tasks.py

   1 #!/usr/bin/env python
   2
   3 import math, os, tqdm
   4
   5 import torch, torchvision
   6
   7 from torch import nn
   8 from torch.nn import functional as F
   9
  10 ######################################################################
  11
  12
  13 def masked_inplace_autoregression(
  14     model,
  15     batch_size,
  16     input,
  17     ar_mask,
  18     deterministic_synthesis,
  19     forbidden_tokens=None,
  20     progress_bar_desc="autoregression",
  21     device=torch.device("cpu"),
  22 ):
  23     assert input.size() == ar_mask.size()
  24
  25     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
  26
  27     if progress_bar_desc is not None:
  28         batches = tqdm.tqdm(
  29             batches,
  30             dynamic_ncols=True,
  31             desc=progress_bar_desc,
  32             #total=input.size(0) // batch_size,
  33         )
  34
  35     with torch.autograd.no_grad():
  36         t = model.training
  37         model.eval()
  38
  39         for input, ar_mask in batches:
  40             model.masked_inplace_autoregression(
  41                 input, ar_mask, forbidden_tokens, deterministic_synthesis
  42             )
  43
  44         model.train(t)
  45
  46
  47 ######################################################################
  48
  49
  50 class Task:
  51     def batches(self, split="train"):
  52         pass
  53
  54     def vocabulary_size(self):
  55         pass
  56
  57     def produce_results(
  58         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  59     ):
  60         pass
  61
  62
  63 ######################################################################
  64
  65 import picoclvr
  66
  67
  68 class PicoCLVR(Task):
  69     # Make a tensor from a list of strings
  70     def tensorize(self, descr):
  71         token_descr = [s.strip().split(" ") for s in descr]
  72         l = max([len(s) for s in token_descr])
  73         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
  74         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
  75         return torch.tensor(id_descr, device=self.device)
  76
  77     # Make a list of strings from a tensor
  78     def detensorize(self, x):
  79         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
  80
  81     # trim all the tensors in the tuple z to remove as much token from
  82     # left and right in the first tensor. If z is a tuple, all its
  83     # elements are trimed according to the triming for the first
  84     def trim(self, z, token="<nul>"):
  85         n = self.token2id[token]
  86         if type(z) == tuple:
  87             x = z[0]
  88             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  89             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  90             return tuple([t[:, a:b] for t in z])
  91         else:
  92             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  93             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  94             return z[:, a:b]
  95
  96     ######################
  97
  98     def __init__(
  99         self,
 100         nb_train_samples,
 101         nb_test_samples,
 102         batch_size,
 103         height,
 104         width,
 105         nb_colors=5,
 106         logger=None,
 107         device=torch.device("cpu"),
 108         pruner_train=None,
 109         pruner_eval=None,
 110     ):
 111         def generate_descr(nb, cache_suffix, pruner):
 112             return picoclvr.generate(
 113                 nb,
 114                 height=self.height,
 115                 width=self.width,
 116                 nb_colors=nb_colors,
 117                 pruner=pruner,
 118             )
 119
 120         self.height = height
 121         self.width = width
 122         self.batch_size = batch_size
 123         self.device = device
 124         self.pruner_train = pruner_train
 125         self.pruner_eval = pruner_eval
 126
 127         if logger is not None:
 128             logger(
 129                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 130             )
 131
 132         self.train_descr = generate_descr(
 133             nb_train_samples, "train", pruner=self.pruner_train
 134         )
 135         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 136
 137         # Build the tokenizer
 138         tokens = {"<nul>", "<img>"}
 139         for d in [self.train_descr, self.test_descr]:
 140             for s in d:
 141                 for t in s.strip().split(" "):
 142                     tokens.add(t)
 143         # make this set a sorted list to get the same tensors given
 144         # the same descr
 145         tokens = list(tokens)
 146         tokens.sort()
 147         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 148         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 149         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 150
 151         # Tokenize the train and test sets
 152         self.train_input = self.tensorize(self.train_descr)
 153         self.test_input = self.tensorize(self.test_descr)
 154
 155     def batches(self, split="train"):
 156         assert split in {"train", "test"}
 157         input = self.train_input if split == "train" else self.test_input
