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     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
  24
  25     if progress_bar_desc is not None:
  26         batches = tqdm.tqdm(
  27             batches,
  28             dynamic_ncols=True,
  29             desc=progress_bar_desc,
  30             total=input.size(0) // batch_size,
  31         )
  32
  33     with torch.autograd.no_grad():
  34         t = model.training
  35         model.eval()
  36
  37         for input, ar_mask in batches:
  38             model.masked_inplace_autoregression(
  39                 input, ar_mask, forbidden_tokens, deterministic_synthesis
  40             )
  41
  42         model.train(t)
  43
  44
  45 ######################################################################
  46
  47
  48 class Task:
  49     def batches(self, split="train"):
  50         pass
  51
  52     def vocabulary_size(self):
  53         pass
  54
  55     def produce_results(
  56         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  57     ):
  58         pass
  59
  60
  61 ######################################################################
  62
  63 import picoclvr
  64
  65
  66 class PicoCLVR(Task):
  67     # Make a tensor from a list of strings
  68     def tensorize(self, descr):
  69         token_descr = [s.strip().split(" ") for s in descr]
  70         l = max([len(s) for s in token_descr])
  71         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
  72         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
  73         return torch.tensor(id_descr, device=self.device)
  74
  75     # Make a list of strings from a tensor
  76     def detensorize(self, x):
  77         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
  78
  79     # trim all the tensors in the tuple z to remove as much token from
  80     # left and right in the first tensor. If z is a tuple, all its
  81     # elements are trimed according to the triming for the first
  82     def trim(self, z, token="<nul>"):
  83         n = self.token2id[token]
  84         if type(z) == tuple:
  85             x = z[0]
  86             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  87             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  88             return tuple([t[:, a:b] for t in z])
  89         else:
  90             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  91             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  92             return z[:, a:b]
  93
  94     ######################
  95
  96     def __init__(
  97         self,
  98         nb_train_samples,
  99         nb_test_samples,
 100         batch_size,
 101         height,
 102         width,
 103         nb_colors=5,
 104         logger=None,
 105         device=torch.device("cpu"),
 106         pruner_train=None,
 107         pruner_eval=None,
 108     ):
 109         def generate_descr(nb, cache_suffix, pruner):
 110             return picoclvr.generate(
 111                 nb,
 112                 height=self.height,
 113                 width=self.width,
 114                 nb_colors=nb_colors,
 115                 pruner=pruner,
 116             )
 117
 118         self.height = height
 119         self.width = width
 120         self.batch_size = batch_size
 121         self.device = device
 122         self.pruner_train = pruner_train
 123         self.pruner_eval = pruner_eval
 124
 125         if logger is not None:
 126             logger(
 127                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 128             )
 129
 130         self.train_descr = generate_descr(
 131             nb_train_samples, "train", pruner=self.pruner_train
 132         )
 133         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 134
 135         # Build the tokenizer
 136         tokens = {"<nul>", "<img>"}
 137         for d in [self.train_descr, self.test_descr]:
 138             for s in d:
 139                 for t in s.strip().split(" "):
 140                     tokens.add(t)
 141         # make this set a sorted list to get the same tensors given
 142         # the same descr
 143         tokens = list(tokens)
 144         tokens.sort()
 145         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 146         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 147         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 148
 149         # Tokenize the train and test sets
 150         self.train_input = self.tensorize(self.train_descr)
 151         self.test_input = self.tensorize(self.test_descr)
 152
 153     def batches(self, split="train"):
 154         assert split in {"train", "test"}
 155         input = self.train_input if split == "train" else self.test_input
 156         for batch in tqdm.tqdm(
 157             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 158         ):
 159             yield self.trim(batch)
 160
 161     def vocabulary_size(self):
 162         return len(self.token2id)
 163
 164     def compute_missing_properties(
 165         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 166     ):
 167         acc_nb_requested_properties = []
 168         acc_nb_missing_properties = []
 169         acc_nb_results = 0
 170
 171         for input in tqdm.tqdm(
 172             self.test_input.split(self.batch_size),
 173             dynamic_ncols=True,
 174             desc=f"test-properties",
 175         ):
 176             result = input.clone()
