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         if save_attention_image is not None:
 185             for k in range(10):
 186                 ns = torch.randint(self.test_input.size(0), (1,)).item()
 187                 input = self.test_input[ns : ns + 1].clone()
 188
 189                 with torch.autograd.no_grad():
 190                     t = model.training
 191                     model.eval()
 192                     model.record_attention(True)
 193                     model(BracketedSequence(input))
 194                     model.train(t)
 195                     ram = model.retrieve_attention()
 196                     model.record_attention(False)
 197
 198                 tokens_output = [c for c in self.problem.seq2str(input[0])]
 199                 tokens_input = ["n/a"] + tokens_output[:-1]
 200                 for n_head in range(ram[0].size(1)):
 201                     filename = os.path.join(
 202                         result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
 203                     )
 204                     attention_matrices = [m[0, n_head] for m in ram]
 205                     save_attention_image(
 206                         filename,
 207                         tokens_input,
 208                         tokens_output,
 209                         attention_matrices,
 210                         k_top=10,
 211                         # min_total_attention=0.9,
 212                         token_gap=12,
 213                         layer_gap=50,
 214                     )
 215                     logger(f"wrote {filename}")
 216
 217
 218 ######################################################################
 219
 220 import picoclvr
 221
 222
 223 class PicoCLVR(Task):
 224     # Make a tensor from a list of strings
 225     def tensorize(self, descr):
 226         token_descr = [s.strip().split(" ") for s in descr]
 227         l = max([len(s) for s in token_descr])
 228         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 229         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 230         return torch.tensor(id_descr, device=self.device)
 231
 232     # Make a list of strings from a tensor
 233     def detensorize(self, x):
 234         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 235
 236     # trim all the tensors in the tuple z to remove as much token from
 237     # left and right in the first tensor. If z is a tuple, all its
 238     # elements are trimed according to the triming for the first
 239     def trim(self, z, token="<nul>"):
 240         n = self.token2id[token]
 241         if type(z) == tuple:
 242             x = z[0]
 243             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 244             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 245             return tuple([t[:, a:b] for t in z])
 246         else:
 247             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 248             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 249             return z[:, a:b]
 250
 251     ######################
 252
 253     def __init__(
 254         self,
 255         nb_train_samples,
 256         nb_test_samples,
 257         batch_size,
 258         height,
 259         width,
 260         nb_colors=5,
 261         logger=None,
 262         device=torch.device("cpu"),
 263         pruner_train=None,
 264         pruner_eval=None,
 265     ):
 266         super().__init__()
 267
 268         def generate_descr(nb, cache_suffix, pruner):
 269             return picoclvr.generate(
 270                 nb,
 271                 height=self.height,
 272                 width=self.width,
 273                 nb_colors=nb_colors,
 274                 pruner=pruner,
 275             )
 276
 277         self.height = height
 278         self.width = width
 279         self.batch_size = batch_size
 280         self.device = device
 281         self.pruner_train = pruner_train
 282         self.pruner_eval = pruner_eval
 283
 284         if logger is not None:
 285             logger(
 286                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 287             )
 288
 289         self.train_descr = generate_descr(
 290             nb_train_samples, "train", pruner=self.pruner_train
 291         )
 292         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 293
 294         # Build the tokenizer
 295         tokens = {"<nul>", "<img>"}
 296         for d in [self.train_descr, self.test_descr]:
 297             for s in d:
 298                 for t in s.strip().split(" "):
 299                     tokens.add(t)
 300         # make this set a sorted list to get the same tensors given
 301         # the same descr
 302         tokens = list(tokens)
 303         tokens.sort()
 304         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 305         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 306         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 307
 308         # Tokenize the train and test sets
 309         self.train_input = self.tensorize(self.train_descr)
 310         self.test_input = self.tensorize(self.test_descr)
 311
 312     def batches(self, split="train"):
 313         assert split in {"train", "test"}
 314         input = self.train_input if split == "train" else self.test_input
 315         for batch in tqdm.tqdm(
 316             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 317         ):
 318             yield self.trim(batch)
 319
 320     def vocabulary_size(self):
 321         return len(self.token2id)
 322
 323     def compute_missing_properties(
 324         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 325     ):
 326         acc_nb_requested_properties = []
 327         acc_nb_missing_properties = []
 328         acc_nb_results = 0
 329
 330         for input in tqdm.tqdm(
 331             self.test_input.split(self.batch_size),
 332             dynamic_ncols=True,
 333             desc=f"test-properties",
 334         ):
 335             result = input.clone()
 336             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 337             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 338             masked_inplace_autoregression(
 339                 model,