 158         for batch in tqdm.tqdm(
 159             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 160         ):
 161             yield self.trim(batch)
 162
 163     def vocabulary_size(self):
 164         return len(self.token2id)
 165
 166     def compute_missing_properties(
 167         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 168     ):
 169         acc_nb_requested_properties = []
 170         acc_nb_missing_properties = []
 171         acc_nb_results = 0
 172
 173         for input in tqdm.tqdm(
 174             self.test_input.split(self.batch_size),
 175             dynamic_ncols=True,
 176             desc=f"test-properties",
 177         ):
 178             result = input.clone()
 179             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 180             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 181             masked_inplace_autoregression(
 182                 model,
 183                 self.batch_size,
 184                 result,
 185                 ar_mask,
 186                 deterministic_synthesis,
 187                 progress_bar_desc=None,
 188                 device=self.device,
 189             )
 190
 191             result_descr = self.detensorize(result)
 192             np = picoclvr.nb_properties(
 193                 result_descr,
 194                 height=self.height,
 195                 width=self.width,
 196                 pruner=pruner,
 197             )
 198             nb_requested_properties, _, nb_missing_properties = zip(*np)
 199             acc_nb_requested_properties += nb_requested_properties
 200             acc_nb_missing_properties += nb_missing_properties
 201             acc_nb_results += len(result_descr)
 202
 203         nb_requested_properties = sum(acc_nb_requested_properties)
 204         nb_missing_properties = sum(acc_nb_missing_properties)
 205
 206         prefix = "" if pruner is None else "pruned_"
 207         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 208         logger(
 209             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 210         )
 211         logger(
 212             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 213         )
 214
 215     ######################################################################
 216
 217     def produce_results(
 218         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 219     ):
 220         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 221
 222         if self.pruner_eval is not None:
 223             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 224
 225         nb_tokens_to_generate = self.height * self.width + 3
 226         result_descr = []
 227         nb_per_primer = 8
 228         primer = []
 229
 230         for primer_descr in [
 231             "red above green <sep> green top <sep> blue right of red",
 232             "there is red <sep> there is yellow <sep> there is blue",
 233             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 234             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 235         ]:
 236             primer += [primer_descr + " <img>"] * nb_per_primer
 237
 238         result = self.tensorize(primer)
 239         fill = result.new_full(
 240             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 241         )
 242         result = torch.cat((result, fill), 1)
 243         ar_mask = (result == self.t_nul).long()
 244         masked_inplace_autoregression(
 245             model,
 246             self.batch_size,
 247             result,
 248             ar_mask,
 249             deterministic_synthesis,
 250             device=self.device,
 251         )
 252         result_descr = self.detensorize(result)
 253
 254         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 255
 256         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 257         acc_nb_results = len(result_descr)
 258
 259         nb_requested_properties = sum(acc_nb_requested_properties)
 260         nb_missing_properties = sum(acc_nb_missing_properties)
 261
 262         prefix = "demo_"
 263         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 264         logger(
 265             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 266         )
 267         logger(
 268             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 269         )
 270
 271         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 272
 273         if img.dim() == 5:
 274             if img.size(1) == 1:
 275                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 276             else:
 277                 img = torch.cat(
 278                     [
 279                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 280                         for x in img
 281                     ],
 282                     0,
 283                 )
 284
 285         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 286         torchvision.utils.save_image(
 287             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 288         )
 289         logger(f"wrote {image_name}")
 290
 291
 292 ######################################################################
 293
 294
 295 class MNIST(Task):
