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     for input, ar_mask in batches:
  34         model.masked_inplace_autoregression(
  35             input, ar_mask, forbidden_tokens, deterministic_synthesis
  36         )
  37
  38
  39 class Task:
  40     def batches(self, split="train"):
  41         pass
  42
  43     def vocabulary_size(self):
  44         pass
  45
  46     def produce_results(
  47         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  48     ):
  49         pass
  50
  51
  52 ######################################################################
  53
  54 import picoclvr
  55
  56
  57 class PicoCLVR(Task):
  58     # Make a tensor from a list of strings
  59     def tensorize(self, descr):
  60         token_descr = [s.strip().split(" ") for s in descr]
  61         l = max([len(s) for s in token_descr])
  62         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
  63         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
  64         return torch.tensor(id_descr, device=self.device)
  65
  66     # Make a list of strings from a tensor
  67     def detensorize(self, x):
  68         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
  69
  70     # trim all the tensors in the tuple z to remove as much token from
  71     # left and right in the first tensor. If z is a tuple, all its
  72     # elements are trimed according to the triming for the first
  73     def trim(self, z, token="<nul>"):
  74         n = self.token2id[token]
  75         if type(z) == tuple:
  76             x = z[0]
  77             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  78             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  79             return tuple([t[:, a:b] for t in z])
  80         else:
  81             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  82             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  83             return z[:, a:b]
  84
  85     ######################
  86
  87     def __init__(
  88         self,
  89         nb_train_samples,
  90         nb_test_samples,
  91         batch_size,
  92         height,
  93         width,
  94         nb_colors=5,
  95         logger=None,
  96         device=torch.device("cpu"),
  97         pruner_train=None,
  98         pruner_eval=None,
  99     ):
 100         def generate_descr(nb, cache_suffix, pruner):
 101             return picoclvr.generate(
 102                 nb,
 103                 height=self.height,
 104                 width=self.width,
 105                 nb_colors=nb_colors,
 106                 pruner=pruner,
 107             )
 108
 109         self.height = height
 110         self.width = width
 111         self.batch_size = batch_size
 112         self.device = device
 113         self.pruner_train = pruner_train
 114         self.pruner_eval = pruner_eval
 115
 116         if logger is not None:
 117             logger(
 118                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 119             )
 120
 121         self.train_descr = generate_descr(
 122             nb_train_samples, "train", pruner=self.pruner_train
 123         )
 124         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 125
 126         # Build the tokenizer
 127         tokens = {"<nul>", "<img>"}
 128         for d in [self.train_descr, self.test_descr]:
 129             for s in d:
 130                 for t in s.strip().split(" "):
 131                     tokens.add(t)
 132         # make this set a sorted list to get the same tensors given
 133         # the same descr
 134         tokens = list(tokens)
 135         tokens.sort()
 136         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 137         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 138         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 139
 140         # Tokenize the train and test sets
 141         self.train_input = self.tensorize(self.train_descr)
 142         self.test_input = self.tensorize(self.test_descr)
 143
 144     def batches(self, split="train"):
 145         assert split in {"train", "test"}
 146         input = self.train_input if split == "train" else self.test_input
 147         for batch in tqdm.tqdm(
 148             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 149         ):
 150             yield self.trim(batch)
 151
 152     def vocabulary_size(self):
 153         return len(self.token2id)
 154
 155     def compute_missing_properties(
 156         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 157     ):
 158         acc_nb_requested_properties = []
 159         acc_nb_missing_properties = []
 160         acc_nb_results = 0
 161
 162         for input in tqdm.tqdm(
 163             self.test_input.split(self.batch_size),
 164             dynamic_ncols=True,
 165             desc=f"test-properties",
 166         ):
 167             result = input.clone()
 168             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 169             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 170             masked_inplace_autoregression(
 171                 model,
 172                 self.batch_size,
 173                 result,
 174                 ar_mask,
 175                 deterministic_synthesis,
 176                 progress_bar_desc=None,
 177                 device=self.device,
 178             )
 179
 180             result_descr = self.detensorize(result)
 181             np = picoclvr.nb_properties(
 182                 result_descr,
 183                 height=self.height,
 184                 width=self.width,
 185                 pruner=pruner,
 186             )
 187             nb_requested_properties, _, nb_missing_properties = zip(*np)
