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
  66 class Problem:
  67     def generate(nb):
  68         pass
  69
  70     def perf(seq, logger):
  71         pass
  72
  73
  74 class ProblemByheart(Problem):
  75     def __init__(self):
  76         nb_seq, len_prompt, len_result = 100, 5, 5
  77         self.seq = torch.randint(10, (nb_seq, len_prompt + 1 + len_result))
  78         self.seq[:, len_prompt] = -1
  79
  80     def generate_sequences(self, nb):
  81         return self.seq[torch.randint(self.seq.size(0), (nb,))]
  82
  83
  84 class SandBox(Task):
  85     def __init__(
  86         self,
  87         nb_train_samples,
  88         nb_test_samples,
  89         batch_size,
  90         logger=None,
  91         device=torch.device("cpu"),
  92     ):
  93         super().__init__()
  94
  95         self.batch_size = batch_size
  96
  97         problems = [ProblemByheart()]
  98         nb_common_codes = 100
  99
 100         def generate_sequences(nb_samples):
 101             problem_indexes = torch.randint(len(problems), (nb_samples,))
 102             nb_samples_per_problem = torch.one_hot(problem_indexes).sum(0)
 103             print(f"{nb_samples_per_problem}")
 104             all_seq = []
 105             for nb, p in zip(nb_samples_per_problem, problems):
 106                 all_seq.append(p.generate_sequences(nb_samples_per_problem[nb]))
 107             return all_seq
 108
 109         train_seq = generate_sequences(nb_train_samples)
 110         test_seq = generate_sequences(nb_test_samples)
 111
 112         for strain, stest in zip(train_seq, test_seq):
 113             s = torch.cat((strain, stest), 0)
 114
 115         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 116
 117     def batches(self, split="train", nb_to_use=-1, desc=None):
 118         assert split in {"train", "test"}
 119         input = self.train_input if split == "train" else self.test_input
 120         if nb_to_use > 0:
 121             input = input[:nb_to_use]
 122         if desc is None:
 123             desc = f"epoch-{split}"
 124         for batch in tqdm.tqdm(
 125             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 126         ):
 127             yield batch
 128
 129     def vocabulary_size(self):
 130         return self.nb_codes
 131
 132     def produce_results(
 133         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 134     ):
 135         # logger(
 136         # 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}%"
 137         # )
 138         pass
 139
 140
 141 ######################################################################
 142
 143 import picoclvr
 144
 145
 146 class PicoCLVR(Task):
 147     # Make a tensor from a list of strings
 148     def tensorize(self, descr):
 149         token_descr = [s.strip().split(" ") for s in descr]
 150         l = max([len(s) for s in token_descr])
 151         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 152         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 153         return torch.tensor(id_descr, device=self.device)
 154
 155     # Make a list of strings from a tensor
 156     def detensorize(self, x):
 157         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 158
 159     # trim all the tensors in the tuple z to remove as much token from
 160     # left and right in the first tensor. If z is a tuple, all its
 161     # elements are trimed according to the triming for the first
 162     def trim(self, z, token="<nul>"):
 163         n = self.token2id[token]
 164         if type(z) == tuple:
 165             x = z[0]
 166             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 167             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 168             return tuple([t[:, a:b] for t in z])
 169         else:
 170             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 171             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 172             return z[:, a:b]
 173
 174     ######################
 175
 176     def __init__(
 177         self,
 178         nb_train_samples,
 179         nb_test_samples,
 180         batch_size,
 181         height,
 182         width,
 183         nb_colors=5,
 184         logger=None,
 185         device=torch.device("cpu"),
 186         pruner_train=None,
 187         pruner_eval=None,
 188     ):
 189         super().__init__()
 190
 191         def generate_descr(nb, cache_suffix, pruner):
 192             return picoclvr.generate(
 193                 nb,
 194                 height=self.height,
 195                 width=self.width,
 196                 nb_colors=nb_colors,
 197                 pruner=pruner,
 198             )
 199
 200         self.height = height
 201         self.width = width
 202         self.batch_size = batch_size
 203         self.device = device
 204         self.pruner_train = pruner_train
 205         self.pruner_eval = pruner_eval
 206
 207         if logger is not None:
 208             logger(
 209                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 210             )
 211
 212         self.train_descr = generate_descr(
 213             nb_train_samples, "train", pruner=self.pruner_train
 214         )
 215         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 216
 217         # Build the tokenizer
 218         tokens = {"<nul>", "<img>"}
 219         for d in [self.train_descr, self.test_descr]:
 220             for s in d:
 221                 for t in s.strip().split(" "):
 222                     tokens.add(t)
 223         # make this set a sorted list to get the same tensors given
