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
  80 class SandBox(Task):
  81     def __init__(
  82         self,
  83         problem,
  84         nb_train_samples,
  85         nb_test_samples,
  86         batch_size,
  87         logger=None,
  88         device=torch.device("cpu"),
  89         max_nb_codes=1024,
  90     ):
  91         super().__init__()
  92
  93         self.batch_size = batch_size
  94         self.device = device
  95         self.problem = problem
  96
  97         self.train_input, self.train_ar_mask = self.problem.generate_sequences(
  98             nb_train_samples
  99         )
 100         self.test_input, self.test_ar_mask = self.problem.generate_sequences(
 101             nb_test_samples
 102         )
 103
 104         self.train_input, self.train_ar_mask = self.train_input.to(
 105             device
 106         ), self.train_ar_mask.to(device)
 107         self.test_input, self.test_ar_mask = self.test_input.to(
 108             device
 109         ), self.test_ar_mask.to(device)
 110
 111         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 112
 113         # A bit of paranoia never hurts
 114         assert (
 115             self.nb_codes <= max_nb_codes
 116             and self.train_input.min() >= 0
 117             and self.test_input.min() >= 0
 118             and tuple(self.train_ar_mask.unique()) == (0, 1)
 119             and tuple(self.test_ar_mask.unique()) == (0, 1)
 120         )
 121
 122     def batches(self, split="train", nb_to_use=-1, desc=None):
 123         assert split in {"train", "test"}
 124         input = self.train_input if split == "train" else self.test_input
 125         if nb_to_use > 0:
 126             input = input[:nb_to_use]
 127         if desc is None:
 128             desc = f"epoch-{split}"
 129         for batch in tqdm.tqdm(
 130             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 131         ):
 132             yield batch
 133
 134     def vocabulary_size(self):
 135         return self.nb_codes
 136
 137     def produce_results(
 138         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 139     ):
 140         def compute_accuracy(input, ar_mask, logger=None):
 141             input, ar_mask = input[:nmax], ar_mask[:nmax]
 142             result = input.clone() * (1 - ar_mask)
 143
 144             masked_inplace_autoregression(
 145                 model,
 146                 self.batch_size,
 147                 result,
 148                 ar_mask,
 149                 deterministic_synthesis,
 150                 progress_bar_desc=None,
 151                 device=self.device,
 152             )
 153
 154             if logger is not None:
 155                 for sp, st in zip(result[:10], input[:10]):
 156                     logger(
 157                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 158                     )
 159                     logger(
 160                         f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 161                     )
 162
 163             nb_total = ar_mask.sum().item()
 164             nb_correct = ((result == input).long() * ar_mask).sum().item()
 165
 166             return nb_total, nb_correct
 167
 168         train_nb_total, train_nb_correct = compute_accuracy(
 169             self.train_input, self.train_ar_mask
 170         )
 171
 172         logger(
 173             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}%"
 174         )
 175
 176         test_nb_total, test_nb_correct = compute_accuracy(
 177             self.test_input, self.test_ar_mask, logger
 178         )
 179
 180         logger(
 181             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}%"
 182         )
 183
 184
 185 ######################################################################
 186
 187 import picoclvr
 188
 189
 190 class PicoCLVR(Task):
 191     # Make a tensor from a list of strings
 192     def tensorize(self, descr):
 193         token_descr = [s.strip().split(" ") for s in descr]
 194         l = max([len(s) for s in token_descr])
 195         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 196         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 197         return torch.tensor(id_descr, device=self.device)
 198
 199     # Make a list of strings from a tensor
 200     def detensorize(self, x):
 201         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 202
 203     # trim all the tensors in the tuple z to remove as much token from
 204     # left and right in the first tensor. If z is a tuple, all its
 205     # elements are trimed according to the triming for the first
 206     def trim(self, z, token="<nul>"):
 207         n = self.token2id[token]
 208         if type(z) == tuple:
 209             x = z[0]
 210             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 211             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 212             return tuple([t[:, a:b] for t in z])
 213         else:
 214             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 215             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 216             return z[:, a:b]
 217
 218     ######################
 219
 220     def __init__(
 221         self,
 222         nb_train_samples,
 223         nb_test_samples,
 224         batch_size,
 225         height,
 226         width,
 227         nb_colors=5,
 228         logger=None,
 229         device=torch.device("cpu"),
 230         pruner_train=None,
 231         pruner_eval=None,
 232     ):
 233         super().__init__()
 234
 235         def generate_descr(nb, cache_suffix, pruner):
 236             return picoclvr.generate(
 237                 nb,
 238                 height=self.height,
 239                 width=self.width,
 240                 nb_colors=nb_colors,
 241                 pruner=pruner,
 242             )
 243
 244         self.height = height
 245         self.width = width
 246         self.batch_size = batch_size
 247         self.device = device
 248         self.pruner_train = pruner_train
 249         self.pruner_eval = pruner_eval
 250
 251         if logger is not None:
 252             logger(
 253                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 254             )
 255
 256         self.train_descr = generate_descr(
 257             nb_train_samples, "train", pruner=self.pruner_train
 258         )
 259         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 260
 261         # Build the tokenizer