 188             acc_nb_requested_properties += nb_requested_properties
 189             acc_nb_missing_properties += nb_missing_properties
 190             acc_nb_results += len(result_descr)
 191
 192         nb_requested_properties = sum(acc_nb_requested_properties)
 193         nb_missing_properties = sum(acc_nb_missing_properties)
 194
 195         prefix = "" if pruner is None else "pruned_"
 196         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 197         logger(
 198             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 199         )
 200         logger(
 201             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 202         )
 203
 204     ######################################################################
 205
 206     def produce_results(
 207         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 208     ):
 209         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 210
 211         if self.pruner_eval is not None:
 212             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 213
 214         nb_tokens_to_generate = self.height * self.width + 3
 215         result_descr = []
 216         nb_per_primer = 8
 217         primer = []
 218
 219         for primer_descr in [
 220             "red above green <sep> green top <sep> blue right of red",
 221             "there is red <sep> there is yellow <sep> there is blue",
 222             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 223             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 224         ]:
 225             primer += [primer_descr + " <img>"] * nb_per_primer
 226
 227         result = self.tensorize(primer)
 228         fill = result.new_full(
 229             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 230         )
 231         result = torch.cat((result, fill), 1)
 232         ar_mask = (result == self.t_nul).long()
 233         masked_inplace_autoregression(
 234             model,
 235             self.batch_size,
 236             result,
 237             ar_mask,
 238             deterministic_synthesis,
 239             device=self.device,
 240         )
 241         result_descr = self.detensorize(result)
 242
 243         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 244
 245         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 246         acc_nb_results = len(result_descr)
 247
 248         nb_requested_properties = sum(acc_nb_requested_properties)
 249         nb_missing_properties = sum(acc_nb_missing_properties)
 250
 251         prefix = "demo_"
 252         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 253         logger(
 254             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 255         )
 256         logger(
 257             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 258         )
 259
 260         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 261
 262         if img.dim() == 5:
 263             if img.size(1) == 1:
 264                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 265             else:
 266                 img = torch.cat(
 267                     [
 268                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 269                         for x in img
 270                     ],
 271                     0,
 272                 )
 273
 274         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 275         torchvision.utils.save_image(
 276             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 277         )
 278         logger(f"wrote {image_name}")
 279
 280
 281 ######################################################################
 282
 283
 284 class MNIST(Task):
 285     def __init__(
 286         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 287     ):
 288         self.nb_train_samples = (nb_train_samples,)
 289         self.nb_test_samples = (nb_test_samples,)
 290         self.batch_size = batch_size
 291         self.device = device
 292         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 293         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 294         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 295         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 296
 297     def batches(self, split="train", nb_to_use=-1, desc=None):
 298         assert split in {"train", "test"}
 299         input = self.train_input if split == "train" else self.test_input
 300         if nb_to_use > 0:
 301             input = input[:nb_to_use]
 302         if desc is None:
 303             desc = f"epoch-{split}"
 304         for batch in tqdm.tqdm(
 305             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 306         ):
 307             yield batch
 308
 309     def vocabulary_size(self):
 310         return 256
 311
 312     def produce_results(
 313         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 314     ):
 315         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 316         ar_mask = torch.full_like(results, 1)
 317         masked_inplace_autoregression(
 318             model,
 319             self.batch_size,
 320             results,
 321             ar_mask,
 322             deterministic_synthesis,
 323             device=self.device,
 324         )
 325         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 326         torchvision.utils.save_image(
 327             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 328             image_name,
 329             nrow=16,
 330             pad_value=0.8,
 331         )
 332         logger(f"wrote {image_name}")
 333
 334
 335 ######################################################################
