]
 804         counts = F.one_hot(counts).sum(0)
 805         logger(f"test_pop_stack_counts {counts}")
 806
 807         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 808
 809     def batches(self, split="train", nb_to_use=-1, desc=None):
 810         assert split in {"train", "test"}
 811         input = self.train_input if split == "train" else self.test_input
 812         if nb_to_use > 0:
 813             input = input[:nb_to_use]
 814         if desc is None:
 815             desc = f"epoch-{split}"
 816         for batch in tqdm.tqdm(
 817             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 818         ):
 819             yield batch
 820
 821     def vocabulary_size(self):
 822         return self.nb_codes
 823
 824     def produce_results(
 825         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 826     ):
 827         def compute_nb_correct(input):
 828             result = input.clone()
 829             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 830             ar_mask = (result != input).long()
 831             masked_inplace_autoregression(
 832                 model,
 833                 self.batch_size,
 834                 result,
 835                 ar_mask,
 836                 deterministic_synthesis,
 837                 device=self.device,
 838             )
 839
 840             errors = ((result != input).long() * ar_mask).reshape(
 841                 -1, 1 + self.nb_digits
 842             )
 843             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 844
 845             nb_total = ar_mask.max(1).values.sum()
 846             nb_correct = nb_total - errors.max(1).values.sum()
 847
 848             return nb_total, nb_correct
 849
 850         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 851
 852         logger(
 853             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 854         )
 855
 856         ##############################################################
 857         # Log a few generated sequences
 858         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 859         result = input.clone()
 860         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 861         ar_mask = (result != input).long()
 862
 863         # for n in range(result.size(0)):
 864         # logger(
 865         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 866         # )
 867
 868         masked_inplace_autoregression(
 869             model,
 870             self.batch_size,
 871             result,
 872             ar_mask,
 873             deterministic_synthesis,
 874             device=self.device,
 875         )
 876
 877         for n in range(result.size(0)):
 878             logger(
 879                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 880             )
 881         ##############################################################
 882
 883
 884 ######################################################################
 885
 886 import rpl
 887
 888
 889 class RPL(Task):
 890     def tensorize(self, sequences):
 891         len_max = max([len(x) for x in sequences])
 892         return torch.cat(
 893             [
 894                 torch.tensor(
 895                     [
 896                         [
 897                             self.token2id[str(c)]
 898                             for c in s + ["<nul>"] * (len_max - len(s))
 899                         ]
 900                         for s in sequences
 901                     ]
 902                 )
 903             ],
 904             0,
 905         )
 906
 907     def seq2str(self, seq):
 908         return " ".join([self.id2token[i] for i in seq])
 909
 910     def __init__(
 911         self,
 912         nb_train_samples,
 913         nb_test_samples,
 914         batch_size,
 915         nb_starting_values=3,
 916         max_input=9,
 917         prog_len=6,
 918         nb_runs=5,
 919         no_prog=False,
 920         logger=None,
 921         device=torch.device("cpu"),
 922     ):
 923         super().__init__()
 924
 925         self.batch_size = batch_size
 926         self.device = device
 927         self.no_prog = no_prog
 928
 929         train_sequences = [
 930             rpl.generate(
 931                 nb_starting_values=nb_starting_values,
 932                 nb_result_values_max=4 * nb_starting_values,
 933                 max_input=max_input,
 934                 prog_len=prog_len,
 935                 nb_runs=nb_runs,
 936             )
 937             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
 938         ]
 939
 940         test_sequences = [
 941             rpl.generate(
 942                 nb_starting_values=nb_starting_values,
 943                 nb_result_values_max=4 * nb_starting_values,
 944                 max_input=max_input,
 945                 prog_len=prog_len,
 946                 nb_runs=nb_runs,
 947             )
 948             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
 949         ]
 950
 951         symbols = list(
 952             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
 953         )
 954         val_max = max([x if type(x) is int else 0 for x in symbols])
 955         symbols = list(filter(lambda x: type(x) is str, symbols))