 177             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 178             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 179             masked_inplace_autoregression(
 180                 model,
 181                 self.batch_size,
 182                 result,
 183                 ar_mask,
 184                 deterministic_synthesis,
 185                 progress_bar_desc=None,
 186                 device=self.device,
 187             )
 188
 189             result_descr = self.detensorize(result)
 190             np = picoclvr.nb_properties(
 191                 result_descr,
 192                 height=self.height,
 193                 width=self.width,
 194                 pruner=pruner,
 195             )
 196             nb_requested_properties, _, nb_missing_properties = zip(*np)
 197             acc_nb_requested_properties += nb_requested_properties
 198             acc_nb_missing_properties += nb_missing_properties
 199             acc_nb_results += len(result_descr)
 200
 201         nb_requested_properties = sum(acc_nb_requested_properties)
 202         nb_missing_properties = sum(acc_nb_missing_properties)
 203
 204         prefix = "" if pruner is None else "pruned_"
 205         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 206         logger(
 207             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 208         )
 209         logger(
 210             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 211         )
 212
 213     ######################################################################
 214
 215     def produce_results(
 216         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 217     ):
 218         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 219
 220         if self.pruner_eval is not None:
 221             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 222
 223         nb_tokens_to_generate = self.height * self.width + 3
 224         result_descr = []
 225         nb_per_primer = 8
 226         primer = []
 227
 228         for primer_descr in [
 229             "red above green <sep> green top <sep> blue right of red",
 230             "there is red <sep> there is yellow <sep> there is blue",
 231             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 232             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 233         ]:
 234             primer += [primer_descr + " <img>"] * nb_per_primer
 235
 236         result = self.tensorize(primer)
 237         fill = result.new_full(
 238             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 239         )
 240         result = torch.cat((result, fill), 1)
 241         ar_mask = (result == self.t_nul).long()
 242         masked_inplace_autoregression(
 243             model,
 244             self.batch_size,
 245             result,
 246             ar_mask,
 247             deterministic_synthesis,
 248             device=self.device,
 249         )
 250         result_descr = self.detensorize(result)
 251
 252         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 253
 254         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 255         acc_nb_results = len(result_descr)
 256
 257         nb_requested_properties = sum(acc_nb_requested_properties)
 258         nb_missing_properties = sum(acc_nb_missing_properties)
 259
 260         prefix = "demo_"
 261         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 262         logger(
 263             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 264         )
 265         logger(
 266             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 267         )
 268
 269         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 270
 271         if img.dim() == 5:
 272             if img.size(1) == 1:
 273                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 274             else:
 275                 img = torch.cat(
 276                     [
 277                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 278                         for x in img
 279                     ],
 280                     0,
 281                 )
 282
 283         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 284         torchvision.utils.save_image(
 285             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 286         )
 287         logger(f"wrote {image_name}")
 288
 289
 290 ######################################################################
 291
 292
 293 class MNIST(Task):
 294     def __init__(
 295         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 296     ):
 297         self.nb_train_samples = (nb_train_samples,)
 298         self.nb_test_samples = (nb_test_samples,)
 299         self.batch_size = batch_size
 300         self.device = device
 301         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 302         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 303         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 304         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 305
 306     def batches(self, split="train", nb_to_use=-1, desc=None):
 307         assert split in {"train", "test"}
 308         input = self.train_input if split == "train" else self.test_input
 309         if nb_to_use > 0:
 310             input = input[:nb_to_use]
 311         if desc is None:
 312             desc = f"epoch-{split}"
 313         for batch in tqdm.tqdm(
 314             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 315         ):
 316             yield batch
 317
 318     def vocabulary_size(self):