 340                 self.batch_size,
 341                 result,
 342                 ar_mask,
 343                 deterministic_synthesis,
 344                 progress_bar_desc=None,
 345                 device=self.device,
 346             )
 347
 348             result_descr = self.detensorize(result)
 349             np = picoclvr.nb_properties(
 350                 result_descr,
 351                 height=self.height,
 352                 width=self.width,
 353                 pruner=pruner,
 354             )
 355             nb_requested_properties, _, nb_missing_properties = zip(*np)
 356             acc_nb_requested_properties += nb_requested_properties
 357             acc_nb_missing_properties += nb_missing_properties
 358             acc_nb_results += len(result_descr)
 359
 360         nb_requested_properties = sum(acc_nb_requested_properties)
 361         nb_missing_properties = sum(acc_nb_missing_properties)
 362
 363         prefix = "" if pruner is None else "pruned_"
 364         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 365         logger(
 366             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 367         )
 368         logger(
 369             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 370         )
 371
 372     ######################################################################
 373
 374     def produce_results(
 375         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 376     ):
 377         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 378
 379         if self.pruner_eval is not None:
 380             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 381
 382         nb_tokens_to_generate = self.height * self.width + 3
 383         result_descr = []
 384         nb_per_primer = 8
 385         primer = []
 386
 387         for primer_descr in [
 388             "red above green <sep> green top <sep> blue right of red",
 389             "there is red <sep> there is yellow <sep> there is blue",
 390             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 391             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 392         ]:
 393             primer += [primer_descr + " <img>"] * nb_per_primer
 394
 395         result = self.tensorize(primer)
 396         fill = result.new_full(
 397             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 398         )
 399         result = torch.cat((result, fill), 1)
 400         ar_mask = (result == self.t_nul).long()
 401         masked_inplace_autoregression(
 402             model,
 403             self.batch_size,
 404             result,
 405             ar_mask,
 406             deterministic_synthesis,
 407             device=self.device,
 408         )
 409         result_descr = self.detensorize(result)
 410
 411         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 412
 413         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 414         acc_nb_results = len(result_descr)
 415
 416         nb_requested_properties = sum(acc_nb_requested_properties)
 417         nb_missing_properties = sum(acc_nb_missing_properties)
 418
 419         prefix = "demo_"
 420         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 421         logger(
 422             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 423         )
 424         logger(
 425             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 426         )
 427
 428         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 429
 430         if img.dim() == 5:
 431             if img.size(1) == 1:
 432                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 433             else:
 434                 img = torch.cat(
 435                     [
 436                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 437                         for x in img
 438                     ],
 439                     0,
 440                 )
 441
 442         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 443         torchvision.utils.save_image(
 444             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 445         )
 446         logger(f"wrote {image_name}")
 447
 448
 449 ######################################################################
 450
 451
 452 class MNIST(Task):
 453     def __init__(
 454         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 455     ):
 456         super().__init__()
 457
 458         self.nb_train_samples = (nb_train_samples,)
 459         self.nb_test_samples = (nb_test_samples,)
 460         self.batch_size = batch_size
 461         self.device = device
 462         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 463         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 464         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 465         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 466
 467     def batches(self, split="train", nb_to_use=-1, desc=None):
 468         assert split in {"train", "test"}
 469         input = self.train_input if split == "train" else self.test_input
 470         if nb_to_use > 0:
 471             input = input[:nb_to_use]
 472         if desc is None:
 473             desc = f"epoch-{split}"
 474         for batch in tqdm.tqdm(
 475             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 476         ):
 477             yield batch
 478
 479     def vocabulary_size(self):
 480         return 256
 481
 482     def produce_results(
 483         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 484     ):
 485         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 486         ar_mask = torch.full_like(results, 1)
 487         masked_inplace_autoregression(
 488             model,
 489             self.batch_size,
 490             results,
 491             ar_mask,
 492             deterministic_synthesis,
 493             device=self.device,
 494         )
 495         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 496         torchvision.utils.save_image(
 497             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 498             image_name,
 499             nrow=16,
 500             pad_value=0.8,
 501         )
 502         logger(f"wrote {image_name}")
 503
 504