 224         # the same descr
 225         tokens = list(tokens)
 226         tokens.sort()
 227         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 228         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 229         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 230
 231         # Tokenize the train and test sets
 232         self.train_input = self.tensorize(self.train_descr)
 233         self.test_input = self.tensorize(self.test_descr)
 234
 235     def batches(self, split="train"):
 236         assert split in {"train", "test"}
 237         input = self.train_input if split == "train" else self.test_input
 238         for batch in tqdm.tqdm(
 239             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 240         ):
 241             yield self.trim(batch)
 242
 243     def vocabulary_size(self):
 244         return len(self.token2id)
 245
 246     def compute_missing_properties(
 247         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 248     ):
 249         acc_nb_requested_properties = []
 250         acc_nb_missing_properties = []
 251         acc_nb_results = 0
 252
 253         for input in tqdm.tqdm(
 254             self.test_input.split(self.batch_size),
 255             dynamic_ncols=True,
 256             desc=f"test-properties",
 257         ):
 258             result = input.clone()
 259             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 260             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 261             masked_inplace_autoregression(
 262                 model,
 263                 self.batch_size,
 264                 result,
 265                 ar_mask,
 266                 deterministic_synthesis,
 267                 progress_bar_desc=None,
 268                 device=self.device,
 269             )
 270
 271             result_descr = self.detensorize(result)
 272             np = picoclvr.nb_properties(
 273                 result_descr,
 274                 height=self.height,
 275                 width=self.width,
 276                 pruner=pruner,
 277             )
 278             nb_requested_properties, _, nb_missing_properties = zip(*np)
 279             acc_nb_requested_properties += nb_requested_properties
 280             acc_nb_missing_properties += nb_missing_properties
 281             acc_nb_results += len(result_descr)
 282
 283         nb_requested_properties = sum(acc_nb_requested_properties)
 284         nb_missing_properties = sum(acc_nb_missing_properties)
 285
 286         prefix = "" if pruner is None else "pruned_"
 287         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 288         logger(
 289             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 290         )
 291         logger(
 292             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 293         )
 294
 295     ######################################################################
 296
 297     def produce_results(
 298         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 299     ):
 300         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 301
 302         if self.pruner_eval is not None:
 303             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 304
 305         nb_tokens_to_generate = self.height * self.width + 3
 306         result_descr = []
 307         nb_per_primer = 8
 308         primer = []
 309
 310         for primer_descr in [
 311             "red above green <sep> green top <sep> blue right of red",
 312             "there is red <sep> there is yellow <sep> there is blue",
 313             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 314             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 315         ]:
 316             primer += [primer_descr + " <img>"] * nb_per_primer
 317
 318         result = self.tensorize(primer)
 319         fill = result.new_full(
 320             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 321         )
 322         result = torch.cat((result, fill), 1)
 323         ar_mask = (result == self.t_nul).long()
 324         masked_inplace_autoregression(
 325             model,
 326             self.batch_size,
 327             result,
 328             ar_mask,
 329             deterministic_synthesis,
 330             device=self.device,
 331         )
 332         result_descr = self.detensorize(result)
 333
 334         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 335
 336         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 337         acc_nb_results = len(result_descr)
 338
 339         nb_requested_properties = sum(acc_nb_requested_properties)
 340         nb_missing_properties = sum(acc_nb_missing_properties)
 341
 342         prefix = "demo_"
 343         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 344         logger(
 345             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 346         )
 347         logger(
 348             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 349         )
 350
 351         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 352
 353         if img.dim() == 5:
 354             if img.size(1) == 1:
 355                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 356             else:
 357                 img = torch.cat(
 358                     [
 359                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 360                         for x in img
 361                     ],
 362                     0,
 363                 )
 364
 365         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 366         torchvision.utils.save_image(