 319         return 256
 320
 321     def produce_results(
 322         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 323     ):
 324         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 325         ar_mask = torch.full_like(results, 1)
 326         masked_inplace_autoregression(
 327             model,
 328             self.batch_size,
 329             results,
 330             ar_mask,
 331             deterministic_synthesis,
 332             device=self.device,
 333         )
 334         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 335         torchvision.utils.save_image(
 336             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 337             image_name,
 338             nrow=16,
 339             pad_value=0.8,
 340         )
 341         logger(f"wrote {image_name}")
 342
 343
 344 ######################################################################
 345
 346 import maze
 347
 348
 349 class Maze(Task):
 350     def map2seq(self, *m):
 351         return torch.cat([x.flatten(1) for x in m], 1)
 352
 353     def seq2map(self, s):
 354         s = s.reshape(s.size(0), -1, self.height, self.width)
 355         return (s[:, k] for k in range(s.size(1)))
 356
 357     def __init__(
 358         self,
 359         nb_train_samples,
 360         nb_test_samples,
 361         batch_size,
 362         height,
 363         width,
 364         nb_walls,
 365         device=torch.device("cpu"),
 366     ):
 367         self.batch_size = batch_size
 368         self.height = height
 369         self.width = width
 370         self.device = device
 371
 372         train_mazes, train_paths, _ = maze.create_maze_data(
 373             nb_train_samples,
 374             height=height,
 375             width=width,
 376             nb_walls=nb_walls,
 377             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 378         )
 379         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 380
 381         test_mazes, test_paths, _ = maze.create_maze_data(
 382             nb_test_samples,
 383             height=height,
 384             width=width,
 385             nb_walls=nb_walls,
 386             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 387         )
 388         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 389
 390         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 391
 392     def batches(self, split="train", nb_to_use=-1, desc=None):
 393         assert split in {"train", "test"}
 394         input = self.train_input if split == "train" else self.test_input
 395         if nb_to_use > 0:
 396             input = input[:nb_to_use]
 397         if desc is None:
 398             desc = f"epoch-{split}"
 399         for batch in tqdm.tqdm(
 400             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 401         ):
 402             yield batch
 403
 404     def vocabulary_size(self):
 405         return self.nb_codes
 406
 407     def compute_error(
 408         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 409     ):
 410         nb_total, nb_correct = 0, 0
 411         count = torch.zeros(
 412             self.width * self.height,
 413             self.width * self.height,
 414             device=self.device,
 415             dtype=torch.int64,
 416         )
 417
 418         for input in self.batches(split, nb_to_use):
 419             result = input.clone()
 420             ar_mask = result.new_zeros(result.size())
 421             ar_mask[:, self.height * self.width :] = 1
 422             result *= 1 - ar_mask
 423             masked_inplace_autoregression(
 424                 model,
 425                 self.batch_size,
 426                 result,
 427                 ar_mask,
 428                 deterministic_synthesis,
 429                 progress_bar_desc=None,
 430                 device=self.device,
 431             )
 432             mazes, paths = self.seq2map(result)
 433             path_correctness = maze.path_correctness(mazes, paths)
 434             nb_correct += path_correctness.long().sum()
 435             nb_total += mazes.size(0)
 436
 437             optimal_path_lengths = (
 438                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 439             )
 440             predicted_path_lengths = (
 441                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 442             )
 443             optimal_path_lengths = optimal_path_lengths[path_correctness]
 444             predicted_path_lengths = predicted_path_lengths[path_correctness]
 445             count[optimal_path_lengths, predicted_path_lengths] += 1
 446
 447         if count.max() == 0:
 448             count = None
 449         else:
 450             count = count[
 451                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 452             ]
 453
 454         return nb_total, nb_correct, count
 455
 456     def produce_results(
 457         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 458     ):
 459         train_nb_total, train_nb_correct, count = self.compute_error(
 460             model,
 461             "train",
 462             nb_to_use=1000,
 463             deterministic_synthesis=deterministic_synthesis,
 464         )
 465         logger(
 466             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}%"
 467         )
 468
 469         test_nb_total, test_nb_correct, count = self.compute_error(
 470             model,
 471             "test",
 472             nb_to_use=1000,