 262         tokens = {"<nul>", "<img>"}
 263         for d in [self.train_descr, self.test_descr]:
 264             for s in d:
 265                 for t in s.strip().split(" "):
 266                     tokens.add(t)
 267         # make this set a sorted list to get the same tensors given
 268         # the same descr
 269         tokens = list(tokens)
 270         tokens.sort()
 271         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 272         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 273         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 274
 275         # Tokenize the train and test sets
 276         self.train_input = self.tensorize(self.train_descr)
 277         self.test_input = self.tensorize(self.test_descr)
 278
 279     def batches(self, split="train"):
 280         assert split in {"train", "test"}
 281         input = self.train_input if split == "train" else self.test_input
 282         for batch in tqdm.tqdm(
 283             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 284         ):
 285             yield self.trim(batch)
 286
 287     def vocabulary_size(self):
 288         return len(self.token2id)
 289
 290     def compute_missing_properties(
 291         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 292     ):
 293         acc_nb_requested_properties = []
 294         acc_nb_missing_properties = []
 295         acc_nb_results = 0
 296
 297         for input in tqdm.tqdm(
 298             self.test_input.split(self.batch_size),
 299             dynamic_ncols=True,
 300             desc=f"test-properties",
 301         ):
 302             result = input.clone()
 303             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 304             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 305             masked_inplace_autoregression(
 306                 model,
 307                 self.batch_size,
 308                 result,
 309                 ar_mask,
 310                 deterministic_synthesis,
 311                 progress_bar_desc=None,
 312                 device=self.device,
 313             )
 314
 315             result_descr = self.detensorize(result)
 316             np = picoclvr.nb_properties(
 317                 result_descr,
 318                 height=self.height,
 319                 width=self.width,
 320                 pruner=pruner,
 321             )
 322             nb_requested_properties, _, nb_missing_properties = zip(*np)
 323             acc_nb_requested_properties += nb_requested_properties
 324             acc_nb_missing_properties += nb_missing_properties
 325             acc_nb_results += len(result_descr)
 326
 327         nb_requested_properties = sum(acc_nb_requested_properties)
 328         nb_missing_properties = sum(acc_nb_missing_properties)
 329
 330         prefix = "" if pruner is None else "pruned_"
 331         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 332         logger(
 333             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 334         )
 335         logger(
 336             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 337         )
 338
 339     ######################################################################
 340
 341     def produce_results(
 342         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 343     ):
 344         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 345
 346         if self.pruner_eval is not None:
 347             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 348
 349         nb_tokens_to_generate = self.height * self.width + 3
 350         result_descr = []
 351         nb_per_primer = 8
 352         primer = []
 353
 354         for primer_descr in [
 355             "red above green <sep> green top <sep> blue right of red",
 356             "there is red <sep> there is yellow <sep> there is blue",
 357             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 358             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 359         ]:
 360             primer += [primer_descr + " <img>"] * nb_per_primer
 361
 362         result = self.tensorize(primer)
 363         fill = result.new_full(
 364             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 365         )
 366         result = torch.cat((result, fill), 1)
 367         ar_mask = (result == self.t_nul).long()
 368         masked_inplace_autoregression(
 369             model,
 370             self.batch_size,
 371             result,
 372             ar_mask,
 373             deterministic_synthesis,
 374             device=self.device,
 375         )
 376         result_descr = self.detensorize(result)
 377
 378         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 379
 380         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 381         acc_nb_results = len(result_descr)
 382
 383         nb_requested_properties = sum(acc_nb_requested_properties)
 384         nb_missing_properties = sum(acc_nb_missing_properties)
 385
 386         prefix = "demo_"
 387         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 388         logger(
 389             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 390         )
 391         logger(
 392             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 393         )
 394
 395         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 396
 397         if img.dim() == 5:
 398             if img.size(1) == 1:
 399                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 400             else:
 401                 img = torch.cat(
 402                     [
 403                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 404                         for x in img
 405                     ],
 406                     0,
 407                 )
 408
 409         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 410         torchvision.utils.save_image(
 411             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 412         )
 413         logger(f"wrote {image_name}")
 414
 415
 416 ######################################################################
 417
 418
 419 class MNIST(Task):
 420     def __init__(
 421         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 422     ):
 423         super().__init__()
 424
 425         self.nb_train_samples = (nb_train_samples,)
 426         self.nb_test_samples = (nb_test_samples,)
 427         self.batch_size = batch_size
 428         self.device = device