 367             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 368         )
 369         logger(f"wrote {image_name}")
 370
 371
 372 ######################################################################
 373
 374
 375 class MNIST(Task):
 376     def __init__(
 377         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 378     ):
 379         super().__init__()
 380
 381         self.nb_train_samples = (nb_train_samples,)
 382         self.nb_test_samples = (nb_test_samples,)
 383         self.batch_size = batch_size
 384         self.device = device
 385         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 386         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 387         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 388         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 389
 390     def batches(self, split="train", nb_to_use=-1, desc=None):
 391         assert split in {"train", "test"}
 392         input = self.train_input if split == "train" else self.test_input
 393         if nb_to_use > 0:
 394             input = input[:nb_to_use]
 395         if desc is None:
 396             desc = f"epoch-{split}"
 397         for batch in tqdm.tqdm(
 398             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 399         ):
 400             yield batch
 401
 402     def vocabulary_size(self):
 403         return 256
 404
 405     def produce_results(
 406         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 407     ):
 408         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 409         ar_mask = torch.full_like(results, 1)
 410         masked_inplace_autoregression(
 411             model,
 412             self.batch_size,
 413             results,
 414             ar_mask,
 415             deterministic_synthesis,
 416             device=self.device,
 417         )
 418         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 419         torchvision.utils.save_image(
 420             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 421             image_name,
 422             nrow=16,
 423             pad_value=0.8,
 424         )
 425         logger(f"wrote {image_name}")
 426
 427
 428 ######################################################################
 429
 430 import maze
 431
 432
 433 class Maze(Task):
 434     def map2seq(self, *m):
 435         return torch.cat([x.flatten(1) for x in m], 1)
 436
 437     def seq2map(self, s):
 438         s = s.reshape(s.size(0), -1, self.height, self.width)
 439         return (s[:, k] for k in range(s.size(1)))
 440
 441     def __init__(
 442         self,
 443         nb_train_samples,
 444         nb_test_samples,
 445         batch_size,
 446         height,
 447         width,
 448         nb_walls,
 449         device=torch.device("cpu"),
 450     ):
 451         super().__init__()
 452
 453         self.batch_size = batch_size
 454         self.height = height
 455         self.width = width
 456         self.device = device
 457
 458         train_mazes, train_paths, _ = maze.create_maze_data(
 459             nb_train_samples,
 460             height=height,
 461             width=width,
 462             nb_walls=nb_walls,
 463             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 464         )
 465         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 466
 467         test_mazes, test_paths, _ = maze.create_maze_data(
 468             nb_test_samples,
 469             height=height,
 470             width=width,
 471             nb_walls=nb_walls,
 472             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 473         )
 474         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 475
 476         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 477
 478     def batches(self, split="train", nb_to_use=-1, desc=None):
 479         assert split in {"train", "test"}
 480         input = self.train_input if split == "train" else self.test_input
 481         if nb_to_use > 0:
 482             input = input[:nb_to_use]
 483         if desc is None:
 484             desc = f"epoch-{split}"
 485         for batch in tqdm.tqdm(
 486             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 487         ):
 488             yield batch
 489
 490     def vocabulary_size(self):
 491         return self.nb_codes
 492
 493     def compute_error(
 494         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 495     ):
 496         nb_total, nb_correct = 0, 0
 497         count = torch.zeros(
 498             self.width * self.height,
 499             self.width * self.height,
 500             device=self.device,
 501             dtype=torch.int64,
 502         )
 503
 504         for input in self.batches(split, nb_to_use):
 505             result = input.clone()
 506             ar_mask = result.new_zeros(result.size())
 507             ar_mask[:, self.height * self.width :] = 1
 508             result *= 1 - ar_mask
 509             masked_inplace_autoregression(
 510                 model,
 511                 self.batch_size,
 512                 result,
 513                 ar_mask,
 514                 deterministic_synthesis,
 515                 progress_bar_desc=None,
 516                 device=self.device,
 517             )
 518             mazes, paths = self.seq2map(result)
 519             path_correctness = maze.path_correctness(mazes, paths)
 520             nb_correct += path_correctness.long().sum()
 521             nb_total += mazes.size(0)
 522
 523             optimal_path_lengths = (
 524                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 525             )