 296     def __init__(
 297         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 298     ):
 299         self.nb_train_samples = (nb_train_samples,)
 300         self.nb_test_samples = (nb_test_samples,)
 301         self.batch_size = batch_size
 302         self.device = device
 303         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 304         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 305         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 306         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 307
 308     def batches(self, split="train", nb_to_use=-1, desc=None):
 309         assert split in {"train", "test"}
 310         input = self.train_input if split == "train" else self.test_input
 311         if nb_to_use > 0:
 312             input = input[:nb_to_use]
 313         if desc is None:
 314             desc = f"epoch-{split}"
 315         for batch in tqdm.tqdm(
 316             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 317         ):
 318             yield batch
 319
 320     def vocabulary_size(self):
 321         return 256
 322
 323     def produce_results(
 324         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 325     ):
 326         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 327         ar_mask = torch.full_like(results, 1)
 328         masked_inplace_autoregression(
 329             model,
 330             self.batch_size,
 331             results,
 332             ar_mask,
 333             deterministic_synthesis,
 334             device=self.device,
 335         )
 336         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 337         torchvision.utils.save_image(
 338             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 339             image_name,
 340             nrow=16,
 341             pad_value=0.8,
 342         )
 343         logger(f"wrote {image_name}")
 344
 345
 346 ######################################################################
 347
 348 import maze
 349
 350
 351 class Maze(Task):
 352     def map2seq(self, *m):
 353         return torch.cat([x.flatten(1) for x in m], 1)
 354
 355     def seq2map(self, s):
 356         s = s.reshape(s.size(0), -1, self.height, self.width)
 357         return (s[:, k] for k in range(s.size(1)))
 358
 359     def __init__(
 360         self,
 361         nb_train_samples,
 362         nb_test_samples,
 363         batch_size,
 364         height,
 365         width,
 366         nb_walls,
 367         device=torch.device("cpu"),
 368     ):
 369         self.batch_size = batch_size
 370         self.height = height
 371         self.width = width
 372         self.device = device
 373
 374         train_mazes, train_paths, _ = maze.create_maze_data(
 375             nb_train_samples,
 376             height=height,
 377             width=width,
 378             nb_walls=nb_walls,
 379             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 380         )
 381         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 382
 383         test_mazes, test_paths, _ = maze.create_maze_data(
 384             nb_test_samples,
 385             height=height,
 386             width=width,
 387             nb_walls=nb_walls,
 388             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 389         )
 390         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 391
 392         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 393
 394     def batches(self, split="train", nb_to_use=-1, desc=None):
 395         assert split in {"train", "test"}
 396         input = self.train_input if split == "train" else self.test_input
 397         if nb_to_use > 0:
 398             input = input[:nb_to_use]
 399         if desc is None:
 400             desc = f"epoch-{split}"
 401         for batch in tqdm.tqdm(
 402             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 403         ):
 404             yield batch
 405
 406     def vocabulary_size(self):
 407         return self.nb_codes
 408
 409     def compute_error(
 410         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 411     ):
 412         nb_total, nb_correct = 0, 0
 413         count = torch.zeros(
 414             self.width * self.height,
 415             self.width * self.height,
 416             device=self.device,
 417             dtype=torch.int64,
 418         )
 419
 420         for input in self.batches(split, nb_to_use):
 421             result = input.clone()
 422             ar_mask = result.new_zeros(result.size())
 423             ar_mask[:, self.height * self.width :] = 1
 424             result *= 1 - ar_mask
 425             masked_inplace_autoregression(
 426                 model,
 427                 self.batch_size,
 428                 result,
 429                 ar_mask,
 430                 deterministic_synthesis,
 431                 progress_bar_desc=None,
 432                 device=self.device,
 433             )
 434             mazes, paths = self.seq2map(result)
 435             path_correctness = maze.path_correctness(mazes, paths)
 436             nb_correct += path_correctness.long().sum()
 437             nb_total += mazes.size(0)
 438
 439             optimal_path_lengths = (