 429         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 430         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 431         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 432         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 433
 434     def batches(self, split="train", nb_to_use=-1, desc=None):
 435         assert split in {"train", "test"}
 436         input = self.train_input if split == "train" else self.test_input
 437         if nb_to_use > 0:
 438             input = input[:nb_to_use]
 439         if desc is None:
 440             desc = f"epoch-{split}"
 441         for batch in tqdm.tqdm(
 442             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 443         ):
 444             yield batch
 445
 446     def vocabulary_size(self):
 447         return 256
 448
 449     def produce_results(
 450         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 451     ):
 452         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 453         ar_mask = torch.full_like(results, 1)
 454         masked_inplace_autoregression(
 455             model,
 456             self.batch_size,
 457             results,
 458             ar_mask,
 459             deterministic_synthesis,
 460             device=self.device,
 461         )
 462         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 463         torchvision.utils.save_image(
 464             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 465             image_name,
 466             nrow=16,
 467             pad_value=0.8,
 468         )
 469         logger(f"wrote {image_name}")
 470
 471
 472 ######################################################################
 473
 474 import maze
 475
 476
 477 class Maze(Task):
 478     def map2seq(self, *m):
 479         return torch.cat([x.flatten(1) for x in m], 1)
 480
 481     def seq2map(self, s):
 482         s = s.reshape(s.size(0), -1, self.height, self.width)
 483         return (s[:, k] for k in range(s.size(1)))
 484
 485     def __init__(
 486         self,
 487         nb_train_samples,
 488         nb_test_samples,
 489         batch_size,
 490         height,
 491         width,
 492         nb_walls,
 493         device=torch.device("cpu"),
 494     ):
 495         super().__init__()
 496
 497         self.batch_size = batch_size
 498         self.height = height
 499         self.width = width
 500         self.device = device
 501
 502         train_mazes, train_paths, _ = maze.create_maze_data(
 503             nb_train_samples,
 504             height=height,
 505             width=width,
 506             nb_walls=nb_walls,
 507             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 508         )
 509         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 510
 511         test_mazes, test_paths, _ = maze.create_maze_data(
 512             nb_test_samples,
 513             height=height,
 514             width=width,
 515             nb_walls=nb_walls,
 516             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 517         )
 518         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 519
 520         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 521
 522     def batches(self, split="train", nb_to_use=-1, desc=None):
 523         assert split in {"train", "test"}
 524         input = self.train_input if split == "train" else self.test_input
 525         if nb_to_use > 0:
 526             input = input[:nb_to_use]
 527         if desc is None:
 528             desc = f"epoch-{split}"
 529         for batch in tqdm.tqdm(
 530             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 531         ):
 532             yield batch
 533
 534     def vocabulary_size(self):
 535         return self.nb_codes
 536
 537     def compute_error(
 538         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 539     ):
 540         nb_total, nb_correct = 0, 0
 541         count = torch.zeros(
 542             self.width * self.height,
 543             self.width * self.height,
 544             device=self.device,
 545             dtype=torch.int64,
 546         )
 547
 548         for input in self.batches(split, nb_to_use):
 549             result = input.clone()
 550             ar_mask = result.new_zeros(result.size())
 551             ar_mask[:, self.height * self.width :] = 1
 552             result *= 1 - ar_mask
 553             masked_inplace_autoregression(
 554                 model,
 555                 self.batch_size,
 556                 result,
 557                 ar_mask,
 558                 deterministic_synthesis,
 559                 progress_bar_desc=None,
 560                 device=self.device,
 561             )
 562             mazes, paths = self.seq2map(result)
 563             path_correctness = maze.path_correctness(mazes, paths)
 564             nb_correct += path_correctness.long().sum()
 565             nb_total += mazes.size(0)
 566
 567             optimal_path_lengths = (
 568                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 569             )
 570             predicted_path_lengths = (
 571                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 572             )
 573             optimal_path_lengths = optimal_path_lengths[path_correctness]
 574             predicted_path_lengths = predicted_path_lengths[path_correctness]
 575             count[optimal_path_lengths, predicted_path_lengths] += 1
 576
 577         if count.max() == 0:
 578             count = None
 579         else:
 580             count = count[
 581                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 582             ]
 583
 584         return nb_total, nb_correct, count
 585
 586     def produce_results(
 587         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 588     ):
 589         train_nb_total, train_nb_correct, count = self.compute_error(
 590             model,
 591             "train",
 592             nb_to_use=1000,
 593             deterministic_synthesis=deterministic_synthesis,
 594         )
 595         logger(
 596             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}%"
 597         )
 598
 599         test_nb_total, test_nb_correct, count = self.compute_error(
 600             model,
 601             "test",
 602             nb_to_use=1000,
 603             deterministic_synthesis=deterministic_synthesis,
 604         )
 605         logger(
 606             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}%"
 607         )
 608
 609         if count is not None:
 610             proportion_optimal = count.diagonal().sum().float() / count.sum()
 611             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 612             with open(
 613                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 614             ) as f:
 615                 for i in range(count.size(0)):
 616                     for j in range(count.size(1)):
 617                         eol = " " if j < count.size(1) - 1 else "\n"
 618                         f.write(f"{count[i,j]}{eol}")
 619
 620         input = self.test_input[:48]
 621         result = input.clone()
 622         ar_mask = result.new_zeros(result.size())
 623         ar_mask[:, self.height * self.width :] = 1
 624         result *= 1 - ar_mask
 625         masked_inplace_autoregression(
 626             model,
 627             self.batch_size,
 628             result,
 629             ar_mask,
 630             deterministic_synthesis,
 631             device=self.device,
 632         )
 633
 634         mazes, paths = self.seq2map(input)
 635         _, predicted_paths = self.seq2map(result)
 636
 637         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 638         maze.save_image(
 639             filename,
 640             mazes=mazes,
 641             target_paths=paths,
 642             predicted_paths=predicted_paths,
 643             path_correct=maze.path_correctness(mazes, predicted_paths),
 644             path_optimal=maze.path_optimality(paths, predicted_paths),
 645         )
 646         logger(f"wrote {filename}")
 647
 648
 649 ######################################################################
 650
 651
 652 import snake
 653
 654
 655 class Snake(Task):
 656     def __init__(
 657         self,
 658         nb_train_samples,
 659         nb_test_samples,
 660         batch_size,
 661         height,
 662         width,
 663         nb_colors,
 664         length,
 665         prompt_length,
 666         device=torch.device("cpu"),
 667     ):
 668         super().__init__()
 669
 670         self.batch_size = batch_size
 671         self.height = height
 672         self.width = width
 673         self.device = device
 674         self.prompt_length = prompt_length
 675
 676         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 677             nb_train_samples,
 678             height,
 679             width,
 680             nb_colors,
 681             length,
 682             prompt_length,
 683             self.device,
 684         )
 685         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 686             nb_test_samples,
 687             height,
 688             width,
 689             nb_colors,
 690             length,
 691             prompt_length,
 692             self.device,
 693         )
 694
 695         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 696
 697     def batches(self, split="train", nb_to_use=-1, desc=None):
 698         assert split in {"train", "test"}
 699         input = self.train_input if split == "train" else self.test_input
 700         if nb_to_use > 0:
 701             input = input[:nb_to_use]
 702         if desc is None:
 703             desc = f"epoch-{split}"
 704         for batch in tqdm.tqdm(
 705             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 706         ):
 707             yield batch
 708
 709     def vocabulary_size(self):
 710         return self.nb_codes
 711
 712     def produce_results(
 713         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 714     ):
 715         def compute_nb_correct(input, prior_visits):
 716             result = input.clone()
 717             i = torch.arange(result.size(1), device=result.device)[None, :]
 718             ar_mask = (
 719                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 720                 .long()
 721                 .expand_as(result)
 722             )
 723             result *= 1 - ar_mask
 724
 725             masked_inplace_autoregression(
 726                 model,
 727                 self.batch_size,
 728                 result,
 729                 ar_mask,
 730                 deterministic_synthesis,
 731                 device=self.device,
 732             )
 733
 734             nb_total = ((prior_visits > 0) * ar_mask).sum()
 735
 736             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 737
 738             return nb_total, nb_correct
 739
 740         test_nb_total, test_nb_correct = compute_nb_correct(
 741             self.test_input[:1000], self.test_prior_visits[:1000]
 742         )
 743
 744         logger(
 745             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}%"
 746         )
 747
 748
 749 ######################################################################
 750
 751
 752 import stack
 753
 754
 755 class Stack(Task):
 756     def __init__(
 757         self,
 758         nb_train_samples,
 759         nb_test_samples,
 760         batch_size,
 761         logger,
 762         nb_steps,
 763         nb_stacks,
 764         nb_digits,
 765         fraction_values_for_train=None,
 766         device=torch.device("cpu"),
 767     ):
 768         super().__init__()
 769
 770         self.batch_size = batch_size
 771         self.nb_steps = nb_steps
 772         self.nb_stacks = nb_stacks
 773         self.nb_digits = nb_digits
 774         self.device = device
 775
 776         if fraction_values_for_train is None:
 777             values_for_train = None
 778             values_for_test = None
 779         else:
 780             all = torch.randperm(10**nb_digits)
 781             nb_for_train = int(all.size(0) * fraction_values_for_train)
 782             values_for_train = all[:nb_for_train]
 783             values_for_test = all[nb_for_train:]
 784
 785         self.train_input, self.train_stack_counts = stack.generate_sequences(
 786             nb_train_samples,
 787             nb_steps,
 788             nb_stacks,
 789             nb_digits,
 790             values_for_train,
 791             self.device,
 792         )
 793
 794         self.test_input, self.test_stack_counts = stack.generate_sequences(
 795             nb_test_samples,
 796             nb_steps,
 797             nb_stacks,
 798             nb_digits,
 799             values_for_test,
 800             self.device,
 801         )
 802
 803         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 804         counts = self.test_stack_counts.flatten()[i.flatten()]
 805         counts = F.one_hot(counts).sum(0)
 806         logger(f"test_pop_stack_counts {counts}")