 336
 337 import maze
 338
 339
 340 class Maze(Task):
 341     def map2seq(self, *m):
 342         return torch.cat([x.flatten(1) for x in m], 1)
 343
 344     def seq2map(self, s):
 345         s = s.reshape(s.size(0), -1, self.height, self.width)
 346         return (s[:, k] for k in range(s.size(1)))
 347
 348     def __init__(
 349         self,
 350         nb_train_samples,
 351         nb_test_samples,
 352         batch_size,
 353         height,
 354         width,
 355         nb_walls,
 356         device=torch.device("cpu"),
 357     ):
 358         self.batch_size = batch_size
 359         self.height = height
 360         self.width = width
 361         self.device = device
 362
 363         train_mazes, train_paths, _ = maze.create_maze_data(
 364             nb_train_samples,
 365             height=height,
 366             width=width,
 367             nb_walls=nb_walls,
 368             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 369         )
 370         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 371
 372         test_mazes, test_paths, _ = maze.create_maze_data(
 373             nb_test_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-test"),
 378         )
 379         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 380
 381         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 382
 383     def batches(self, split="train", nb_to_use=-1, desc=None):
 384         assert split in {"train", "test"}
 385         input = self.train_input if split == "train" else self.test_input
 386         if nb_to_use > 0:
 387             input = input[:nb_to_use]
 388         if desc is None:
 389             desc = f"epoch-{split}"
 390         for batch in tqdm.tqdm(
 391             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 392         ):
 393             yield batch
 394
 395     def vocabulary_size(self):
 396         return self.nb_codes
 397
 398     def compute_error(
 399         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 400     ):
 401         nb_total, nb_correct = 0, 0
 402         count = torch.zeros(
 403             self.width * self.height,
 404             self.width * self.height,
 405             device=self.device,
 406             dtype=torch.int64,
 407         )
 408
 409         for input in self.batches(split, nb_to_use):
 410             result = input.clone()
 411             ar_mask = result.new_zeros(result.size())
 412             ar_mask[:, self.height * self.width :] = 1
 413             result *= 1 - ar_mask
 414             masked_inplace_autoregression(
 415                 model,
 416                 self.batch_size,
 417                 result,
 418                 ar_mask,
 419                 deterministic_synthesis,
 420                 progress_bar_desc=None,
 421                 device=self.device,
 422             )
 423             mazes, paths = self.seq2map(result)
 424             path_correctness = maze.path_correctness(mazes, paths)
 425             nb_correct += path_correctness.long().sum()
 426             nb_total += mazes.size(0)
 427
 428             optimal_path_lengths = (
 429                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 430             )
 431             predicted_path_lengths = (
 432                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 433             )
 434             optimal_path_lengths = optimal_path_lengths[path_correctness]
 435             predicted_path_lengths = predicted_path_lengths[path_correctness]
 436             count[optimal_path_lengths, predicted_path_lengths] += 1
 437
 438         if count.max() == 0:
 439             count = None
 440         else:
 441             count = count[
 442                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 443             ]
 444
 445         return nb_total, nb_correct, count
 446
 447     def produce_results(
 448         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 449     ):
 450         with torch.autograd.no_grad():
 451             t = model.training
 452             model.eval()
 453
 454             train_nb_total, train_nb_correct, count = self.compute_error(
 455                 model,
 456                 "train",
 457                 nb_to_use=1000,
 458                 deterministic_synthesis=deterministic_synthesis,
 459             )
 460             logger(
 461                 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}%"
 462             )
 463
 464             test_nb_total, test_nb_correct, count = self.compute_error(
 465                 model,
 466                 "test",
 467                 nb_to_use=1000,
 468                 deterministic_synthesis=deterministic_synthesis,
 469             )
 470             logger(
 471                 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}%"
 472             )
 473
 474             if count is not None:
 475                 proportion_optimal = count.diagonal().sum().float() / count.sum()
 476                 logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 477                 with open(
 478                     os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 479                 ) as f:
 480                     for i in range(count.size(0)):
 481                         for j in range(count.size(1)):
 482                             eol = " " if j < count.size(1) - 1 else "\n"
 483                             f.write(f"{count[i,j]}{eol}")
 484
 485             input = self.test_input[:48]
 486             result = input.clone()
 487             ar_mask = result.new_zeros(result.size())
 488             ar_mask[:, self.height * self.width :] = 1
 489             result *= 1 - ar_mask