 505 ######################################################################
 506
 507 import maze
 508
 509
 510 class Maze(Task):
 511     def map2seq(self, *m):
 512         return torch.cat([x.flatten(1) for x in m], 1)
 513
 514     def seq2map(self, s):
 515         s = s.reshape(s.size(0), -1, self.height, self.width)
 516         return (s[:, k] for k in range(s.size(1)))
 517
 518     def __init__(
 519         self,
 520         nb_train_samples,
 521         nb_test_samples,
 522         batch_size,
 523         height,
 524         width,
 525         nb_walls,
 526         device=torch.device("cpu"),
 527     ):
 528         super().__init__()
 529
 530         self.batch_size = batch_size
 531         self.height = height
 532         self.width = width
 533         self.device = device
 534
 535         train_mazes, train_paths, _ = maze.create_maze_data(
 536             nb_train_samples,
 537             height=height,
 538             width=width,
 539             nb_walls=nb_walls,
 540             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 541         )
 542         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 543
 544         test_mazes, test_paths, _ = maze.create_maze_data(
 545             nb_test_samples,
 546             height=height,
 547             width=width,
 548             nb_walls=nb_walls,
 549             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 550         )
 551         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 552
 553         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 554
 555     def batches(self, split="train", nb_to_use=-1, desc=None):
 556         assert split in {"train", "test"}
 557         input = self.train_input if split == "train" else self.test_input
 558         if nb_to_use > 0:
 559             input = input[:nb_to_use]
 560         if desc is None:
 561             desc = f"epoch-{split}"
 562         for batch in tqdm.tqdm(
 563             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 564         ):
 565             yield batch
 566
 567     def vocabulary_size(self):
 568         return self.nb_codes
 569
 570     def compute_error(
 571         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 572     ):
 573         nb_total, nb_correct = 0, 0
 574         count = torch.zeros(
 575             self.width * self.height,
 576             self.width * self.height,
 577             device=self.device,
 578             dtype=torch.int64,
 579         )
 580
 581         for input in self.batches(split, nb_to_use):
 582             result = input.clone()
 583             ar_mask = result.new_zeros(result.size())
 584             ar_mask[:, self.height * self.width :] = 1
 585             result *= 1 - ar_mask
 586             masked_inplace_autoregression(
 587                 model,
 588                 self.batch_size,
 589                 result,
 590                 ar_mask,
 591                 deterministic_synthesis,
 592                 progress_bar_desc=None,
 593                 device=self.device,
 594             )
 595             mazes, paths = self.seq2map(result)
 596             path_correctness = maze.path_correctness(mazes, paths)
 597             nb_correct += path_correctness.long().sum()
 598             nb_total += mazes.size(0)
 599
 600             optimal_path_lengths = (
 601                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 602             )
 603             predicted_path_lengths = (
 604                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 605             )
 606             optimal_path_lengths = optimal_path_lengths[path_correctness]
 607             predicted_path_lengths = predicted_path_lengths[path_correctness]
 608             count[optimal_path_lengths, predicted_path_lengths] += 1
 609
 610         if count.max() == 0:
 611             count = None
 612         else:
 613             count = count[
 614                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 615             ]
 616
 617         return nb_total, nb_correct, count
 618
 619     def produce_results(
 620         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 621     ):
 622         train_nb_total, train_nb_correct, count = self.compute_error(
 623             model,
 624             "train",
 625             nb_to_use=1000,
 626             deterministic_synthesis=deterministic_synthesis,
 627         )
 628         logger(
 629             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}%"
 630         )
 631
 632         test_nb_total, test_nb_correct, count = self.compute_error(
 633             model,
 634             "test",
 635             nb_to_use=1000,
 636             deterministic_synthesis=deterministic_synthesis,
 637         )
 638         logger(
 639             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}%"
 640         )
 641
 642         if count is not None:
 643             proportion_optimal = count.diagonal().sum().float() / count.sum()
 644             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 645             with open(
 646                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 647             ) as f:
 648                 for i in range(count.size(0)):
 649                     for j in range(count.size(1)):
 650                         eol = " " if j < count.size(1) - 1 else "\n"
 651                         f.write(f"{count[i,j]}{eol}")
 652
 653         input = self.test_input[:48]
 654         result = input.clone()
 655         ar_mask = result.new_zeros(result.size())
 656         ar_mask[:, self.height * self.width :] = 1
 657         result *= 1 - ar_mask
 658         masked_inplace_autoregression(
 659             model,
 660             self.batch_size,
 661             result,
 662             ar_mask,
 663             deterministic_synthesis,
 664             device=self.device,
 665         )
 666
 667         mazes, paths = self.seq2map(input)
 668         _, predicted_paths = self.seq2map(result)
 669
 670         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 671         maze.save_image(
 672             filename,
 673             mazes=mazes,
 674             target_paths=paths,
 675             predicted_paths=predicted_paths,