 807
 808         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 809
 810     def batches(self, split="train", nb_to_use=-1, desc=None):
 811         assert split in {"train", "test"}
 812         input = self.train_input if split == "train" else self.test_input
 813         if nb_to_use > 0:
 814             input = input[:nb_to_use]
 815         if desc is None:
 816             desc = f"epoch-{split}"
 817         for batch in tqdm.tqdm(
 818             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 819         ):
 820             yield batch
 821
 822     def vocabulary_size(self):
 823         return self.nb_codes
 824
 825     def produce_results(
 826         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 827     ):
 828         def compute_nb_correct(input):
 829             result = input.clone()
 830             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 831             ar_mask = (result != input).long()
 832             masked_inplace_autoregression(
 833                 model,
 834                 self.batch_size,
 835                 result,
 836                 ar_mask,
 837                 deterministic_synthesis,
 838                 device=self.device,
 839             )
 840
 841             errors = ((result != input).long() * ar_mask).reshape(
 842                 -1, 1 + self.nb_digits
 843             )
 844             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 845
 846             nb_total = ar_mask.max(1).values.sum()
 847             nb_correct = nb_total - errors.max(1).values.sum()
 848
 849             return nb_total, nb_correct
 850
 851         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 852
 853         logger(
 854             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}%"
 855         )
 856
 857         ##############################################################
 858         # Log a few generated sequences
 859         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 860         result = input.clone()
 861         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 862         ar_mask = (result != input).long()
 863
 864         # for n in range(result.size(0)):
 865         # logger(
 866         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 867         # )
 868
 869         masked_inplace_autoregression(
 870             model,
 871             self.batch_size,
 872             result,
 873             ar_mask,
 874             deterministic_synthesis,
 875             device=self.device,
 876         )
 877
 878         for n in range(result.size(0)):
 879             logger(
 880                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 881             )
 882         ##############################################################
 883
 884
 885 ######################################################################
 886
 887 import rpl
 888
 889
 890 class RPL(Task):
 891     def tensorize(self, sequences):
 892         len_max = max([len(x) for x in sequences])
 893         return torch.cat(
 894             [
 895                 torch.tensor(
 896                     [
 897                         [
 898                             self.token2id[str(c)]
 899                             for c in s + ["<nul>"] * (len_max - len(s))
 900                         ]
 901                         for s in sequences
 902                     ]
 903                 )
 904             ],
 905             0,
 906         )
 907
 908     def seq2str(self, seq):
 909         return " ".join([self.id2token[i] for i in seq])
 910
 911     def __init__(
 912         self,
 913         nb_train_samples,
 914         nb_test_samples,
 915         batch_size,
 916         nb_starting_values=3,
 917         max_input=9,
 918         prog_len=6,
 919         nb_runs=5,
 920         no_prog=False,
 921         logger=None,
 922         device=torch.device("cpu"),
 923     ):
 924         super().__init__()
 925
 926         self.batch_size = batch_size
 927         self.device = device
 928         self.no_prog = no_prog
 929
 930         train_sequences = [
 931             rpl.generate(
 932                 nb_starting_values=nb_starting_values,
 933                 nb_result_values_max=4 * nb_starting_values,
 934                 max_input=max_input,
 935                 prog_len=prog_len,
 936                 nb_runs=nb_runs,
 937             )
 938             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
 939         ]
 940
 941         test_sequences = [
 942             rpl.generate(
 943                 nb_starting_values=nb_starting_values,
 944                 nb_result_values_max=4 * nb_starting_values,
 945                 max_input=max_input,
 946                 prog_len=prog_len,
 947                 nb_runs=nb_runs,
 948             )
 949             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
 950         ]
 951
 952         symbols = list(
 953             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
 954         )
 955         val_max = max([x if type(x) is int else 0 for x in symbols])
 956         symbols = list(filter(lambda x: type(x) is str, symbols))
 957         symbols.sort()
 958         symbols += [str(n) for n in range(val_max + 1)]
 959         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
 960         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
 961
 962         self.t_nul = self.token2id["<nul>"]
 963         self.t_input = self.token2id["<in>"]
 964         self.t_output = self.token2id["<out>"]
 965         self.t_prog = self.token2id["<prg>"]
 966         self.t_end = self.token2id["<end>"]
 967
 968         self.train_input = self.tensorize(train_sequences)
 969         self.test_input = self.tensorize(test_sequences)
 970
 971         if no_prog:
 972             # Excise the program from every train and test example
 973             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
 974                 None, :
 975             ]
 976             p = (
 977                 ((self.train_input == self.t_prog).long() * k)
 978                 .max(1, keepdim=True)
 979                 .values
 980             )
 981             self.train_input = (
 982                 self.train_input * (k <= p).long()
 983                 + self.t_end * (k == p + 1).long()
 984                 + self.t_nul * (k > p + 1).long()
 985             )
 986             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
 987                 None, :
 988             ]
 989             p = (