 440                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 441             )
 442             predicted_path_lengths = (
 443                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 444             )
 445             optimal_path_lengths = optimal_path_lengths[path_correctness]
 446             predicted_path_lengths = predicted_path_lengths[path_correctness]
 447             count[optimal_path_lengths, predicted_path_lengths] += 1
 448
 449         if count.max() == 0:
 450             count = None
 451         else:
 452             count = count[
 453                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 454             ]
 455
 456         return nb_total, nb_correct, count
 457
 458     def produce_results(
 459         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 460     ):
 461         train_nb_total, train_nb_correct, count = self.compute_error(
 462             model,
 463             "train",
 464             nb_to_use=1000,
 465             deterministic_synthesis=deterministic_synthesis,
 466         )
 467         logger(
 468             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}%"
 469         )
 470
 471         test_nb_total, test_nb_correct, count = self.compute_error(
 472             model,
 473             "test",
 474             nb_to_use=1000,
 475             deterministic_synthesis=deterministic_synthesis,
 476         )
 477         logger(
 478             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}%"
 479         )
 480
 481         if count is not None:
 482             proportion_optimal = count.diagonal().sum().float() / count.sum()
 483             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 484             with open(
 485                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 486             ) as f:
 487                 for i in range(count.size(0)):
 488                     for j in range(count.size(1)):
 489                         eol = " " if j < count.size(1) - 1 else "\n"
 490                         f.write(f"{count[i,j]}{eol}")
 491
 492         input = self.test_input[:48]
 493         result = input.clone()
 494         ar_mask = result.new_zeros(result.size())
 495         ar_mask[:, self.height * self.width :] = 1
 496         result *= 1 - ar_mask
 497         masked_inplace_autoregression(
 498             model,
 499             self.batch_size,
 500             result,
 501             ar_mask,
 502             deterministic_synthesis,
 503             device=self.device,
 504         )
 505
 506         mazes, paths = self.seq2map(input)
 507         _, predicted_paths = self.seq2map(result)
 508
 509         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 510         maze.save_image(
 511             filename,
 512             mazes=mazes,
 513             target_paths=paths,
 514             predicted_paths=predicted_paths,
 515             path_correct=maze.path_correctness(mazes, predicted_paths),
 516             path_optimal=maze.path_optimality(paths, predicted_paths),
 517         )
 518         logger(f"wrote {filename}")
 519
 520
 521 ######################################################################
 522
 523
 524 import snake
 525
 526
 527 class Snake(Task):
 528     def __init__(
 529         self,
 530         nb_train_samples,
 531         nb_test_samples,
 532         batch_size,
 533         height,
 534         width,
 535         nb_colors,
 536         length,
 537         prompt_length,
 538         device=torch.device("cpu"),
 539     ):
 540         self.batch_size = batch_size
 541         self.height = height
 542         self.width = width
 543         self.device = device
 544         self.prompt_length = prompt_length
 545
 546         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 547             nb_train_samples,
 548             height,
 549             width,
 550             nb_colors,
 551             length,
 552             prompt_length,
 553             self.device,
 554         )
 555         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 556             nb_test_samples,
 557             height,
 558             width,
 559             nb_colors,
 560             length,
 561             prompt_length,
 562             self.device,
 563         )
 564
 565         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 566
 567     def batches(self, split="train", nb_to_use=-1, desc=None):
 568         assert split in {"train", "test"}
 569         input = self.train_input if split == "train" else self.test_input
 570         if nb_to_use > 0:
 571             input = input[:nb_to_use]
 572         if desc is None:
 573             desc = f"epoch-{split}"
 574         for batch in tqdm.tqdm(
 575             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 576         ):
 577             yield batch
 578
 579     def vocabulary_size(self):
 580         return self.nb_codes
 581
 582     def produce_results(
 583         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 584     ):
 585         def compute_nb_correct(input, prior_visits):
 586             result = input.clone()
 587             i = torch.arange(result.size(1), device=result.device)[None, :]
 588             ar_mask = (