 490             masked_inplace_autoregression(
 491                 model,
 492                 self.batch_size,
 493                 result,
 494                 ar_mask,
 495                 deterministic_synthesis,
 496                 device=self.device,
 497             )
 498
 499             mazes, paths = self.seq2map(input)
 500             _, predicted_paths = self.seq2map(result)
 501
 502             filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 503             maze.save_image(
 504                 filename,
 505                 mazes=mazes,
 506                 target_paths=paths,
 507                 predicted_paths=predicted_paths,
 508                 path_correct=maze.path_correctness(mazes, predicted_paths),
 509                 path_optimal=maze.path_optimality(paths, predicted_paths),
 510             )
 511             logger(f"wrote {filename}")
 512
 513             model.train(t)
 514
 515
 516 ######################################################################
 517
 518
 519 import snake
 520
 521
 522 class Snake(Task):
 523     def __init__(
 524         self,
 525         nb_train_samples,
 526         nb_test_samples,
 527         batch_size,
 528         height,
 529         width,
 530         nb_colors,
 531         length,
 532         prompt_length,
 533         device=torch.device("cpu"),
 534     ):
 535         self.batch_size = batch_size
 536         self.height = height
 537         self.width = width
 538         self.device = device
 539         self.prompt_length = prompt_length
 540
 541         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 542             nb_train_samples,
 543             height,
 544             width,
 545             nb_colors,
 546             length,
 547             prompt_length,
 548             self.device,
 549         )
 550         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 551             nb_test_samples,
 552             height,
 553             width,
 554             nb_colors,
 555             length,
 556             prompt_length,
 557             self.device,
 558         )
 559
 560         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 561
 562     def batches(self, split="train", nb_to_use=-1, desc=None):
 563         assert split in {"train", "test"}
 564         input = self.train_input if split == "train" else self.test_input
 565         if nb_to_use > 0:
 566             input = input[:nb_to_use]
 567         if desc is None:
 568             desc = f"epoch-{split}"
 569         for batch in tqdm.tqdm(
 570             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 571         ):
 572             yield batch
 573
 574     def vocabulary_size(self):
 575         return self.nb_codes
 576
 577     def produce_results(
 578         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 579     ):
 580         with torch.autograd.no_grad():
 581             t = model.training
 582             model.eval()
 583
 584             def compute_nb_correct(input, prior_visits):
 585                 result = input.clone()
 586                 i = torch.arange(result.size(1), device=result.device)[None, :]
 587                 ar_mask = (
 588                     torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 589                     .long()
 590                     .expand_as(result)
 591                 )
 592                 result *= 1 - ar_mask
 593
 594                 # snake.solver(result,ar_mask)
 595
 596                 masked_inplace_autoregression(
 597                     model,
 598                     self.batch_size,
 599                     result,
 600                     ar_mask,
 601                     deterministic_synthesis,
 602                     device=self.device,
 603                 )
 604
 605                 nb_total = ((prior_visits > 0) * ar_mask).sum()
 606
 607                 nb_correct = (
 608                     (result == input).long() * (prior_visits > 0) * ar_mask
 609                 ).sum()
 610
 611                 # nb_total = result.size(0)
 612                 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 613
 614                 return nb_total, nb_correct
 615
 616             # train_nb_total, train_nb_correct = compute_nb_correct(
 617             # self.train_input, self.train_prior_visits
 618             # )
 619
 620             # logger(
 621             # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 622             # )
 623
 624             test_nb_total, test_nb_correct = compute_nb_correct(
 625                 self.test_input[:1000], self.test_prior_visits[:1000]
 626             )
 627
 628             logger(
 629                 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}%"
 630             )
 631
 632             model.train(t)
 633
 634
 635 ######################################################################
 636
 637
 638 import stack
 639
 640
 641 class Stack(Task):
 642     def __init__(
 643         self,
 644         nb_train_samples,
 645         nb_test_samples,
 646         batch_size,
 647         logger,
 648         nb_steps,
 649         nb_stacks,
 650         nb_digits,
 651         fraction_values_for_train=None,
 652         device=torch.device("cpu"),
 653     ):
 654         self.batch_size = batch_size
 655         self.nb_steps = nb_steps
 656         self.nb_stacks = nb_stacks
 657         self.nb_digits = nb_digits
 658         self.device = device
 659
 660         if fraction_values_for_train is None:
 661             values_for_train = None
 662             values_for_test = None
 663         else:
 664             all = torch.randperm(10**nb_digits)