 956         symbols.sort()
 957         symbols += [str(n) for n in range(val_max + 1)]
 958         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
 959         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
 960
 961         self.t_nul = self.token2id["<nul>"]
 962         self.t_input = self.token2id["<in>"]
 963         self.t_output = self.token2id["<out>"]
 964         self.t_prog = self.token2id["<prg>"]
 965         self.t_end = self.token2id["<end>"]
 966
 967         self.train_input = self.tensorize(train_sequences)
 968         self.test_input = self.tensorize(test_sequences)
 969
 970         if no_prog:
 971             # Excise the program from every train and test example
 972             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
 973                 None, :
 974             ]
 975             p = (
 976                 ((self.train_input == self.t_prog).long() * k)
 977                 .max(1, keepdim=True)
 978                 .values
 979             )
 980             self.train_input = (
 981                 self.train_input * (k <= p).long()
 982                 + self.t_end * (k == p + 1).long()
 983                 + self.t_nul * (k > p + 1).long()
 984             )
 985             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
 986                 None, :
 987             ]
 988             p = (
 989                 ((self.test_input == self.t_prog).long() * k)
 990                 .max(1, keepdim=True)
 991                 .values
 992             )
 993             self.test_input = (
 994                 self.test_input * (k <= p).long()
 995                 + self.t_end * (k == p + 1).long()
 996                 + self.t_nul * (k > p + 1).long()
 997             )
 998
 999         if logger is not None:
1000             logger(f"value_max {val_max}")
1001             for x in self.train_input[:25]:
1002                 end = (x != self.t_nul).nonzero().max().item() + 1
1003                 seq = [self.id2token[i.item()] for i in x[:end]]
1004                 s = " ".join(seq)
1005                 logger(f"example_seq {s}")
1006
1007         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1008
1009     def batches(self, split="train", nb_to_use=-1, desc=None):
1010         assert split in {"train", "test"}
1011         input = self.train_input if split == "train" else self.test_input
1012         if nb_to_use > 0:
1013             input = input[:nb_to_use]
1014         if desc is None:
1015             desc = f"epoch-{split}"
1016         for batch in tqdm.tqdm(
1017             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1018         ):
1019             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1020             batch = batch[:, :last].to(self.device)
1021             yield batch
1022
1023     def vocabulary_size(self):
1024         return self.nb_codes
1025
1026     def produce_results(
1027         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1028     ):
1029         # --------------------------------------------------------------------
1030         def compute_nb_errors_prog(input, nb_to_log=0):
1031             result = input.clone()
1032             s = (result == self.t_prog).long()
1033             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1034             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1035
1036             masked_inplace_autoregression(
1037                 model,
1038                 self.batch_size,
1039                 result,
1040                 ar_mask,
1041                 deterministic_synthesis,
1042                 device=self.device,
1043             )
1044
1045             sum_nb_total, sum_nb_errors = 0, 0
1046             for one_input, one_result in zip(input, result):
1047                 seq = [self.id2token[i.item()] for i in one_result]
1048                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1049                 sum_nb_total += 1
1050                 sum_nb_errors += 0 if nb_errors == 0 else 1
1051                 if nb_to_log > 0:
1052                     gt_seq = [self.id2token[i.item()] for i in one_input]
1053                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1054                     gt_prog = " ".join([str(x) for x in gt_prog])
1055                     prog = " ".join([str(x) for x in prog])
1056                     comment = "*" if nb_errors == 0 else "-"
1057                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1058                     for start_stack, target_stack, result_stack, correct in stacks:
1059                         comment = "*" if correct else "-"
1060                         start_stack = " ".join([str(x) for x in start_stack])
1061                         target_stack = " ".join([str(x) for x in target_stack])
1062                         result_stack = " ".join([str(x) for x in result_stack])
1063                         logger(
1064                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1065                         )
1066                     nb_to_log -= 1
1067
1068             return sum_nb_total, sum_nb_errors
1069
1070         # --------------------------------------------------------------------