 473             deterministic_synthesis=deterministic_synthesis,
 474         )
 475         logger(
 476             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}%"
 477         )
 478
 479         if count is not None:
 480             proportion_optimal = count.diagonal().sum().float() / count.sum()
 481             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 482             with open(
 483                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 484             ) as f:
 485                 for i in range(count.size(0)):
 486                     for j in range(count.size(1)):
 487                         eol = " " if j < count.size(1) - 1 else "\n"
 488                         f.write(f"{count[i,j]}{eol}")
 489
 490         input = self.test_input[:48]
 491         result = input.clone()
 492         ar_mask = result.new_zeros(result.size())
 493         ar_mask[:, self.height * self.width :] = 1
 494         result *= 1 - ar_mask
 495         masked_inplace_autoregression(
 496             model,
 497             self.batch_size,
 498             result,
 499             ar_mask,
 500             deterministic_synthesis,
 501             device=self.device,
 502         )
 503
 504         mazes, paths = self.seq2map(input)
 505         _, predicted_paths = self.seq2map(result)
 506
 507         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 508         maze.save_image(
 509             filename,
 510             mazes=mazes,
 511             target_paths=paths,
 512             predicted_paths=predicted_paths,
 513             path_correct=maze.path_correctness(mazes, predicted_paths),
 514             path_optimal=maze.path_optimality(paths, predicted_paths),
 515         )
 516         logger(f"wrote {filename}")
 517
 518
 519 ######################################################################
 520
 521
 522 import snake
 523
 524
 525 class Snake(Task):
 526     def __init__(
 527         self,
 528         nb_train_samples,
 529         nb_test_samples,
 530         batch_size,
 531         height,
 532         width,
 533         nb_colors,
 534         length,
 535         prompt_length,
 536         device=torch.device("cpu"),
 537     ):
 538         self.batch_size = batch_size
 539         self.height = height
 540         self.width = width
 541         self.device = device
 542         self.prompt_length = prompt_length
 543
 544         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 545             nb_train_samples,
 546             height,
 547             width,
 548             nb_colors,
 549             length,
 550             prompt_length,
 551             self.device,
 552         )
 553         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 554             nb_test_samples,
 555             height,
 556             width,
 557             nb_colors,
 558             length,
 559             prompt_length,
 560             self.device,
 561         )
 562
 563         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 564
 565     def batches(self, split="train", nb_to_use=-1, desc=None):
 566         assert split in {"train", "test"}
 567         input = self.train_input if split == "train" else self.test_input
 568         if nb_to_use > 0:
 569             input = input[:nb_to_use]
 570         if desc is None:
 571             desc = f"epoch-{split}"
 572         for batch in tqdm.tqdm(
 573             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 574         ):
 575             yield batch
 576
 577     def vocabulary_size(self):
 578         return self.nb_codes
 579
 580     def produce_results(
 581         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 582     ):
 583         def compute_nb_correct(input, prior_visits):
 584             result = input.clone()
 585             i = torch.arange(result.size(1), device=result.device)[None, :]
 586             ar_mask = (
 587                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 588                 .long()
 589                 .expand_as(result)
 590             )
 591             result *= 1 - ar_mask
 592
 593             # snake.solver(result,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             # nb_total = result.size(0)
 609             # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 610
 611             return nb_total, nb_correct
 612
 613         # train_nb_total, train_nb_correct = compute_nb_correct(
 614         # self.train_input, self.train_prior_visits
 615         # )
 616
 617         # logger(
 618         # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 619         # )
 620
 621         test_nb_total, test_nb_correct = compute_nb_correct(
 622             self.test_input[:1000], self.test_prior_visits[:1000]
 623         )
 624
 625         logger(
 626             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}%"
 627         )
 628
 629
 630 ######################################################################
 631
 632
 633 import stack
 634
 635
 636 class Stack(Task):
 637     def __init__(
 638         self,
 639         nb_train_samples,
 640         nb_test_samples,
 641         batch_size,
 642         logger,
 643         nb_steps,
 644         nb_stacks,
 645         nb_digits,
 646         fraction_values_for_train=None,
 647         device=torch.device("cpu"),
 648     ):
 649         self.batch_size = batch_size
 650         self.nb_steps = nb_steps
 651         self.nb_stacks = nb_stacks
 652         self.nb_digits = nb_digits