 665             nb_for_train = int(all.size(0) * fraction_values_for_train)
 666             values_for_train = all[:nb_for_train]
 667             values_for_test = all[nb_for_train:]
 668
 669         self.train_input, self.train_stack_counts = stack.generate_sequences(
 670             nb_train_samples,
 671             nb_steps,
 672             nb_stacks,
 673             nb_digits,
 674             values_for_train,
 675             self.device,
 676         )
 677
 678         self.test_input, self.test_stack_counts = stack.generate_sequences(
 679             nb_test_samples,
 680             nb_steps,
 681             nb_stacks,
 682             nb_digits,
 683             values_for_test,
 684             self.device,
 685         )
 686
 687         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 688         counts = self.test_stack_counts.flatten()[i.flatten()]
 689         counts = F.one_hot(counts).sum(0)
 690         logger(f"test_pop_stack_counts {counts}")
 691
 692         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 693
 694     def batches(self, split="train", nb_to_use=-1, desc=None):
 695         assert split in {"train", "test"}
 696         input = self.train_input if split == "train" else self.test_input
 697         if nb_to_use > 0:
 698             input = input[:nb_to_use]
 699         if desc is None:
 700             desc = f"epoch-{split}"
 701         for batch in tqdm.tqdm(
 702             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 703         ):
 704             yield batch
 705
 706     def vocabulary_size(self):
 707         return self.nb_codes
 708
 709     def produce_results(
 710         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 711     ):
 712         with torch.autograd.no_grad():
 713             t = model.training
 714             model.eval()
 715
 716             def compute_nb_correct(input):
 717                 result = input.clone()
 718                 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 719                 ar_mask = (result != input).long()
 720                 masked_inplace_autoregression(
 721                     model,
 722                     self.batch_size,
 723                     result,
 724                     ar_mask,
 725                     deterministic_synthesis,
 726                     device=self.device,
 727                 )
 728
 729                 errors = ((result != input).long() * ar_mask).reshape(
 730                     -1, 1 + self.nb_digits
 731                 )
 732                 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 733
 734                 nb_total = ar_mask.max(1).values.sum()
 735                 nb_correct = nb_total - errors.max(1).values.sum()
 736
 737                 return nb_total, nb_correct
 738
 739             test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 740
 741             logger(
 742                 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}%"
 743             )
 744
 745             ##############################################################
 746             # Log a few generated sequences
 747             input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 748             result = input.clone()
 749             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 750             ar_mask = (result != input).long()
 751
 752             # for n in range(result.size(0)):
 753             # logger(
 754             # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 755             # )
 756
 757             masked_inplace_autoregression(
 758                 model,
 759                 self.batch_size,
 760                 result,
 761                 ar_mask,
 762                 deterministic_synthesis,
 763                 device=self.device,
 764             )
 765
 766             for n in range(result.size(0)):
 767                 logger(
 768                     f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 769                 )
 770             ##############################################################
 771
 772             model.train(t)
 773
 774
 775 ######################################################################
 776
 777
 778 import expr
 779
 780
 781 class Expr(Task):
 782     def tensorize(self, sequences):
 783         len_max = max([len(x) for x in sequences])
 784         return torch.cat(
 785             [
 786                 torch.tensor(
 787                     [
 788                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
 789                         for s in sequences
 790                     ]
 791                 )
 792             ],
 793             0,
 794         ).to(self.device)
 795
 796     def __init__(
 797         self,
 798         nb_train_samples,
 799         nb_test_samples,
 800         nb_variables,
 801         sequence_length,
 802         batch_size,
 803         device=torch.device("cpu"),
 804     ):
 805         self.batch_size = batch_size
 806         self.device = device
 807
 808         train_sequences = expr.generate_sequences(
 809             nb_train_samples,
 810             nb_variables=nb_variables,
 811             length=sequence_length,
 812             # length=2 * sequence_length,
 813             # randomize_length=True,
 814         )
 815         test_sequences = expr.generate_sequences(
 816             nb_test_samples,
 817             nb_variables=nb_variables,
 818             length=sequence_length,
 819         )
 820
 821         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
 822         symbols.sort()
 823
 824         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
 825         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 826