 990                 ((self.test_input == self.t_prog).long() * k)
 991                 .max(1, keepdim=True)
 992                 .values
 993             )
 994             self.test_input = (
 995                 self.test_input * (k <= p).long()
 996                 + self.t_end * (k == p + 1).long()
 997                 + self.t_nul * (k > p + 1).long()
 998             )
 999
1000         if logger is not None:
1001             logger(f"value_max {val_max}")
1002             for x in self.train_input[:25]:
1003                 end = (x != self.t_nul).nonzero().max().item() + 1
1004                 seq = [self.id2token[i.item()] for i in x[:end]]
1005                 s = " ".join(seq)
1006                 logger(f"example_seq {s}")
1007
1008         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1009
1010     def batches(self, split="train", nb_to_use=-1, desc=None):
1011         assert split in {"train", "test"}
1012         input = self.train_input if split == "train" else self.test_input
1013         if nb_to_use > 0:
1014             input = input[:nb_to_use]
1015         if desc is None:
1016             desc = f"epoch-{split}"
1017         for batch in tqdm.tqdm(
1018             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1019         ):
1020             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1021             batch = batch[:, :last].to(self.device)
1022             yield batch
1023
1024     def vocabulary_size(self):
1025         return self.nb_codes
1026
1027     def produce_results(
1028         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1029     ):
1030         # --------------------------------------------------------------------
1031         def compute_nb_errors_prog(input, nb_to_log=0):
1032             result = input.clone()
1033             s = (result == self.t_prog).long()
1034             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1035             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1036
1037             masked_inplace_autoregression(
1038                 model,
1039                 self.batch_size,
1040                 result,
1041                 ar_mask,
1042                 deterministic_synthesis,
1043                 device=self.device,
1044             )
1045
1046             sum_nb_total, sum_nb_errors = 0, 0
1047             for one_input, one_result in zip(input, result):
1048                 seq = [self.id2token[i.item()] for i in one_result]
1049                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1050                 sum_nb_total += 1
1051                 sum_nb_errors += 0 if nb_errors == 0 else 1
1052                 if nb_to_log > 0:
1053                     gt_seq = [self.id2token[i.item()] for i in one_input]
1054                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1055                     gt_prog = " ".join([str(x) for x in gt_prog])
1056                     prog = " ".join([str(x) for x in prog])
1057                     comment = "*" if nb_errors == 0 else "-"
1058                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1059                     for start_stack, target_stack, result_stack, correct in stacks:
1060                         comment = "*" if correct else "-"
1061                         start_stack = " ".join([str(x) for x in start_stack])
1062                         target_stack = " ".join([str(x) for x in target_stack])
1063                         result_stack = " ".join([str(x) for x in result_stack])
1064                         logger(
1065                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1066                         )
1067                     nb_to_log -= 1
1068
1069             return sum_nb_total, sum_nb_errors
1070
1071         # --------------------------------------------------------------------
1072         def compute_nb_errors_output(input, nb_to_log=0):
1073             result = input.clone()
1074             k = torch.arange(result.size(1), device=result.device)[None, :]
1075             last_output_idx = (
1076                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1077             )
1078             first_prog_idx = (
1079                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1080             )
1081             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1082             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1083
1084             masked_inplace_autoregression(
1085                 model,
1086                 self.batch_size,
1087                 result,
1088                 ar_mask,
1089                 deterministic_synthesis,
1090                 device=self.device,
1091             )
1092
1093             sum_nb_total, sum_nb_errors = 0, 0
1094             for one_input, one_result, i, j in zip(
1095                 input, result, last_output_idx, first_prog_idx
1096             ):
1097                 seq = [self.id2token[i.item()] for i in one_result]
1098                 sum_nb_total += 1
1099                 correct = (one_input - one_result).abs().max() == 0
1100                 sum_nb_errors += 0 if correct else 1
1101                 if nb_to_log > 0:
1102                     result_stack = [
1103                         self.id2token[i.item()] for i in one_result[i : j + 1]
1104                     ]
1105                     target_stack = [
1106                         self.id2token[i.item()] for i in one_input[i : j + 1]
1107                     ]
1108                     comment = "*" if correct else "-"
1109                     result_stack = " ".join([str(x) for x in result_stack])
1110                     target_stack = " ".join([str(x) for x in target_stack])
1111                     logger(
1112                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1113                     )
1114                     nb_to_log -= 1
1115
1116             return sum_nb_total, sum_nb_errors
1117
1118         # --------------------------------------------------------------------
1119
1120         if not self.no_prog:
1121             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1122                 self.test_input[:1000].to(self.device), nb_to_log=10
1123             )
1124
1125             logger(
1126                 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}%"
1127             )
1128
1129         test_nb_total, test_nb_errors = compute_nb_errors_output(
1130             self.test_input[:1000].to(self.device), nb_to_log=10
1131         )
1132
1133         logger(