 526             predicted_path_lengths = (
 527                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 528             )
 529             optimal_path_lengths = optimal_path_lengths[path_correctness]
 530             predicted_path_lengths = predicted_path_lengths[path_correctness]
 531             count[optimal_path_lengths, predicted_path_lengths] += 1
 532
 533         if count.max() == 0:
 534             count = None
 535         else:
 536             count = count[
 537                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 538             ]
 539
 540         return nb_total, nb_correct, count
 541
 542     def produce_results(
 543         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 544     ):
 545         train_nb_total, train_nb_correct, count = self.compute_error(
 546             model,
 547             "train",
 548             nb_to_use=1000,
 549             deterministic_synthesis=deterministic_synthesis,
 550         )
 551         logger(
 552             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}%"
 553         )
 554
 555         test_nb_total, test_nb_correct, count = self.compute_error(
 556             model,
 557             "test",
 558             nb_to_use=1000,
 559             deterministic_synthesis=deterministic_synthesis,
 560         )
 561         logger(
 562             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}%"
 563         )
 564
 565         if count is not None:
 566             proportion_optimal = count.diagonal().sum().float() / count.sum()
 567             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 568             with open(
 569                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 570             ) as f:
 571                 for i in range(count.size(0)):
 572                     for j in range(count.size(1)):
 573                         eol = " " if j < count.size(1) - 1 else "\n"
 574                         f.write(f"{count[i,j]}{eol}")
 575
 576         input = self.test_input[:48]
 577         result = input.clone()
 578         ar_mask = result.new_zeros(result.size())
 579         ar_mask[:, self.height * self.width :] = 1
 580         result *= 1 - ar_mask
 581         masked_inplace_autoregression(
 582             model,
 583             self.batch_size,
 584             result,
 585             ar_mask,
 586             deterministic_synthesis,
 587             device=self.device,
 588         )
 589
 590         mazes, paths = self.seq2map(input)
 591         _, predicted_paths = self.seq2map(result)
 592
 593         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 594         maze.save_image(
 595             filename,
 596             mazes=mazes,
 597             target_paths=paths,
 598             predicted_paths=predicted_paths,
 599             path_correct=maze.path_correctness(mazes, predicted_paths),
 600             path_optimal=maze.path_optimality(paths, predicted_paths),
 601         )
 602         logger(f"wrote {filename}")
 603
 604
 605 ######################################################################
 606
 607
 608 import snake
 609
 610
 611 class Snake(Task):
 612     def __init__(
 613         self,
 614         nb_train_samples,
 615         nb_test_samples,
 616         batch_size,
 617         height,
 618         width,
 619         nb_colors,
 620         length,
 621         prompt_length,
 622         device=torch.device("cpu"),
 623     ):
 624         super().__init__()
 625
 626         self.batch_size = batch_size
 627         self.height = height
 628         self.width = width
 629         self.device = device
 630         self.prompt_length = prompt_length
 631
 632         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 633             nb_train_samples,
 634             height,
 635             width,
 636             nb_colors,
 637             length,
 638             prompt_length,
 639             self.device,
 640         )
 641         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 642             nb_test_samples,
 643             height,
 644             width,
 645             nb_colors,
 646             length,
 647             prompt_length,
 648             self.device,
 649         )
 650
 651         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 652
 653     def batches(self, split="train", nb_to_use=-1, desc=None):
 654         assert split in {"train", "test"}
 655         input = self.train_input if split == "train" else self.test_input
 656         if nb_to_use > 0:
 657             input = input[:nb_to_use]
 658         if desc is None:
 659             desc = f"epoch-{split}"
 660         for batch in tqdm.tqdm(
 661             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 662         ):
 663             yield batch
 664
 665     def vocabulary_size(self):
 666         return self.nb_codes
 667
 668     def produce_results(
 669         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 670     ):
 671         def compute_nb_correct(input, prior_visits):
 672             result = input.clone()
 673             i = torch.arange(result.size(1), device=result.device)[None, :]
 674             ar_mask = (
 675                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 676                 .long()
 677                 .expand_as(result)
 678             )
 679             result *= 1 - ar_mask
 680
 681             masked_inplace_autoregression(
 682                 model,
 683                 self.batch_size,
 684                 result,
 685                 ar_mask,