 589                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 590                 .long()
 591                 .expand_as(result)
 592             )
 593             result *= 1 - ar_mask
 594
 595             masked_inplace_autoregression(
 596                 model,
 597                 self.batch_size,
 598                 result,
 599                 ar_mask,
 600                 deterministic_synthesis,
 601                 device=self.device,
 602             )
 603
 604             nb_total = ((prior_visits > 0) * ar_mask).sum()
 605
 606             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 607
 608             return nb_total, nb_correct
 609
 610         test_nb_total, test_nb_correct = compute_nb_correct(
 611             self.test_input[:1000], self.test_prior_visits[:1000]
 612         )
 613
 614         logger(
 615             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}%"
 616         )
 617
 618
 619 ######################################################################
 620
 621
 622 import stack
 623
 624
 625 class Stack(Task):
 626     def __init__(
 627         self,
 628         nb_train_samples,
 629         nb_test_samples,
 630         batch_size,
 631         logger,
 632         nb_steps,
 633         nb_stacks,
 634         nb_digits,
 635         fraction_values_for_train=None,
 636         device=torch.device("cpu"),
 637     ):
 638         self.batch_size = batch_size
 639         self.nb_steps = nb_steps
 640         self.nb_stacks = nb_stacks
 641         self.nb_digits = nb_digits
 642         self.device = device
 643
 644         if fraction_values_for_train is None:
 645             values_for_train = None
 646             values_for_test = None
 647         else:
 648             all = torch.randperm(10**nb_digits)
 649             nb_for_train = int(all.size(0) * fraction_values_for_train)
 650             values_for_train = all[:nb_for_train]
 651             values_for_test = all[nb_for_train:]
 652
 653         self.train_input, self.train_stack_counts = stack.generate_sequences(
 654             nb_train_samples,
 655             nb_steps,
 656             nb_stacks,
 657             nb_digits,
 658             values_for_train,
 659             self.device,
 660         )
 661
 662         self.test_input, self.test_stack_counts = stack.generate_sequences(
 663             nb_test_samples,
 664             nb_steps,
 665             nb_stacks,
 666             nb_digits,
 667             values_for_test,
 668             self.device,
 669         )
 670
 671         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 672         counts = self.test_stack_counts.flatten()[i.flatten()]
 673         counts = F.one_hot(counts).sum(0)
 674         logger(f"test_pop_stack_counts {counts}")
 675
 676         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 677
 678     def batches(self, split="train", nb_to_use=-1, desc=None):
 679         assert split in {"train", "test"}
 680         input = self.train_input if split == "train" else self.test_input
 681         if nb_to_use > 0:
 682             input = input[:nb_to_use]
 683         if desc is None:
 684             desc = f"epoch-{split}"
 685         for batch in tqdm.tqdm(
 686             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 687         ):
 688             yield batch
 689
 690     def vocabulary_size(self):
 691         return self.nb_codes
 692
 693     def produce_results(
 694         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 695     ):
 696         def compute_nb_correct(input):
 697             result = input.clone()
 698             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 699             ar_mask = (result != input).long()
 700             masked_inplace_autoregression(
 701                 model,
 702                 self.batch_size,
 703                 result,
 704                 ar_mask,
 705                 deterministic_synthesis,
 706                 device=self.device,
 707             )
 708
 709             errors = ((result != input).long() * ar_mask).reshape(
 710                 -1, 1 + self.nb_digits
 711             )
 712             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 713
 714             nb_total = ar_mask.max(1).values.sum()
 715             nb_correct = nb_total - errors.max(1).values.sum()
 716
 717             return nb_total, nb_correct
 718
 719         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 720
 721         logger(
 722             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}%"
 723         )
 724
 725         ##############################################################
 726         # Log a few generated sequences
 727         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 728         result = input.clone()
 729         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 730         ar_mask = (result != input).long()
 731
 732         # for n in range(result.size(0)):
 733         # logger(
 734         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 735         # )
 736
 737         masked_inplace_autoregression(
 738             model,
 739             self.batch_size,
 740             result,
 741             ar_mask,
 742             deterministic_synthesis,