 676             path_correct=maze.path_correctness(mazes, predicted_paths),
 677             path_optimal=maze.path_optimality(paths, predicted_paths),
 678         )
 679         logger(f"wrote {filename}")
 680
 681
 682 ######################################################################
 683
 684
 685 import snake
 686
 687
 688 class Snake(Task):
 689     def __init__(
 690         self,
 691         nb_train_samples,
 692         nb_test_samples,
 693         batch_size,
 694         height,
 695         width,
 696         nb_colors,
 697         length,
 698         prompt_length,
 699         device=torch.device("cpu"),
 700     ):
 701         super().__init__()
 702
 703         self.batch_size = batch_size
 704         self.height = height
 705         self.width = width
 706         self.device = device
 707         self.prompt_length = prompt_length
 708
 709         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 710             nb_train_samples,
 711             height,
 712             width,
 713             nb_colors,
 714             length,
 715             prompt_length,
 716             self.device,
 717         )
 718         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 719             nb_test_samples,
 720             height,
 721             width,
 722             nb_colors,
 723             length,
 724             prompt_length,
 725             self.device,
 726         )
 727
 728         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 729
 730     def batches(self, split="train", nb_to_use=-1, desc=None):
 731         assert split in {"train", "test"}
 732         input = self.train_input if split == "train" else self.test_input
 733         if nb_to_use > 0:
 734             input = input[:nb_to_use]
 735         if desc is None:
 736             desc = f"epoch-{split}"
 737         for batch in tqdm.tqdm(
 738             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 739         ):
 740             yield batch
 741
 742     def vocabulary_size(self):
 743         return self.nb_codes
 744
 745     def produce_results(
 746         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 747     ):
 748         def compute_nb_correct(input, prior_visits):
 749             result = input.clone()
 750             i = torch.arange(result.size(1), device=result.device)[None, :]
 751             ar_mask = (
 752                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 753                 .long()
 754                 .expand_as(result)
 755             )
 756             result *= 1 - ar_mask
 757
 758             masked_inplace_autoregression(
 759                 model,
 760                 self.batch_size,
 761                 result,
 762                 ar_mask,
 763                 deterministic_synthesis,
 764                 device=self.device,
 765             )
 766
 767             nb_total = ((prior_visits > 0) * ar_mask).sum()
 768
 769             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 770
 771             return nb_total, nb_correct
 772
 773         test_nb_total, test_nb_correct = compute_nb_correct(
 774             self.test_input[:1000], self.test_prior_visits[:1000]
 775         )
 776
 777         logger(
 778             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}%"
 779         )
 780
 781
 782 ######################################################################
 783
 784
 785 import stack
 786
 787
 788 class Stack(Task):
 789     def __init__(
 790         self,
 791         nb_train_samples,
 792         nb_test_samples,
 793         batch_size,
 794         logger,
 795         nb_steps,
 796         nb_stacks,
 797         nb_digits,
 798         fraction_values_for_train=None,
 799         device=torch.device("cpu"),
 800     ):
 801         super().__init__()
 802
 803         self.batch_size = batch_size
 804         self.nb_steps = nb_steps
 805         self.nb_stacks = nb_stacks
 806         self.nb_digits = nb_digits
 807         self.device = device
 808
 809         if fraction_values_for_train is None:
 810             values_for_train = None
 811             values_for_test = None
 812         else:
 813             all = torch.randperm(10**nb_digits)
 814             nb_for_train = int(all.size(0) * fraction_values_for_train)
 815             values_for_train = all[:nb_for_train]
 816             values_for_test = all[nb_for_train:]
 817
 818         self.train_input, self.train_stack_counts = stack.generate_sequences(
 819             nb_train_samples,
 820             nb_steps,
 821             nb_stacks,
 822             nb_digits,
 823             values_for_train,
 824             self.device,
 825         )
 826
 827         self.test_input, self.test_stack_counts = stack.generate_sequences(
 828             nb_test_samples,
 829             nb_steps,
 830             nb_stacks,
 831             nb_digits,
 832             values_for_test,
 833             self.device,
 834         )
 835
 836         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 837         counts = self.test_stack_counts.flatten()[i.flatten()]
 838         counts = F.one_hot(counts).sum(0)
 839         logger(f"test_pop_stack_counts {counts}")
 840
 841         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 842
 843     def batches(self, split="train", nb_to_use=-1, desc=None):
 844         assert split in {"train", "test"}
 845         input = self.train_input if split == "train" else self.test_input
 846         if nb_to_use > 0:
 847             input = input[:nb_to_use]
 848         if desc is None:
 849             desc = f"epoch-{split}"
 850         for batch in tqdm.tqdm(
 851             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 852         ):
 853             yield batch
 854
 855     def vocabulary_size(self):
 856         return self.nb_codes
 857
 858     def produce_results(
 859         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 860     ):
 861         def compute_nb_correct(input):
 862             result = input.clone()
 863             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 864             ar_mask = (result != input).long()
 865             masked_inplace_autoregression(
 866                 model,
 867                 self.batch_size,