1071         def compute_nb_errors_output(input, nb_to_log=0):
1072             result = input.clone()
1073             k = torch.arange(result.size(1), device=result.device)[None, :]
1074             last_output_idx = (
1075                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1076             )
1077             first_prog_idx = (
1078                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1079             )
1080             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1081             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1082
1083             masked_inplace_autoregression(
1084                 model,
1085                 self.batch_size,
1086                 result,
1087                 ar_mask,
1088                 deterministic_synthesis,
1089                 device=self.device,
1090             )
1091
1092             sum_nb_total, sum_nb_errors = 0, 0
1093             for one_input, one_result, i, j in zip(
1094                 input, result, last_output_idx, first_prog_idx
1095             ):
1096                 seq = [self.id2token[i.item()] for i in one_result]
1097                 sum_nb_total += 1
1098                 correct = (one_input - one_result).abs().max() == 0
1099                 sum_nb_errors += 0 if correct else 1
1100                 if nb_to_log > 0:
1101                     result_stack = [
1102                         self.id2token[i.item()] for i in one_result[i : j + 1]
1103                     ]
1104                     target_stack = [
1105                         self.id2token[i.item()] for i in one_input[i : j + 1]
1106                     ]
1107                     comment = "*" if correct else "-"
1108                     result_stack = " ".join([str(x) for x in result_stack])
1109                     target_stack = " ".join([str(x) for x in target_stack])
1110                     logger(
1111                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1112                     )
1113                     nb_to_log -= 1
1114
1115             return sum_nb_total, sum_nb_errors
1116
1117         # --------------------------------------------------------------------
1118
1119         if not self.no_prog:
1120             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1121                 self.test_input[:1000].to(self.device), nb_to_log=10
1122             )
1123
1124             logger(
1125                 f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1126             )
1127
1128         test_nb_total, test_nb_errors = compute_nb_errors_output(
1129             self.test_input[:1000].to(self.device), nb_to_log=10
1130         )
1131
1132         logger(
1133             f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1134         )
1135
1136         if save_attention_image is not None:
1137             ns=torch.randint(self.test_input.size(0),(1,)).item()
1138             input = self.test_input[ns:ns+1].clone()
1139             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1140             input = input[:, :last].to(self.device)
1141
1142             with torch.autograd.no_grad():
1143                 t = model.training
1144                 model.eval()
1145                 model.record_attention(True)
1146                 model(BracketedSequence(input))
1147                 model.train(t)
1148                 ram = model.retrieve_attention()
1149                 model.record_attention(False)
1150
1151             tokens_output = [self.id2token[i.item()] for i in input[0]]
1152             tokens_input = ["n/a"] + tokens_output[:-1]
1153             for n_head in range(ram[0].size(1)):
1154                 filename = os.path.join(
1155                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1156                 )
1157                 attention_matrices = [m[0, n_head] for m in ram]
1158                 save_attention_image(
1159                     filename,
1160                     tokens_input,
1161                     tokens_output,
1162                     attention_matrices,
1163                     k_top=10,
1164                     # min_total_attention=0.9,
1165                     token_gap=12,
1166                     layer_gap=50,
1167                 )
1168                 logger(f"wrote {filename}")
1169
1170
1171 ######################################################################
1172
1173
1174 import expr
1175
1176
1177 class Expr(Task):
1178     def tensorize(self, sequences):
1179         len_max = max([len(x) for x in sequences])
1180         return torch.cat(
1181             [
1182                 torch.tensor(
1183                     [
1184                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1185                         for s in sequences
1186                     ]
1187                 )
1188             ],
1189             0,
1190         ).to(self.device)
1191
1192     def __init__(
1193         self,
1194         nb_train_samples,
1195         nb_test_samples,
1196         nb_variables,
1197         sequence_length,
1198         operand_max,
1199         result_max,
1200         batch_size,
1201         device=torch.device("cpu"),
1202     ):
1203         super().__init__()
1204
1205         self.batch_size = batch_size