 743             device=self.device,
 744         )
 745
 746         for n in range(result.size(0)):
 747             logger(
 748                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 749             )
 750         ##############################################################
 751
 752
 753 ######################################################################
 754
 755
 756 import expr
 757
 758
 759 class Expr(Task):
 760     def tensorize(self, sequences):
 761         len_max = max([len(x) for x in sequences])
 762         return torch.cat(
 763             [
 764                 torch.tensor(
 765                     [
 766                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
 767                         for s in sequences
 768                     ]
 769                 )
 770             ],
 771             0,
 772         ).to(self.device)
 773
 774     def __init__(
 775         self,
 776         nb_train_samples,
 777         nb_test_samples,
 778         nb_variables,
 779         sequence_length,
 780         operand_max,
 781         result_max,
 782         batch_size,
 783         device=torch.device("cpu"),
 784     ):
 785         self.batch_size = batch_size
 786         self.device = device
 787
 788         train_sequences = expr.generate_sequences(
 789             nb_train_samples,
 790             nb_variables=nb_variables,
 791             length=sequence_length,
 792             operand_max=operand_max,
 793             result_max=result_max,
 794         )
 795
 796         test_sequences = expr.generate_sequences(
 797             nb_test_samples,
 798             nb_variables=nb_variables,
 799             length=sequence_length,
 800             operand_max=operand_max,
 801             result_max=result_max,
 802         )
 803
 804         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
 805         symbols.sort()
 806
 807         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
 808         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 809
 810         self.filler, self.space = self.char2id["#"], self.char2id[" "]
 811
 812         self.train_input = self.tensorize(train_sequences)
 813         self.test_input = self.tensorize(test_sequences)
 814
 815         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 816
 817     def batches(self, split="train", nb_to_use=-1, desc=None):
 818         assert split in {"train", "test"}
 819         input = self.train_input if split == "train" else self.test_input
 820         if nb_to_use > 0:
 821             input = input[:nb_to_use]
 822         if desc is None:
 823             desc = f"epoch-{split}"
 824         for batch in tqdm.tqdm(
 825             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 826         ):
 827             last = (batch != self.filler).max(0).values.nonzero().max() + 3
 828             batch = batch[:, :last]
 829             yield batch
 830
 831     def vocabulary_size(self):
 832         return self.nb_codes
 833
 834     def seq2str(self, s):
 835         return "".join([self.id2char[k.item()] for k in s])
 836
 837     def produce_results(
 838         self,
 839         n_epoch,
 840         model,
 841         result_dir,
 842         logger,
 843         deterministic_synthesis,
 844         input_file=None,
 845     ):
 846         def compute_nb_correct(input):
 847             result = input.clone()
 848             s = (result == self.space).long()
 849             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
 850             result = (1 - ar_mask) * result + ar_mask * self.filler
 851             masked_inplace_autoregression(
 852                 model,
 853                 self.batch_size,
 854                 result,
 855                 ar_mask,
 856                 deterministic_synthesis,
 857                 device=self.device,
 858             )
 859
 860             nb_total = input.size(0)
 861             nb_correct = (input == result).long().min(1).values.sum()
 862
 863             #######################################################################
 864             # Comput predicted vs. true variable values
 865
 866             nb_delta = torch.zeros(5, dtype=torch.int64)
 867             nb_missed = 0
 868
 869             values_input = expr.extract_results([self.seq2str(s) for s in input])
 870             values_result = expr.extract_results([self.seq2str(s) for s in result])
 871
 872             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
 873
 874             with open(filename, "w") as f:
 875                 for i, r in zip(values_input, values_result):
 876                     for n, vi in i.items():
 877                         vr = r.get(n)
 878                         f.write(f"{vi} {-1 if vr is None else vr}\n")
 879
 880                         if vr is None or vr < 0:
 881                             nb_missed += 1
 882                         else:
 883                             d = abs(vr - vi)
 884                             if d >= nb_delta.size(0):
 885                                 nb_missed += 1
 886                             else:
 887                                 nb_delta[d] += 1
 888
 889             ######################################################################