 868                 result,
 869                 ar_mask,
 870                 deterministic_synthesis,
 871                 device=self.device,
 872             )
 873
 874             errors = ((result != input).long() * ar_mask).reshape(
 875                 -1, 1 + self.nb_digits
 876             )
 877             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 878
 879             nb_total = ar_mask.max(1).values.sum()
 880             nb_correct = nb_total - errors.max(1).values.sum()
 881
 882             return nb_total, nb_correct
 883
 884         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 885
 886         logger(
 887             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}%"
 888         )
 889
 890         ##############################################################
 891         # Log a few generated sequences
 892         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 893         result = input.clone()
 894         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 895         ar_mask = (result != input).long()
 896
 897         # for n in range(result.size(0)):
 898         # logger(
 899         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 900         # )
 901
 902         masked_inplace_autoregression(
 903             model,
 904             self.batch_size,
 905             result,
 906             ar_mask,
 907             deterministic_synthesis,
 908             device=self.device,
 909         )
 910
 911         for n in range(result.size(0)):
 912             logger(
 913                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 914             )
 915         ##############################################################
 916
 917
 918 ######################################################################
 919
 920 import rpl
 921
 922
 923 class RPL(Task):
 924     def tensorize(self, sequences):
 925         len_max = max([len(x) for x in sequences])
 926         return torch.cat(
 927             [
 928                 torch.tensor(
 929                     [
 930                         [
 931                             self.token2id[str(c)]
 932                             for c in s + ["<nul>"] * (len_max - len(s))
 933                         ]
 934                         for s in sequences
 935                     ]
 936                 )
 937             ],
 938             0,
 939         )
 940
 941     def seq2str(self, seq):
 942         return " ".join([self.id2token[i] for i in seq])
 943
 944     def __init__(
 945         self,
 946         nb_train_samples,
 947         nb_test_samples,
 948         batch_size,
 949         nb_starting_values=3,
 950         max_input=9,
 951         prog_len=6,
 952         nb_runs=5,
 953         no_prog=False,
 954         logger=None,
 955         device=torch.device("cpu"),
 956     ):
 957         super().__init__()
 958
 959         self.batch_size = batch_size
 960         self.device = device
 961         self.no_prog = no_prog
 962
 963         train_sequences = [
 964             rpl.generate(
 965                 nb_starting_values=nb_starting_values,
 966                 nb_result_values_max=4 * nb_starting_values,
 967                 max_input=max_input,
 968                 prog_len=prog_len,
 969                 nb_runs=nb_runs,
 970             )
 971             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
 972         ]
 973
 974         test_sequences = [
 975             rpl.generate(
 976                 nb_starting_values=nb_starting_values,
 977                 nb_result_values_max=4 * nb_starting_values,
 978                 max_input=max_input,
 979                 prog_len=prog_len,
 980                 nb_runs=nb_runs,
 981             )
 982             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
 983         ]
 984
 985         symbols = list(
 986             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
 987         )
 988         val_max = max([x if type(x) is int else 0 for x in symbols])
 989         symbols = list(filter(lambda x: type(x) is str, symbols))
 990         symbols.sort()
 991         symbols += [str(n) for n in range(val_max + 1)]
 992         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
 993         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
 994
 995         self.t_nul = self.token2id["<nul>"]
 996         self.t_input = self.token2id["<in>"]
 997         self.t_output = self.token2id["<out>"]
 998         self.t_prog = self.token2id["<prg>"]
 999         self.t_end = self.token2id["<end>"]
1000
1001         self.train_input = self.tensorize(train_sequences)
1002         self.test_input = self.tensorize(test_sequences)
1003
1004         if no_prog:
1005             # Excise the program from every train and test example
1006             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
1007                 None, :
1008             ]
1009             p = (
1010                 ((self.train_input == self.t_prog).long() * k)
1011                 .max(1, keepdim=True)
1012                 .values
1013             )
1014             self.train_input = (
1015                 self.train_input * (k <= p).long()
1016                 + self.t_end * (k == p + 1).long()
1017                 + self.t_nul * (k > p + 1).long()
1018             )
1019             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
1020                 None, :
1021             ]
1022             p = (
1023                 ((self.test_input == self.t_prog).long() * k)
1024                 .max(1, keepdim=True)
1025                 .values
1026             )
1027             self.test_input = (
1028                 self.test_input * (k <= p).long()
1029                 + self.t_end * (k == p + 1).long()
1030                 + self.t_nul * (k > p + 1).long()
1031             )
1032
1033         if logger is not None:
1034             logger(f"value_max {val_max}")
1035             for x in self.train_input[:25]:
1036                 end = (x != self.t_nul).nonzero().max().item() + 1
1037                 seq = [self.id2token[i.item()] for i in x[:end]]
1038                 s = " ".join(seq)
1039                 logger(f"example_seq {s}")
1040
1041         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1042