1206         self.device = device
1207
1208         train_sequences = expr.generate_sequences(
1209             nb_train_samples,
1210             nb_variables=nb_variables,
1211             length=sequence_length,
1212             operand_max=operand_max,
1213             result_max=result_max,
1214         )
1215
1216         test_sequences = expr.generate_sequences(
1217             nb_test_samples,
1218             nb_variables=nb_variables,
1219             length=sequence_length,
1220             operand_max=operand_max,
1221             result_max=result_max,
1222         )
1223
1224         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1225         symbols.sort()
1226
1227         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1228         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1229
1230         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1231
1232         self.train_input = self.tensorize(train_sequences)
1233         self.test_input = self.tensorize(test_sequences)
1234
1235         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1236
1237     def batches(self, split="train", nb_to_use=-1, desc=None):
1238         assert split in {"train", "test"}
1239         input = self.train_input if split == "train" else self.test_input
1240         if nb_to_use > 0:
1241             input = input[:nb_to_use]
1242         if desc is None:
1243             desc = f"epoch-{split}"
1244         for batch in tqdm.tqdm(
1245             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1246         ):
1247             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1248             batch = batch[:, :last]
1249             yield batch
1250
1251     def vocabulary_size(self):
1252         return self.nb_codes
1253
1254     def seq2str(self, s):
1255         return "".join([self.id2char[k.item()] for k in s])
1256
1257     def produce_results(
1258         self,
1259         n_epoch,
1260         model,
1261         result_dir,
1262         logger,
1263         deterministic_synthesis,
1264         input_file=None,
1265     ):
1266         def compute_nb_correct(input):
1267             result = input.clone()
1268             s = (result == self.space).long()
1269             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1270             result = (1 - ar_mask) * result + ar_mask * self.filler
1271             masked_inplace_autoregression(
1272                 model,
1273                 self.batch_size,
1274                 result,
1275                 ar_mask,
1276                 deterministic_synthesis,
1277                 device=self.device,
1278             )
1279
1280             nb_total = input.size(0)
1281             nb_correct = (input == result).long().min(1).values.sum()
1282
1283             #######################################################################
1284             # Comput predicted vs. true variable values
1285
1286             nb_delta = torch.zeros(5, dtype=torch.int64)
1287             nb_missed = 0
1288
1289             values_input = expr.extract_results([self.seq2str(s) for s in input])
1290             values_result = expr.extract_results([self.seq2str(s) for s in result])
1291
1292             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1293
1294             with open(filename, "w") as f:
1295                 for i, r in zip(values_input, values_result):
1296                     for n, vi in i.items():
1297                         vr = r.get(n)
1298                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1299
1300                         if vr is None or vr < 0:
1301                             nb_missed += 1
1302                         else:
1303                             d = abs(vr - vi)
1304                             if d >= nb_delta.size(0):
1305                                 nb_missed += 1
1306                             else:
1307                                 nb_delta[d] += 1
1308
1309             ######################################################################
1310
1311             return nb_total, nb_correct, nb_delta, nb_missed
1312
1313         (
1314             test_nb_total,
1315             test_nb_correct,
1316             test_nb_delta,
1317             test_nb_missed,
1318         ) = compute_nb_correct(self.test_input[:10000])
1319
1320         logger(
1321             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
1322         )
1323
1324         nb_total = test_nb_delta.sum() + test_nb_missed
1325         for d in range(test_nb_delta.size(0)):
1326             logger(
1327                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1328             )
1329         logger(
1330             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1331         )
1332
1333         ##############################################################
1334         # Log a few generated sequences
1335         if input_file is None:
1336             input = self.test_input[:10]
1337         else:
1338             with open(input_file, "r") as f:
1339                 sequences = [e.strip() for e in f.readlines()]
1340                 sequences = [s + " " + "#" * 50 for s in sequences]
1341                 input = self.tensorize(sequences)
