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