1043     def batches(self, split="train", nb_to_use=-1, desc=None):
1044         assert split in {"train", "test"}
1045         input = self.train_input if split == "train" else self.test_input
1046         if nb_to_use > 0:
1047             input = input[:nb_to_use]
1048         if desc is None:
1049             desc = f"epoch-{split}"
1050         for batch in tqdm.tqdm(
1051             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1052         ):
1053             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1054             batch = batch[:, :last].to(self.device)
1055             yield batch
1056
1057     def vocabulary_size(self):
1058         return self.nb_codes
1059
1060     def produce_results(
1061         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1062     ):
1063         # --------------------------------------------------------------------
1064         def compute_nb_errors_prog(input, nb_to_log=0):
1065             result = input.clone()
1066             s = (result == self.t_prog).long()
1067             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1068             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1069
1070             masked_inplace_autoregression(
1071                 model,
1072                 self.batch_size,
1073                 result,
1074                 ar_mask,
1075                 deterministic_synthesis,
1076                 device=self.device,
1077             )
1078
1079             sum_nb_total, sum_nb_errors = 0, 0
1080             for one_input, one_result in zip(input, result):
1081                 seq = [self.id2token[i.item()] for i in one_result]
1082                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1083                 sum_nb_total += 1
1084                 sum_nb_errors += 0 if nb_errors == 0 else 1
1085                 if nb_to_log > 0:
1086                     gt_seq = [self.id2token[i.item()] for i in one_input]
1087                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1088                     gt_prog = " ".join([str(x) for x in gt_prog])
1089                     prog = " ".join([str(x) for x in prog])
1090                     comment = "*" if nb_errors == 0 else "-"
1091                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1092                     for start_stack, target_stack, result_stack, correct in stacks:
1093                         comment = "*" if correct else "-"
1094                         start_stack = " ".join([str(x) for x in start_stack])
1095                         target_stack = " ".join([str(x) for x in target_stack])
1096                         result_stack = " ".join([str(x) for x in result_stack])
1097                         logger(
1098                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1099                         )
1100                     nb_to_log -= 1
1101
1102             return sum_nb_total, sum_nb_errors
1103
1104         # --------------------------------------------------------------------
1105         def compute_nb_errors_output(input, nb_to_log=0):
1106             result = input.clone()
1107             k = torch.arange(result.size(1), device=result.device)[None, :]
1108             last_output_idx = (
1109                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1110             )
1111             first_prog_idx = (
1112                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1113             )
1114             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1115             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1116
1117             masked_inplace_autoregression(
1118                 model,
1119                 self.batch_size,
1120                 result,
1121                 ar_mask,
1122                 deterministic_synthesis,
1123                 device=self.device,
1124             )
1125
1126             sum_nb_total, sum_nb_errors = 0, 0
1127             for one_input, one_result, i, j in zip(
1128                 input, result, last_output_idx, first_prog_idx
1129             ):
1130                 seq = [self.id2token[i.item()] for i in one_result]
1131                 sum_nb_total += 1
1132                 correct = (one_input - one_result).abs().max() == 0
1133                 sum_nb_errors += 0 if correct else 1
1134                 if nb_to_log > 0:
1135                     result_stack = [
1136                         self.id2token[i.item()] for i in one_result[i : j + 1]
1137                     ]
1138                     target_stack = [
1139                         self.id2token[i.item()] for i in one_input[i : j + 1]
1140                     ]
1141                     comment = "*" if correct else "-"
1142                     result_stack = " ".join([str(x) for x in result_stack])
1143                     target_stack = " ".join([str(x) for x in target_stack])
1144                     logger(
1145                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1146                     )
1147                     nb_to_log -= 1
1148
1149             return sum_nb_total, sum_nb_errors
1150
1151         # --------------------------------------------------------------------
1152
1153         if not self.no_prog:
1154             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1155                 self.test_input[:1000].to(self.device), nb_to_log=10
1156             )
1157
1158             logger(
1159                 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}%"
1160             )
1161
1162         test_nb_total, test_nb_errors = compute_nb_errors_output(
1163             self.test_input[:1000].to(self.device), nb_to_log=10
1164         )
1165
1166         logger(
1167             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}%"
1168         )
1169
1170         if save_attention_image is not None:
1171             ns = torch.randint(self.test_input.size(0), (1,)).item()
1172             input = self.test_input[ns : ns + 1].clone()
1173             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1174             input = input[:, :last].to(self.device)
1175
1176             with torch.autograd.no_grad():
1177                 t = model.training
1178                 model.eval()