 827         self.filler, self.space = self.char2id["#"], self.char2id[" "]
 828
 829         self.train_input = self.tensorize(train_sequences)
 830         self.test_input = self.tensorize(test_sequences)
 831
 832         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 833
 834     def batches(self, split="train", nb_to_use=-1, desc=None):
 835         assert split in {"train", "test"}
 836         input = self.train_input if split == "train" else self.test_input
 837         if nb_to_use > 0:
 838             input = input[:nb_to_use]
 839         if desc is None:
 840             desc = f"epoch-{split}"
 841         for batch in tqdm.tqdm(
 842             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 843         ):
 844             if split == "train":
 845                 last = (batch != self.filler).max(0).values.nonzero().max() + 3
 846                 batch = batch[:, :last]
 847             yield batch
 848
 849     def vocabulary_size(self):
 850         return self.nb_codes
 851
 852     def seq2str(self, s):
 853         return "".join([self.id2char[k.item()] for k in s])
 854
 855     def produce_results(
 856         self,
 857         n_epoch,
 858         model,
 859         result_dir,
 860         logger,
 861         deterministic_synthesis,
 862         input_file=None,
 863     ):
 864         with torch.autograd.no_grad():
 865             t = model.training
 866             model.eval()
 867
 868             def compute_nb_correct(input):
 869                 result = input.clone()
 870                 ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
 871                 result = (1 - ar_mask) * result + ar_mask * self.filler
 872                 masked_inplace_autoregression(
 873                     model,
 874                     self.batch_size,
 875                     result,
 876                     ar_mask,
 877                     deterministic_synthesis,
 878                     device=self.device,
 879                 )
 880
 881                 nb_total = input.size(0)
 882                 nb_correct = (input == result).long().min(1).values.sum()
 883
 884                 #######################################################################
 885                 # Comput predicted vs. true variable values
 886
 887                 nb_delta = torch.zeros(5, dtype=torch.int64)
 888                 nb_missed = 0
 889
 890                 values_input = expr.extract_results([self.seq2str(s) for s in input])
 891                 values_result = expr.extract_results([self.seq2str(s) for s in result])
 892
 893                 for i, r in zip(values_input, values_result):
 894                     for n, vi in i.items():
 895                         vr = r.get(n)
 896                         if vr is None or vr < 0:
 897                             nb_missed += 1
 898                         else:
 899                             d = abs(vr - vi)
 900                             if d >= nb_delta.size(0):
 901                                 nb_missed += 1
 902                             else:
 903                                 nb_delta[d] += 1
 904
 905                 ######################################################################
 906
 907                 return nb_total, nb_correct, nb_delta, nb_missed
 908
 909             (
 910                 test_nb_total,
 911                 test_nb_correct,
 912                 test_nb_delta,
 913                 test_nb_missed,
 914             ) = compute_nb_correct(self.test_input[:1000])
 915
 916             logger(
 917                 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}%"
 918             )
 919
 920             nb_total = test_nb_delta.sum() + test_nb_missed
 921             for d in range(test_nb_delta.size(0)):
 922                 logger(
 923                     f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
 924                 )
 925             logger(
 926                 f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
 927             )
 928
 929             ##############################################################
 930             # Log a few generated sequences
 931             if input_file is None:
 932                 input = self.test_input[:10]
 933             else:
 934                 with open(input_file, "r") as f:
 935                     sequences = [e.strip() for e in f.readlines()]
 936                     sequences = [s + " " + "#" * 50 for s in sequences]
 937                     input = self.tensorize(sequences)
 938
 939             result = input.clone()
 940             ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
 941             result = (1 - ar_mask) * result + ar_mask * self.filler
 942
 943             # for n in range(result.size(0)):
 944             # logger(f"test_before {self.seq2str(result[n])}")
 945
 946             masked_inplace_autoregression(
 947                 model,
 948                 self.batch_size,
 949                 result,
 950                 ar_mask,
 951                 deterministic_synthesis,
 952                 device=self.device,
 953             )
 954
 955             correct = (1 - ar_mask) * self.space + ar_mask * input
 956             for n in range(result.size(0)):
 957                 comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
 958                 logger(f"test_after  {self.seq2str(result[n])} {comment}")
 959                 logger(f"correct     {self.seq2str(correct[n])}")
 960             ##############################################################
 961
 962             model.train(t)
 963
 964
 965 ######################################################################