tasks.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math, os, tqdm
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 ######################################################################
  16
  17
  18 def masked_inplace_autoregression(
  19     model,
  20     batch_size,
  21     input,
  22     ar_mask,
  23     deterministic_synthesis,
  24     forbidden_tokens=None,
  25     progress_bar_desc="autoregression",
  26     device=torch.device("cpu"),
  27 ):
  28     assert input.size() == ar_mask.size()
  29
  30     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
  31
  32     if progress_bar_desc is not None:
  33         batches = tqdm.tqdm(
  34             batches,
  35             dynamic_ncols=True,
  36             desc=progress_bar_desc,
  37             total=(input.size(0) + batch_size - 1) // batch_size,
  38         )
  39
  40     with torch.autograd.no_grad():
  41         t = model.training
  42         model.eval()
  43
  44         for input, ar_mask in batches:
  45             model.masked_inplace_autoregression(
  46                 input, ar_mask, forbidden_tokens, deterministic_synthesis
  47             )
  48
  49         model.train(t)
  50
  51
  52 ######################################################################
  53
  54
  55 class Task:
  56     def batches(self, split="train"):
  57         pass
  58
  59     def vocabulary_size(self):
  60         pass
  61
  62     def produce_results(
  63         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  64     ):
  65         pass
  66
  67
  68 ######################################################################
  69
  70
  71 class Problem:
  72     def generate_sequences(self, nb):
  73         pass
  74
  75     def seq2str(self, seq):
  76         return "[NOT IMPLEMENTED]"
  77
  78
  79 ####################
  80
  81
  82 class ProblemLevel0(Problem):
  83     def __init__(self, nb_sentences=100, len_prompt=5, len_result=5):
  84         self.seq = torch.randint(10, (nb_sentences, len_prompt + 1 + len_result))
  85         self.seq[:, len_prompt] = 10
  86
  87     def generate_sequences(self, nb):
  88         sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
  89         ar_mask = (sequences == 10).long()
  90         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
  91         return sequences, ar_mask
  92
  93
  94 class ProblemLevel1(Problem):
  95     def __init__(self, nb_operators=100, len_source=5, len_result=8):
  96         self.len_source = len_source
  97         self.len_result = len_result
  98         self.len_nb_operator = int(math.log(nb_operators) / math.log(10)) + 1
  99         self.operators = F.one_hot(
 100             torch.rand(nb_operators, len_result, len_source).argmax(-1),
 101             num_classes=len_source,
 102         )
 103
 104     def generate_sequences(self, nb):
 105         nb_operators = torch.randint(self.operators.size(0), (nb,))
 106         operators = self.operators[nb_operators]
 107         nb_operators = (
 108             nb_operators[:, None]
 109             // 10 ** torch.arange(self.len_nb_operator - 1, -1, -1)
 110         ) % 10
 111         marker1 = torch.full((nb, 1), 10)
 112         # source = torch.randint(10, (nb, self.len_source))
 113         source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
 114         marker2 = torch.full((nb, 1), 11)
 115         result = operators.bmm(source[:, :, None]).squeeze(-1)
 116         sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
 117         ar_mask = (sequences == 11).long()
 118         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
 119         return sequences, ar_mask
 120
 121     def seq2str(self, seq):
 122         return "".join("0123456789|>"[x.item()] for x in seq)
 123
 124
 125 class ProblemLevel2(Problem):
 126     def __init__(self, len_source=5, len_result=8):
 127         self.len_source = len_source
 128         self.len_result = len_result
 129
 130     def generate_sequences(self, nb):
 131         operators = F.one_hot(
 132             torch.rand(nb, self.len_result, self.len_source).argmax(-1),
 133             num_classes=self.len_source,
 134         )
 135         source1 = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
 136         # source1 = torch.randint(10, (nb, self.len_source))
 137         marker1 = torch.full((nb, 1), 10)
 138         result1 = operators.bmm(source1[:, :, None]).squeeze(-1)
 139         marker2 = torch.full((nb, 1), 11)
 140         source2 = torch.randint(10, (nb, self.len_source))
 141         marker3 = torch.full((nb, 1), 12)