1134             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}%"
1135         )
1136
1137         if save_attention_image is not None:
1138             ns = torch.randint(self.test_input.size(0), (1,)).item()
1139             input = self.test_input[ns : ns + 1].clone()
1140             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1141             input = input[:, :last].to(self.device)
1142
1143             with torch.autograd.no_grad():
1144                 t = model.training
1145                 model.eval()
1146                 model.record_attention(True)
1147                 model(BracketedSequence(input))
1148                 model.train(t)
1149                 ram = model.retrieve_attention()
1150                 model.record_attention(False)
1151
1152             tokens_output = [self.id2token[i.item()] for i in input[0]]
1153             tokens_input = ["n/a"] + tokens_output[:-1]
1154             for n_head in range(ram[0].size(1)):
1155                 filename = os.path.join(
1156                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1157                 )
1158                 attention_matrices = [m[0, n_head] for m in ram]
1159                 save_attention_image(
1160                     filename,
1161                     tokens_input,
1162                     tokens_output,
1163                     attention_matrices,
1164                     k_top=10,
1165                     # min_total_attention=0.9,
1166                     token_gap=12,
1167                     layer_gap=50,
1168                 )
1169                 logger(f"wrote {filename}")
1170
1171
1172 ######################################################################
1173
1174
1175 import expr
1176
1177
1178 class Expr(Task):
1179     def tensorize(self, sequences):
1180         len_max = max([len(x) for x in sequences])
1181         return torch.cat(
1182             [
1183                 torch.tensor(
1184                     [
1185                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1186                         for s in sequences
1187                     ]
1188                 )
1189             ],
1190             0,
1191         ).to(self.device)
1192
1193     def __init__(
1194         self,
1195         nb_train_samples,
1196         nb_test_samples,
1197         nb_variables,
1198         sequence_length,
1199         operand_max,
1200         result_max,
1201         batch_size,
1202         device=torch.device("cpu"),
1203     ):
1204         super().__init__()
1205
1206         self.batch_size = batch_size
1207         self.device = device
1208
1209         train_sequences = expr.generate_sequences(
1210             nb_train_samples,
1211             nb_variables=nb_variables,
1212             length=sequence_length,
1213             operand_max=operand_max,
1214             result_max=result_max,
1215         )
1216
1217         test_sequences = expr.generate_sequences(
1218             nb_test_samples,
1219             nb_variables=nb_variables,
1220             length=sequence_length,
1221             operand_max=operand_max,
1222             result_max=result_max,
1223         )
1224
1225         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1226         symbols.sort()
1227
1228         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1229         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1230
1231         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1232
1233         self.train_input = self.tensorize(train_sequences)
1234         self.test_input = self.tensorize(test_sequences)
1235
1236         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1237
1238     def batches(self, split="train", nb_to_use=-1, desc=None):
1239         assert split in {"train", "test"}
1240         input = self.train_input if split == "train" else self.test_input
1241         if nb_to_use > 0:
1242             input = input[:nb_to_use]
1243         if desc is None:
1244             desc = f"epoch-{split}"
1245         for batch in tqdm.tqdm(
1246             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1247         ):
1248             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1249             batch = batch[:, :last]
1250             yield batch
1251
1252     def vocabulary_size(self):
1253         return self.nb_codes
1254
1255     def seq2str(self, s):
1256         return "".join([self.id2char[k.item()] for k in s])
1257
1258     def produce_results(
1259         self,
1260         n_epoch,
1261         model,
1262         result_dir,
1263         logger,
1264         deterministic_synthesis,
1265         input_file=None,
1266     ):
1267         def compute_nb_correct(input):
1268             result = input.clone()
1269             s = (result == self.space).long()
1270             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1271             result = (1 - ar_mask) * result + ar_mask * self.filler
1272             masked_inplace_autoregression(
1273                 model,
1274                 self.batch_size,
1275                 result,
1276                 ar_mask,
1277                 deterministic_synthesis,
1278                 device=self.device,
1279             )
1280
1281             nb_total = input.size(0)
1282             nb_correct = (input == result).long().min(1).values.sum()
1283
1284             #######################################################################
1285             # Comput predicted vs. true variable values
1286
1287             nb_delta = torch.zeros(5, dtype=torch.int64)
1288             nb_missed = 0
1289
1290             values_input = expr.extract_results([self.seq2str(s) for s in input])
1291             values_result = expr.extract_results([self.seq2str(s) for s in result])
1292
1293             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1294
1295             with open(filename, "w") as f:
1296                 for i, r in zip(values_input, values_result):
1297                     for n, vi in i.items():
1298                         vr = r.get(n)
1299                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1300
1301                         if vr is None or vr < 0:
1302                             nb_missed += 1
1303                         else:
1304                             d = abs(vr - vi)
1305                             if d >= nb_delta.size(0):
1306                                 nb_missed += 1
1307                             else:
1308                                 nb_delta[d] += 1
1309