1179                 model.record_attention(True)
1180                 model(BracketedSequence(input))
1181                 model.train(t)
1182                 ram = model.retrieve_attention()
1183                 model.record_attention(False)
1184
1185             tokens_output = [self.id2token[i.item()] for i in input[0]]
1186             tokens_input = ["n/a"] + tokens_output[:-1]
1187             for n_head in range(ram[0].size(1)):
1188                 filename = os.path.join(
1189                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1190                 )
1191                 attention_matrices = [m[0, n_head] for m in ram]
1192                 save_attention_image(
1193                     filename,
1194                     tokens_input,
1195                     tokens_output,
1196                     attention_matrices,
1197                     k_top=10,
1198                     # min_total_attention=0.9,
1199                     token_gap=12,
1200                     layer_gap=50,
1201                 )
1202                 logger(f"wrote {filename}")
1203
1204
1205 ######################################################################
1206
1207
1208 import expr
1209
1210
1211 class Expr(Task):
1212     def tensorize(self, sequences):
1213         len_max = max([len(x) for x in sequences])
1214         return torch.cat(
1215             [
1216                 torch.tensor(
1217                     [
1218                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1219                         for s in sequences
1220                     ]
1221                 )
1222             ],
1223             0,
1224         ).to(self.device)
1225
1226     def __init__(
1227         self,
1228         nb_train_samples,
1229         nb_test_samples,
1230         nb_variables,
1231         sequence_length,
1232         operand_max,
1233         result_max,
1234         batch_size,
1235         device=torch.device("cpu"),
1236     ):
1237         super().__init__()
1238
1239         self.batch_size = batch_size
1240         self.device = device
1241
1242         train_sequences = expr.generate_sequences(
1243             nb_train_samples,
1244             nb_variables=nb_variables,
1245             length=sequence_length,
1246             operand_max=operand_max,
1247             result_max=result_max,
1248         )
1249
1250         test_sequences = expr.generate_sequences(
1251             nb_test_samples,
1252             nb_variables=nb_variables,
1253             length=sequence_length,
1254             operand_max=operand_max,
1255             result_max=result_max,
1256         )
1257
1258         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1259         symbols.sort()
1260
1261         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1262         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1263
1264         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1265
1266         self.train_input = self.tensorize(train_sequences)
1267         self.test_input = self.tensorize(test_sequences)
1268
1269         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1270
1271     def batches(self, split="train", nb_to_use=-1, desc=None):
1272         assert split in {"train", "test"}
1273         input = self.train_input if split == "train" else self.test_input
1274         if nb_to_use > 0:
1275             input = input[:nb_to_use]
1276         if desc is None:
1277             desc = f"epoch-{split}"
1278         for batch in tqdm.tqdm(
1279             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1280         ):
1281             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1282             batch = batch[:, :last]
1283             yield batch
1284
1285     def vocabulary_size(self):
1286         return self.nb_codes
1287
1288     def seq2str(self, s):
1289         return "".join([self.id2char[k.item()] for k in s])
1290
1291     def produce_results(
1292         self,
1293         n_epoch,
1294         model,
1295         result_dir,
1296         logger,
1297         deterministic_synthesis,
1298         input_file=None,
1299     ):
1300         def compute_nb_correct(input):
1301             result = input.clone()
1302             s = (result == self.space).long()
1303             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1304             result = (1 - ar_mask) * result + ar_mask * self.filler
1305             masked_inplace_autoregression(
1306                 model,
1307                 self.batch_size,
1308                 result,
1309                 ar_mask,
1310                 deterministic_synthesis,
1311                 device=self.device,
1312             )
1313
1314             nb_total = input.size(0)
1315             nb_correct = (input == result).long().min(1).values.sum()
1316
1317             #######################################################################
1318             # Comput predicted vs. true variable values
1319
1320             nb_delta = torch.zeros(5, dtype=torch.int64)
1321             nb_missed = 0
1322
1323             values_input = expr.extract_results([self.seq2str(s) for s in input])
1324             values_result = expr.extract_results([self.seq2str(s) for s in result])
1325
1326             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1327
1328             with open(filename, "w") as f:
1329                 for i, r in zip(values_input, values_result):
1330                     for n, vi in i.items():
1331                         vr = r.get(n)
1332                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1333
1334                         if vr is None or vr < 0:
1335                             nb_missed += 1
1336                         else:
1337                             d = abs(vr - vi)
1338                             if d >= nb_delta.size(0):
1339                                 nb_missed += 1
1340                             else:
1341                                 nb_delta[d] += 1
1342
1343             ######################################################################
1344
1345             return nb_total, nb_correct, nb_delta, nb_missed
1346
1347         (
1348             test_nb_total,
1349             test_nb_correct,
1350             test_nb_delta,