 653         self.device = device
 654
 655         if fraction_values_for_train is None:
 656             values_for_train = None
 657             values_for_test = None
 658         else:
 659             all = torch.randperm(10**nb_digits)
 660             nb_for_train = int(all.size(0) * fraction_values_for_train)
 661             values_for_train = all[:nb_for_train]
 662             values_for_test = all[nb_for_train:]
 663
 664         self.train_input, self.train_stack_counts = stack.generate_sequences(
 665             nb_train_samples,
 666             nb_steps,
 667             nb_stacks,
 668             nb_digits,
 669             values_for_train,
 670             self.device,
 671         )
 672
 673         self.test_input, self.test_stack_counts = stack.generate_sequences(
 674             nb_test_samples,
 675             nb_steps,
 676             nb_stacks,
 677             nb_digits,
 678             values_for_test,
 679             self.device,
 680         )
 681
 682         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 683         counts = self.test_stack_counts.flatten()[i.flatten()]
 684         counts = F.one_hot(counts).sum(0)
 685         logger(f"test_pop_stack_counts {counts}")
 686
 687         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 688
 689     def batches(self, split="train", nb_to_use=-1, desc=None):
 690         assert split in {"train", "test"}
 691         input = self.train_input if split == "train" else self.test_input
 692         if nb_to_use > 0:
 693             input = input[:nb_to_use]
 694         if desc is None:
 695             desc = f"epoch-{split}"
 696         for batch in tqdm.tqdm(
 697             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 698         ):
 699             yield batch
 700
 701     def vocabulary_size(self):
 702         return self.nb_codes
 703
 704     def produce_results(
 705         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 706     ):
 707         def compute_nb_correct(input):
 708             result = input.clone()
 709             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 710             ar_mask = (result != input).long()
 711             masked_inplace_autoregression(
 712                 model,
 713                 self.batch_size,
 714                 result,
 715                 ar_mask,
 716                 deterministic_synthesis,
 717                 device=self.device,
 718             )
 719
 720             errors = ((result != input).long() * ar_mask).reshape(
 721                 -1, 1 + self.nb_digits
 722             )
 723             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 724
 725             nb_total = ar_mask.max(1).values.sum()
 726             nb_correct = nb_total - errors.max(1).values.sum()
 727
 728             return nb_total, nb_correct
 729
 730         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 731
 732         logger(
 733             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}%"
 734         )
 735
 736         ##############################################################
 737         # Log a few generated sequences
 738         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 739         result = input.clone()
 740         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 741         ar_mask = (result != input).long()
 742
 743         # for n in range(result.size(0)):
 744         # logger(
 745         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 746         # )
 747
 748         masked_inplace_autoregression(
 749             model,
 750             self.batch_size,
 751             result,
 752             ar_mask,
 753             deterministic_synthesis,
 754             device=self.device,
 755         )
 756
 757         for n in range(result.size(0)):
 758             logger(
 759                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 760             )
 761         ##############################################################
 762
 763
 764 ######################################################################
 765
 766
 767 import expr
 768
 769
 770 class Expr(Task):
 771     def tensorize(self, sequences):
 772         len_max = max([len(x) for x in sequences])
 773         return torch.cat(
 774             [
 775                 torch.tensor(
 776                     [
 777                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
 778                         for s in sequences
 779                     ]
 780                 )
 781             ],
 782             0,
 783         ).to(self.device)
 784
 785     def __init__(
 786         self,
 787         nb_train_samples,
 788         nb_test_samples,
 789         nb_variables,
 790         sequence_length,
 791         operand_max,
 792         result_max,
 793         batch_size,
 794         device=torch.device("cpu"),
 795     ):
 796         self.batch_size = batch_size
 797         self.device = device
 798
 799         train_sequences = expr.generate_sequences(
 800             nb_train_samples,
 801             nb_variables=nb_variables,
 802             length=sequence_length,
 803             operand_max=operand_max,
 804             result_max=result_max,
 805         )
 806
 807         test_sequences = expr.generate_sequences(
 808             nb_test_samples,
 809             nb_variables=nb_variables,
 810             length=sequence_length,
 811             operand_max=operand_max,
 812             result_max=result_max,
 813         )