1310             ######################################################################
1311
1312             return nb_total, nb_correct, nb_delta, nb_missed
1313
1314         (
1315             test_nb_total,
1316             test_nb_correct,
1317             test_nb_delta,
1318             test_nb_missed,
1319         ) = compute_nb_correct(self.test_input[:10000])
1320
1321         logger(
1322             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}%"
1323         )
1324
1325         nb_total = test_nb_delta.sum() + test_nb_missed
1326         for d in range(test_nb_delta.size(0)):
1327             logger(
1328                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1329             )
1330         logger(
1331             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1332         )
1333
1334         ##############################################################
1335         # Log a few generated sequences
1336         if input_file is None:
1337             input = self.test_input[:10]
1338         else:
1339             with open(input_file, "r") as f:
1340                 sequences = [e.strip() for e in f.readlines()]
1341                 sequences = [s + " " + "#" * 50 for s in sequences]
1342                 input = self.tensorize(sequences)
1343
1344         result = input.clone()
1345         s = (result == self.space).long()
1346         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1347         result = (1 - ar_mask) * result + ar_mask * self.filler
1348
1349         for n in range(result.size(0)):
1350             logger(f"test_before {self.seq2str(result[n])}")
1351
1352         masked_inplace_autoregression(
1353             model,
1354             self.batch_size,
1355             result,
1356             ar_mask,
1357             deterministic_synthesis,
1358             device=self.device,
1359         )
1360
1361         correct = (1 - ar_mask) * self.space + ar_mask * input
1362         for n in range(result.size(0)):
1363             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1364             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1365             logger(f"truth       {self.seq2str(correct[n])}")
1366         ##############################################################
1367
1368
1369 ######################################################################
1370
1371 import world
1372
1373
1374 class World(Task):
1375     def __init__(
1376         self,
1377         nb_train_samples,
1378         nb_test_samples,
1379         batch_size,
1380         vqae_nb_epochs,
1381         logger=None,
1382         device=torch.device("cpu"),
1383         device_storage=torch.device("cpu"),
1384     ):
1385         super().__init__()
1386
1387         self.batch_size = batch_size
1388         self.device = device
1389
1390         (
1391             train_frames,
1392             train_action_seq,
1393             test_frames,
1394             test_action_seq,
1395             self.frame2seq,
1396             self.seq2frame,
1397         ) = world.create_data_and_processors(
1398             nb_train_samples,
1399             nb_test_samples,
1400             mode="first_last",
1401             nb_steps=30,
1402             nb_epochs=vqae_nb_epochs,
1403             logger=logger,
1404             device=device,
1405             device_storage=device_storage,
1406         )
1407
1408         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
1409         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
1410
1411         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
1412         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
1413
1414         self.len_frame_seq = train_frame_seq.size(1)
1415         self.len_action_seq = train_action_seq.size(1)
1416         self.nb_codes = nb_frame_codes + nb_action_codes
1417
1418         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
1419
1420         train_action_seq += nb_frame_codes
1421         self.train_input = torch.cat(
1422             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
1423         )
1424
1425         test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
1426         test_action_seq += nb_frame_codes
1427         self.test_input = torch.cat(
1428             (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
1429         )
1430
1431     def batches(self, split="train", nb_to_use=-1, desc=None):
1432         assert split in {"train", "test"}
1433         input = self.train_input if split == "train" else self.test_input
1434         if nb_to_use > 0:
1435             input = input[:nb_to_use]
1436         if desc is None:
1437             desc = f"epoch-{split}"
1438         for batch in tqdm.tqdm(
1439             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1440         ):
1441             yield batch.to(self.device)
1442
1443     def vocabulary_size(self):
1444         return self.nb_codes
1445
1446     def produce_results(
1447         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1448     ):
1449         k = torch.arange(
1450             2 * self.len_frame_seq + self.len_action_seq, device=self.device
1451         )[None, :]
1452
1453         input = self.test_input[:64].to(self.device)
1454         result = input.clone()
1455
1456         ar_mask = (
1457             (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
1458         )
1459         result *= 1 - ar_mask
1460
1461         masked_inplace_autoregression(
1462             model,
1463             self.batch_size,
1464             result,
1465             ar_mask,
1466             deterministic_synthesis,
1467             device=self.device,
1468         )
1469
1470         seq_start = input[:, : self.len_frame_seq]
1471         seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
1472         seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
1473
1474         result = torch.cat(
1475             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
1476         )
1477         result = result.reshape(-1, result.size(-1))
1478
1479         frames = self.seq2frame(result)
1480         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1481         torchvision.utils.save_image(
1482             frames.float() / (world.Box.nb_rgb_levels - 1),
1483             image_name,
1484             nrow=12,
1485             padding=1,
1486             pad_value=0.0,
1487         )
1488         logger(f"wrote {image_name}")
1489
1490
1491 ######################################################################