 686                 deterministic_synthesis,
 687                 device=self.device,
 688             )
 689
 690             nb_total = ((prior_visits > 0) * ar_mask).sum()
 691
 692             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 693
 694             return nb_total, nb_correct
 695
 696         test_nb_total, test_nb_correct = compute_nb_correct(
 697             self.test_input[:1000], self.test_prior_visits[:1000]
 698         )
 699
 700         logger(
 701             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}%"
 702         )
 703
 704
 705 ######################################################################
 706
 707
 708 import stack
 709
 710
 711 class Stack(Task):
 712     def __init__(
 713         self,
 714         nb_train_samples,
 715         nb_test_samples,
 716         batch_size,
 717         logger,
 718         nb_steps,
 719         nb_stacks,
 720         nb_digits,
 721         fraction_values_for_train=None,
 722         device=torch.device("cpu"),
 723     ):
 724         super().__init__()
 725
 726         self.batch_size = batch_size
 727         self.nb_steps = nb_steps
 728         self.nb_stacks = nb_stacks
 729         self.nb_digits = nb_digits
 730         self.device = device
 731
 732         if fraction_values_for_train is None:
 733             values_for_train = None
 734             values_for_test = None
 735         else:
 736             all = torch.randperm(10**nb_digits)
 737             nb_for_train = int(all.size(0) * fraction_values_for_train)
 738             values_for_train = all[:nb_for_train]
 739             values_for_test = all[nb_for_train:]
 740
 741         self.train_input, self.train_stack_counts = stack.generate_sequences(
 742             nb_train_samples,
 743             nb_steps,
 744             nb_stacks,
 745             nb_digits,
 746             values_for_train,
 747             self.device,
 748         )
 749
 750         self.test_input, self.test_stack_counts = stack.generate_sequences(
 751             nb_test_samples,
 752             nb_steps,
 753             nb_stacks,
 754             nb_digits,
 755             values_for_test,
 756             self.device,
 757         )
 758
 759         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 760         counts = self.test_stack_counts.flatten()[i.flatten()]
 761         counts = F.one_hot(counts).sum(0)
 762         logger(f"test_pop_stack_counts {counts}")
 763
 764         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 765
 766     def batches(self, split="train", nb_to_use=-1, desc=None):
 767         assert split in {"train", "test"}
 768         input = self.train_input if split == "train" else self.test_input
 769         if nb_to_use > 0:
 770             input = input[:nb_to_use]
 771         if desc is None:
 772             desc = f"epoch-{split}"
 773         for batch in tqdm.tqdm(
 774             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 775         ):
 776             yield batch
 777
 778     def vocabulary_size(self):
 779         return self.nb_codes
 780
 781     def produce_results(
 782         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 783     ):
 784         def compute_nb_correct(input):
 785             result = input.clone()
 786             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 787             ar_mask = (result != input).long()
 788             masked_inplace_autoregression(
 789                 model,
 790                 self.batch_size,
 791                 result,
 792                 ar_mask,
 793                 deterministic_synthesis,
 794                 device=self.device,
 795             )
 796
 797             errors = ((result != input).long() * ar_mask).reshape(
 798                 -1, 1 + self.nb_digits
 799             )
 800             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 801
 802             nb_total = ar_mask.max(1).values.sum()
 803             nb_correct = nb_total - errors.max(1).values.sum()
 804
 805             return nb_total, nb_correct
 806
 807         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 808
 809         logger(
 810             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}%"
 811         )
 812
 813         ##############################################################
 814         # Log a few generated sequences
 815         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 816         result = input.clone()
 817         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 818         ar_mask = (result != input).long()
 819
 820         # for n in range(result.size(0)):
 821         # logger(
 822         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 823         # )
 824
 825         masked_inplace_autoregression(
 826             model,
 827             self.batch_size,
 828             result,
 829             ar_mask,
 830             deterministic_synthesis,
 831             device=self.device,
 832         )
 833
 834         for n in range(result.size(0)):
 835             logger(
 836                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 837             )
 838         ##############################################################
 839
 840
 841 ######################################################################
 842
 843
 844 import expr
 845
 846
 847 class Expr(Task):
 848     def tensorize(self, sequences):
 849         len_max = max([len(x) for x in sequences])
 850         return torch.cat(
 851             [
 852                 torch.tensor(
 853                     [
 854                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