 814
 815         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
 816         symbols.sort()
 817
 818         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
 819         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 820
 821         self.filler, self.space = self.char2id["#"], self.char2id[" "]
 822
 823         self.train_input = self.tensorize(train_sequences)
 824         self.test_input = self.tensorize(test_sequences)
 825
 826         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 827
 828     def batches(self, split="train", nb_to_use=-1, desc=None):
 829         assert split in {"train", "test"}
 830         input = self.train_input if split == "train" else self.test_input
 831         if nb_to_use > 0:
 832             input = input[:nb_to_use]
 833         if desc is None:
 834             desc = f"epoch-{split}"
 835         for batch in tqdm.tqdm(
 836             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 837         ):
 838             last = (batch != self.filler).max(0).values.nonzero().max() + 3
 839             batch = batch[:, :last]
 840             yield batch
 841
 842     def vocabulary_size(self):
 843         return self.nb_codes
 844
 845     def seq2str(self, s):
 846         return "".join([self.id2char[k.item()] for k in s])
 847
 848     def produce_results(
 849         self,
 850         n_epoch,
 851         model,
 852         result_dir,
 853         logger,
 854         deterministic_synthesis,
 855         input_file=None,
 856     ):
 857         def compute_nb_correct(input):
 858             result = input.clone()
 859             s = (result == self.space).long()
 860             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
 861             result = (1 - ar_mask) * result + ar_mask * self.filler
 862             masked_inplace_autoregression(
 863                 model,
 864                 self.batch_size,
 865                 result,
 866                 ar_mask,
 867                 deterministic_synthesis,
 868                 device=self.device,
 869             )
 870
 871             nb_total = input.size(0)
 872             nb_correct = (input == result).long().min(1).values.sum()
 873
 874             #######################################################################
 875             # Comput predicted vs. true variable values
 876
 877             nb_delta = torch.zeros(5, dtype=torch.int64)
 878             nb_missed = 0
 879
 880             values_input = expr.extract_results([self.seq2str(s) for s in input])
 881             values_result = expr.extract_results([self.seq2str(s) for s in result])
 882
 883             for i, r in zip(values_input, values_result):
 884                 for n, vi in i.items():
 885                     vr = r.get(n)
 886                     if vr is None or vr < 0:
 887                         nb_missed += 1
 888                     else:
 889                         d = abs(vr - vi)
 890                         if d >= nb_delta.size(0):
 891                             nb_missed += 1
 892                         else:
 893                             nb_delta[d] += 1
 894
 895             ######################################################################
 896
 897             return nb_total, nb_correct, nb_delta, nb_missed
 898
 899         (
 900             test_nb_total,
 901             test_nb_correct,
 902             test_nb_delta,
 903             test_nb_missed,
 904         ) = compute_nb_correct(self.test_input[:10000])
 905
 906         logger(
 907             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}%"
 908         )
 909
 910         nb_total = test_nb_delta.sum() + test_nb_missed
 911         for d in range(test_nb_delta.size(0)):
 912             logger(
 913                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
 914             )
 915         logger(
 916             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
 917         )
 918
 919         ##############################################################
 920         # Log a few generated sequences
 921         if input_file is None:
 922             input = self.test_input[:10]
 923         else:
 924             with open(input_file, "r") as f:
 925                 sequences = [e.strip() for e in f.readlines()]
 926                 sequences = [s + " " + "#" * 50 for s in sequences]
 927                 input = self.tensorize(sequences)
 928
 929         result = input.clone()
 930         s = (result == self.space).long()
 931         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
 932         result = (1 - ar_mask) * result + ar_mask * self.filler
 933
 934         for n in range(result.size(0)):
 935             logger(f"test_before {self.seq2str(result[n])}")
 936
 937         masked_inplace_autoregression(
 938             model,
 939             self.batch_size,
 940             result,
 941             ar_mask,
 942             deterministic_synthesis,
 943             device=self.device,
 944         )
 945
 946         correct = (1 - ar_mask) * self.space + ar_mask * input
 947         for n in range(result.size(0)):
 948             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
 949             logger(f"test_after  {self.seq2str(result[n])} {comment}")
 950             logger(f"truth       {self.seq2str(correct[n])}")
 951         ##############################################################
 952
 953
 954 ######################################################################