1342
1343         result = input.clone()
1344         s = (result == self.space).long()
1345         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1346         result = (1 - ar_mask) * result + ar_mask * self.filler
1347
1348         for n in range(result.size(0)):
1349             logger(f"test_before {self.seq2str(result[n])}")
1350
1351         masked_inplace_autoregression(
1352             model,
1353             self.batch_size,
1354             result,
1355             ar_mask,
1356             deterministic_synthesis,
1357             device=self.device,
1358         )
1359
1360         correct = (1 - ar_mask) * self.space + ar_mask * input
1361         for n in range(result.size(0)):
1362             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1363             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1364             logger(f"truth       {self.seq2str(correct[n])}")
1365         ##############################################################
1366
1367
1368 ######################################################################
1369
1370 import world
1371
1372
1373 class World(Task):
1374     def __init__(
1375         self,
1376         nb_train_samples,
1377         nb_test_samples,
1378         batch_size,
1379         vqae_nb_epochs,
1380         logger=None,
1381         device=torch.device("cpu"),
1382         device_storage=torch.device("cpu"),
1383     ):
1384         super().__init__()
1385
1386         self.batch_size = batch_size
1387         self.device = device
1388
1389         (
1390             train_frames,
1391             train_action_seq,
1392             test_frames,
1393             test_action_seq,
1394             self.frame2seq,
1395             self.seq2frame,
1396         ) = world.create_data_and_processors(
1397             nb_train_samples,
1398             nb_test_samples,
1399             mode="first_last",
1400             nb_steps=30,
1401             nb_epochs=vqae_nb_epochs,
1402             logger=logger,
1403             device=device,
1404             device_storage=device_storage,
1405         )
1406
1407         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
1408         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
1409
1410         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
1411         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
1412
1413         self.len_frame_seq = train_frame_seq.size(1)
1414         self.len_action_seq = train_action_seq.size(1)
1415         self.nb_codes = nb_frame_codes + nb_action_codes
1416
1417         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
1418
1419         train_action_seq += nb_frame_codes
1420         self.train_input = torch.cat(
1421             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
1422         )
1423
1424         test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
1425         test_action_seq += nb_frame_codes
1426         self.test_input = torch.cat(
1427             (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
1428         )
1429
1430     def batches(self, split="train", nb_to_use=-1, desc=None):
1431         assert split in {"train", "test"}
1432         input = self.train_input if split == "train" else self.test_input
1433         if nb_to_use > 0:
1434             input = input[:nb_to_use]
1435         if desc is None:
1436             desc = f"epoch-{split}"
1437         for batch in tqdm.tqdm(
1438             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1439         ):
1440             yield batch.to(self.device)
1441
1442     def vocabulary_size(self):
1443         return self.nb_codes
1444
1445     def produce_results(
1446         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1447     ):
1448         k = torch.arange(
1449             2 * self.len_frame_seq + self.len_action_seq, device=self.device
1450         )[None, :]
1451
1452         input = self.test_input[:64].to(self.device)
1453         result = input.clone()
1454
1455         ar_mask = (
1456             (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
1457         )
1458         result *= 1 - ar_mask
1459
1460         masked_inplace_autoregression(
1461             model,
1462             self.batch_size,
1463             result,
1464             ar_mask,
1465             deterministic_synthesis,
1466             device=self.device,
1467         )
1468
1469         seq_start = input[:, : self.len_frame_seq]
1470         seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
1471         seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
1472
1473         result = torch.cat(
1474             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
1475         )
1476         result = result.reshape(-1, result.size(-1))
1477
1478         frames = self.seq2frame(result)
1479         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1480         torchvision.utils.save_image(
1481             frames.float() / (world.Box.nb_rgb_levels - 1),
1482             image_name,
1483             nrow=12,
1484             padding=1,
1485             pad_value=0.0,
1486         )
1487         logger(f"wrote {image_name}")
1488
1489
1490 ######################################################################