1351             test_nb_missed,
1352         ) = compute_nb_correct(self.test_input[:10000])
1353
1354         logger(
1355             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}%"
1356         )
1357
1358         nb_total = test_nb_delta.sum() + test_nb_missed
1359         for d in range(test_nb_delta.size(0)):
1360             logger(
1361                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1362             )
1363         logger(
1364             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1365         )
1366
1367         ##############################################################
1368         # Log a few generated sequences
1369         if input_file is None:
1370             input = self.test_input[:10]
1371         else:
1372             with open(input_file, "r") as f:
1373                 sequences = [e.strip() for e in f.readlines()]
1374                 sequences = [s + " " + "#" * 50 for s in sequences]
1375                 input = self.tensorize(sequences)
1376
1377         result = input.clone()
1378         s = (result == self.space).long()
1379         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1380         result = (1 - ar_mask) * result + ar_mask * self.filler
1381
1382         for n in range(result.size(0)):
1383             logger(f"test_before {self.seq2str(result[n])}")
1384
1385         masked_inplace_autoregression(
1386             model,
1387             self.batch_size,
1388             result,
1389             ar_mask,
1390             deterministic_synthesis,
1391             device=self.device,
1392         )
1393
1394         correct = (1 - ar_mask) * self.space + ar_mask * input
1395         for n in range(result.size(0)):
1396             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1397             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1398             logger(f"truth       {self.seq2str(correct[n])}")
1399         ##############################################################
1400
1401
1402 ######################################################################
1403
1404 import world
1405
1406
1407 class World(Task):
1408     def __init__(
1409         self,
1410         nb_train_samples,
1411         nb_test_samples,
1412         batch_size,
1413         vqae_nb_epochs,
1414         logger=None,
1415         device=torch.device("cpu"),
1416         device_storage=torch.device("cpu"),
1417     ):
1418         super().__init__()
1419
1420         self.batch_size = batch_size
1421         self.device = device
1422
1423         (
1424             train_frames,
1425             train_action_seq,
1426             test_frames,
1427             test_action_seq,
1428             self.frame2seq,
1429             self.seq2frame,
1430         ) = world.create_data_and_processors(
1431             nb_train_samples,
1432             nb_test_samples,
1433             mode="first_last",
1434             nb_steps=30,
1435             nb_epochs=vqae_nb_epochs,
1436             logger=logger,
1437             device=device,
1438             device_storage=device_storage,
1439         )
1440
1441         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
1442         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
1443
1444         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
1445         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
1446
1447         self.len_frame_seq = train_frame_seq.size(1)
1448         self.len_action_seq = train_action_seq.size(1)
1449         self.nb_codes = nb_frame_codes + nb_action_codes
1450
1451         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
1452
1453         train_action_seq += nb_frame_codes
1454         self.train_input = torch.cat(
1455             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
1456         )
1457
1458         test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
1459         test_action_seq += nb_frame_codes
1460         self.test_input = torch.cat(
1461             (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
1462         )
1463
1464     def batches(self, split="train", nb_to_use=-1, desc=None):
1465         assert split in {"train", "test"}
1466         input = self.train_input if split == "train" else self.test_input
1467         if nb_to_use > 0:
1468             input = input[:nb_to_use]
1469         if desc is None:
1470             desc = f"epoch-{split}"
1471         for batch in tqdm.tqdm(
1472             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1473         ):
1474             yield batch.to(self.device)
1475
1476     def vocabulary_size(self):
1477         return self.nb_codes
1478
1479     def produce_results(
1480         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1481     ):
1482         k = torch.arange(
1483             2 * self.len_frame_seq + self.len_action_seq, device=self.device
1484         )[None, :]
1485
1486         input = self.test_input[:64].to(self.device)
1487         result = input.clone()
1488
1489         ar_mask = (
1490             (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
1491         )
1492         result *= 1 - ar_mask
1493
1494         masked_inplace_autoregression(
1495             model,
1496             self.batch_size,
1497             result,
1498             ar_mask,
1499             deterministic_synthesis,
1500             device=self.device,
1501         )
1502
1503         seq_start = input[:, : self.len_frame_seq]
1504         seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
1505         seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
1506
1507         result = torch.cat(
1508             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
1509         )
1510         result = result.reshape(-1, result.size(-1))
1511
1512         frames = self.seq2frame(result)
1513         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1514         torchvision.utils.save_image(
1515             frames.float() / (world.Box.nb_rgb_levels - 1),
1516             image_name,
1517             nrow=12,
1518             padding=1,
1519             pad_value=0.0,
1520         )
1521         logger(f"wrote {image_name}")
1522
1523
1524 ######################################################################