 890
 891             return nb_total, nb_correct, nb_delta, nb_missed
 892
 893         (
 894             test_nb_total,
 895             test_nb_correct,
 896             test_nb_delta,
 897             test_nb_missed,
 898         ) = compute_nb_correct(self.test_input[:10000])
 899
 900         logger(
 901             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}%"
 902         )
 903
 904         nb_total = test_nb_delta.sum() + test_nb_missed
 905         for d in range(test_nb_delta.size(0)):
 906             logger(
 907                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
 908             )
 909         logger(
 910             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
 911         )
 912
 913         ##############################################################
 914         # Log a few generated sequences
 915         if input_file is None:
 916             input = self.test_input[:10]
 917         else:
 918             with open(input_file, "r") as f:
 919                 sequences = [e.strip() for e in f.readlines()]
 920                 sequences = [s + " " + "#" * 50 for s in sequences]
 921                 input = self.tensorize(sequences)
 922
 923         result = input.clone()
 924         s = (result == self.space).long()
 925         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
 926         result = (1 - ar_mask) * result + ar_mask * self.filler
 927
 928         for n in range(result.size(0)):
 929             logger(f"test_before {self.seq2str(result[n])}")
 930
 931         masked_inplace_autoregression(
 932             model,
 933             self.batch_size,
 934             result,
 935             ar_mask,
 936             deterministic_synthesis,
 937             device=self.device,
 938         )
 939
 940         correct = (1 - ar_mask) * self.space + ar_mask * input
 941         for n in range(result.size(0)):
 942             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
 943             logger(f"test_after  {self.seq2str(result[n])} {comment}")
 944             logger(f"truth       {self.seq2str(correct[n])}")
 945         ##############################################################
 946
 947
 948 ######################################################################
 949
 950 import world
 951
 952
 953 class World(Task):
 954     def __init__(
 955         self,
 956         nb_train_samples,
 957         nb_test_samples,
 958         batch_size,
 959         vqae_nb_epochs,
 960         device=torch.device("cpu"),
 961     ):
 962         self.batch_size = batch_size
 963         self.device = device
 964
 965         (
 966             train_frames,
 967             self.train_actions,
 968             test_frames,
 969             self.test_actions,
 970             self.frame2seq,
 971             self.seq2frame,
 972         ) = world.create_data_and_processors(
 973             nb_train_samples,
 974             nb_test_samples,
 975             mode="first_last",
 976             nb_steps=30,
 977             nb_epochs=vqae_nb_epochs,
 978             device=device,
 979         )
 980
 981         self.train_input = self.frame2seq(train_frames)
 982         self.train_input = self.train_input.reshape(self.train_input.size(0) // 2, -1)
 983         self.test_input = self.frame2seq(test_frames)
 984         self.test_input = self.test_input.reshape(self.test_input.size(0) // 2, -1)
 985
 986         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 987
 988     def batches(self, split="train", nb_to_use=-1, desc=None):
 989         assert split in {"train", "test"}
 990         input = self.train_input if split == "train" else self.test_input
 991         if nb_to_use > 0:
 992             input = input[:nb_to_use]
 993         if desc is None:
 994             desc = f"epoch-{split}"
 995         for batch in tqdm.tqdm(
 996             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 997         ):
 998             yield batch
 999
1000     def vocabulary_size(self):
1001         return self.nb_codes
1002
1003     def produce_results(
1004         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1005     ):
1006         l = self.train_input.size(1)
1007         k = torch.arange(l, device=self.device)[None, :]
1008         result = self.test_input[:64].clone()
1009
1010         ar_mask = (k >= l // 2).long().expand_as(result)
1011         result *= 1 - ar_mask
1012
1013         masked_inplace_autoregression(
1014             model,
1015             self.batch_size,
1016             result,
1017             ar_mask,
1018             deterministic_synthesis,
1019             device=self.device,
1020         )
1021
1022         result = result.reshape(result.size(0) * 2, -1)
1023
1024         frames = self.seq2frame(result)
1025         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1026         torchvision.utils.save_image(
1027             frames.float() / (world.Box.nb_rgb_levels - 1),
1028             image_name,
1029             nrow=8,
1030             padding=1,
1031             pad_value=0.0,
1032         )
1033         logger(f"wrote {image_name}")
1034
1035
1036 ######################################################################