 855                         for s in sequences
 856                     ]
 857                 )
 858             ],
 859             0,
 860         ).to(self.device)
 861
 862     def __init__(
 863         self,
 864         nb_train_samples,
 865         nb_test_samples,
 866         nb_variables,
 867         sequence_length,
 868         operand_max,
 869         result_max,
 870         batch_size,
 871         device=torch.device("cpu"),
 872     ):
 873         super().__init__()
 874
 875         self.batch_size = batch_size
 876         self.device = device
 877
 878         train_sequences = expr.generate_sequences(
 879             nb_train_samples,
 880             nb_variables=nb_variables,
 881             length=sequence_length,
 882             operand_max=operand_max,
 883             result_max=result_max,
 884         )
 885
 886         test_sequences = expr.generate_sequences(
 887             nb_test_samples,
 888             nb_variables=nb_variables,
 889             length=sequence_length,
 890             operand_max=operand_max,
 891             result_max=result_max,
 892         )
 893
 894         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
 895         symbols.sort()
 896
 897         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
 898         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 899
 900         self.filler, self.space = self.char2id["#"], self.char2id[" "]
 901
 902         self.train_input = self.tensorize(train_sequences)
 903         self.test_input = self.tensorize(test_sequences)
 904
 905         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 906
 907     def batches(self, split="train", nb_to_use=-1, desc=None):
 908         assert split in {"train", "test"}
 909         input = self.train_input if split == "train" else self.test_input
 910         if nb_to_use > 0:
 911             input = input[:nb_to_use]
 912         if desc is None:
 913             desc = f"epoch-{split}"
 914         for batch in tqdm.tqdm(
 915             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 916         ):
 917             last = (batch != self.filler).max(0).values.nonzero().max() + 3
 918             batch = batch[:, :last]
 919             yield batch
 920
 921     def vocabulary_size(self):
 922         return self.nb_codes
 923
 924     def seq2str(self, s):
 925         return "".join([self.id2char[k.item()] for k in s])
 926
 927     def produce_results(
 928         self,
 929         n_epoch,
 930         model,
 931         result_dir,
 932         logger,
 933         deterministic_synthesis,
 934         input_file=None,
 935     ):
 936         def compute_nb_correct(input):
 937             result = input.clone()
 938             s = (result == self.space).long()
 939             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
 940             result = (1 - ar_mask) * result + ar_mask * self.filler
 941             masked_inplace_autoregression(
 942                 model,
 943                 self.batch_size,
 944                 result,
 945                 ar_mask,
 946                 deterministic_synthesis,
 947                 device=self.device,
 948             )
 949
 950             nb_total = input.size(0)
 951             nb_correct = (input == result).long().min(1).values.sum()
 952
 953             #######################################################################
 954             # Comput predicted vs. true variable values
 955
 956             nb_delta = torch.zeros(5, dtype=torch.int64)
 957             nb_missed = 0
 958
 959             values_input = expr.extract_results([self.seq2str(s) for s in input])
 960             values_result = expr.extract_results([self.seq2str(s) for s in result])
 961
 962             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
 963
 964             with open(filename, "w") as f:
 965                 for i, r in zip(values_input, values_result):
 966                     for n, vi in i.items():
 967                         vr = r.get(n)
 968                         f.write(f"{vi} {-1 if vr is None else vr}\n")
 969
 970                         if vr is None or vr < 0:
 971                             nb_missed += 1
 972                         else:
 973                             d = abs(vr - vi)
 974                             if d >= nb_delta.size(0):
 975                                 nb_missed += 1
 976                             else:
 977                                 nb_delta[d] += 1
 978
 979             ######################################################################
 980
 981             return nb_total, nb_correct, nb_delta, nb_missed
 982
 983         (
 984             test_nb_total,
 985             test_nb_correct,
 986             test_nb_delta,
 987             test_nb_missed,
 988         ) = compute_nb_correct(self.test_input[:10000])
 989
 990         logger(
 991             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}%"
 992         )
 993
 994         nb_total = test_nb_delta.sum() + test_nb_missed
 995         for d in range(test_nb_delta.size(0)):
 996             logger(
 997                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
 998             )
 999         logger(
1000             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1001         )
1002
1003         ##############################################################
1004         # Log a few generated sequences
1005         if input_file is None:
1006             input = self.test_input[:10]
1007         else:
1008             with open(input_file, "r") as f:
1009                 sequences = [e.strip() for e in f.readlines()]
1010                 sequences = [s + " " + "#" * 50 for s in sequences]
1011                 input = self.tensorize(sequences)
1012
1013         result = input.clone()
1014         s = (result == self.space).long()
1015         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1016         result = (1 - ar_mask) * result + ar_mask * self.filler
1017
1018         for n in range(result.size(0)):
1019             logger(f"test_before {self.seq2str(result[n])}")
1020
1021         masked_inplace_autoregression(
1022             model,
1023             self.batch_size,
1024             result,
1025             ar_mask,
1026             deterministic_synthesis,
1027             device=self.device,
1028         )
1029
1030         correct = (1 - ar_mask) * self.space + ar_mask * input
1031         for n in range(result.size(0)):
1032             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1033             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1034             logger(f"truth       {self.seq2str(correct[n])}")
1035         ##############################################################
1036
1037
1038 ######################################################################
1039
1040 import world
1041
1042
1043 class World(Task):
1044     def __init__(
1045         self,
1046         nb_train_samples,
1047         nb_test_samples,
1048         batch_size,
1049         vqae_nb_epochs,
1050         logger=None,
1051         device=torch.device("cpu"),
1052         device_storage=torch.device("cpu"),
1053     ):
1054         super().__init__()
1055
1056         self.batch_size = batch_size
1057         self.device = device
1058
1059         (
1060             train_frames,
1061             train_action_seq,
1062             test_frames,
1063             test_action_seq,
1064             self.frame2seq,
1065             self.seq2frame,
1066         ) = world.create_data_and_processors(
1067             nb_train_samples,
1068             nb_test_samples,
1069             mode="first_last",
1070             nb_steps=30,
1071             nb_epochs=vqae_nb_epochs,
1072             logger=logger,
1073             device=device,
1074             device_storage=device_storage,
1075         )
1076
1077         print(f"{train_action_seq.size()=}")
1078
1079         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
1080         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
1081
1082         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
1083         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
1084
1085         self.len_frame_seq = train_frame_seq.size(1)
1086         self.len_action_seq = train_action_seq.size(1)
1087         self.nb_codes = nb_frame_codes + nb_action_codes
1088
1089         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
1090         print(f"{train_action_seq.device=} {nb_frame_codes.device=}")
1091         train_action_seq += nb_frame_codes
1092         self.train_input = torch.cat(
1093             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
1094         )
1095
1096         test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
1097         test_action_seq += nb_frame_codes
1098         self.test_input = torch.cat(
1099             (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
1100         )
1101
1102     def batches(self, split="train", nb_to_use=-1, desc=None):
1103         assert split in {"train", "test"}
1104         input = self.train_input if split == "train" else self.test_input
1105         if nb_to_use > 0:
1106             input = input[:nb_to_use]
1107         if desc is None:
1108             desc = f"epoch-{split}"
1109         for batch in tqdm.tqdm(
1110             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1111         ):
1112             yield batch.to(self.device)
1113
1114     def vocabulary_size(self):
1115         return self.nb_codes
1116
1117     def produce_results(
1118         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1119     ):
1120         k = torch.arange(
1121             2 * self.len_frame_seq + self.len_action_seq, device=self.device
1122         )[None, :]
1123
1124         input = self.test_input[:64].to(self.device)
1125         result = input.clone()
1126
1127         ar_mask = (
1128             (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
1129         )
1130         result *= 1 - ar_mask
1131
1132         masked_inplace_autoregression(
1133             model,
1134             self.batch_size,
1135             result,
1136             ar_mask,
1137             deterministic_synthesis,
1138             device=self.device,
1139         )
1140
1141         seq_start = input[:, : self.len_frame_seq]
1142         seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
1143         seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
1144
1145         result = torch.cat(
1146             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
1147         )
1148         result = result.reshape(-1, result.size(-1))
1149         print(f"{result.size()=}")
1150
1151         frames = self.seq2frame(result)
1152         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1153         torchvision.utils.save_image(
1154             frames.float() / (world.Box.nb_rgb_levels - 1),
1155             image_name,
1156             nrow=12,
1157             padding=1,
1158             pad_value=0.0,
1159         )
1160         logger(f"wrote {image_name}")
1161
1162
1163 ######################################################################