 142         result2 = operators.bmm(source2[:, :, None]).squeeze(-1)
 143
 144         sequences = torch.cat(
 145             (source1, marker1, result1, marker2, source2, marker3, result2), 1
 146         )
 147         ar_mask = (sequences == 12).long()
 148         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
 149         return sequences, ar_mask
 150
 151     def seq2str(self, seq):
 152         return "".join("0123456789>|~"[x.item()] for x in seq)
 153
 154
 155 ####################
 156
 157
 158 class ProblemAddition(Problem):
 159     def __init__(self, nb_digits=10, zero_padded=False, inverted_result=False):
 160         self.nb_digits = nb_digits
 161         self.zero_padded = zero_padded
 162         self.inverted_result = inverted_result
 163         self.char2id = dict([(c, n) for n, c in enumerate("0123456789+=$")])
 164         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 165
 166     def tensorize(self, strings):
 167         len_max = max([len(x) for x in strings])
 168         return torch.cat(
 169             [
 170                 torch.tensor(
 171                     [
 172                         [self.char2id[c] for c in s + "$" * (len_max - len(s))]
 173                         for s in strings
 174                     ]
 175                 )
 176             ],
 177             0,
 178         )
 179
 180     def generate_sequences(self, nb):
 181         sequences = []
 182         for k in range(nb):
 183             a, b = torch.randint(10**self.nb_digits, (2,))
 184             c = a + b
 185             a, b, c = str(a.item()), str(b.item()), str(c.item())
 186             if self.zero_padded:
 187                 a = "0" * (self.nb_digits - len(a)) + a
 188                 b = "0" * (self.nb_digits - len(b)) + b
 189                 c = "0" * (self.nb_digits + 1 - len(c)) + c
 190             if self.inverted_result:
 191                 c = c[::-1]
 192             sequences.append(f"{a}+{b}={c}$")
 193
 194         sequences = self.tensorize(sequences)
 195         ar_mask = (sequences == self.char2id["="]).long()
 196         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
 197         return sequences, ar_mask
 198
 199     def seq2str(self, seq):
 200         return "".join(self.id2char[x.item()] for x in seq)
 201
 202
 203 # class ProblemUnion(Problem):
 204 # problems = [ProblemByheart()]
 205 # nb_common_codes = 100
 206
 207 # def generate_sequences(nb_samples):
 208 # problem_indexes = torch.randint(len(problems), (nb_samples,))
 209 # nb_samples_per_problem = torch.one_hot(problem_indexes).sum(0)
 210 # print(f"{nb_samples_per_problem}")
 211 # all_seq = []
 212 # for nb, p in zip(nb_samples_per_problem, problems):
 213 # all_seq.append(p.generate_sequences(nb_samples_per_problem[nb]))
 214 # return all_seq
 215
 216 # for strain, stest in zip(train_seq, test_seq):
 217 # s = torch.cat((strain, stest), 0)
 218
 219 ####################
 220
 221
 222 class SandBox(Task):
 223     def __init__(
 224         self,
 225         problem,
 226         nb_train_samples,
 227         nb_test_samples,
 228         batch_size,
 229         logger=None,
 230         device=torch.device("cpu"),
 231         max_nb_codes=1024,
 232     ):
 233         super().__init__()
 234
 235         self.batch_size = batch_size
 236         self.device = device
 237         self.problem = problem
 238
 239         self.train_input, self.train_ar_mask = self.problem.generate_sequences(
 240             nb_train_samples
 241         )
 242         self.test_input, self.test_ar_mask = self.problem.generate_sequences(
 243             nb_test_samples
 244         )
 245
 246         self.train_input, self.train_ar_mask = self.train_input.to(
 247             device
 248         ), self.train_ar_mask.to(device)
 249         self.test_input, self.test_ar_mask = self.test_input.to(
 250             device
 251         ), self.test_ar_mask.to(device)
 252
 253         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 254
 255         # A bit of paranoia never hurts
 256         assert (
 257             self.nb_codes <= max_nb_codes
 258             and self.train_input.min() >= 0
 259             and self.test_input.min() >= 0
 260             and tuple(self.train_ar_mask.unique()) == (0, 1)
 261             and tuple(self.test_ar_mask.unique()) == (0, 1)
 262         )
 263
 264     def batches(self, split="train", nb_to_use=-1, desc=None):
 265         assert split in {"train", "test"}
 266         input = self.train_input if split == "train" else self.test_input
 267         if nb_to_use > 0:
 268             input = input[:nb_to_use]
 269         if desc is None:
 270             desc = f"epoch-{split}"
 271         for batch in tqdm.tqdm(
 272             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 273         ):
 274             yield batch
 275
 276     def vocabulary_size(self):
 277         return self.nb_codes
 278
 279     def produce_results(
 280         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 281     ):
 282         def compute_accuracy(input, ar_mask, logger=None):
 283             input, ar_mask = input[:nmax], ar_mask[:nmax]
 284             result = input.clone() * (1 - ar_mask)
 285
 286             masked_inplace_autoregression(
 287                 model,
 288                 self.batch_size,
 289                 result,
 290                 ar_mask,
 291                 deterministic_synthesis,
 292                 progress_bar_desc=None,
 293                 device=self.device,
 294             )
 295
 296             if logger is not None:
 297                 for sp, st in zip(result[:10], input[:10]):
 298                     logger(
 299                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 300                     )
 301                     logger(
 302                         f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 303                     )
 304
 305             nb_total = ar_mask.sum().item()
 306             nb_correct = ((result == input).long() * ar_mask).sum().item()
 307
 308             return nb_total, nb_correct
 309
 310         train_nb_total, train_nb_correct = compute_accuracy(
 311             self.train_input, self.train_ar_mask
 312         )
 313
 314         logger(
 315             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}%"
 316         )
 317
 318         test_nb_total, test_nb_correct = compute_accuracy(
 319             self.test_input, self.test_ar_mask, logger
 320         )
 321
 322         logger(
 323             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}%"
 324         )
 325
 326
 327 ######################################################################
 328
 329 import picoclvr
 330
 331
 332 class PicoCLVR(Task):
 333     # Make a tensor from a list of strings
 334     def tensorize(self, descr):
 335         token_descr = [s.strip().split(" ") for s in descr]
 336         l = max([len(s) for s in token_descr])
 337         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 338         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 339         return torch.tensor(id_descr, device=self.device)
 340
 341     # Make a list of strings from a tensor
 342     def detensorize(self, x):
 343         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 344
 345     # trim all the tensors in the tuple z to remove as much token from
 346     # left and right in the first tensor. If z is a tuple, all its
 347     # elements are trimed according to the triming for the first
 348     def trim(self, z, token="<nul>"):
 349         n = self.token2id[token]
 350         if type(z) == tuple:
 351             x = z[0]
 352             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 353             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 354             return tuple([t[:, a:b] for t in z])
 355         else:
 356             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 357             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 358             return z[:, a:b]
 359
 360     ######################
 361
 362     def __init__(
 363         self,
 364         nb_train_samples,
 365         nb_test_samples,
 366         batch_size,
 367         height,
 368         width,
 369         nb_colors=5,
 370         logger=None,
 371         device=torch.device("cpu"),
 372         pruner_train=None,
 373         pruner_eval=None,
 374     ):
 375         super().__init__()
 376
 377         def generate_descr(nb, cache_suffix, pruner):
 378             return picoclvr.generate(
 379                 nb,
 380                 height=self.height,
 381                 width=self.width,
 382                 nb_colors=nb_colors,
 383                 pruner=pruner,
 384             )
 385
 386         self.height = height
 387         self.width = width
 388         self.batch_size = batch_size
 389         self.device = device
 390         self.pruner_train = pruner_train
 391         self.pruner_eval = pruner_eval
 392
 393         if logger is not None:
 394             logger(
 395                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 396             )
 397
 398         self.train_descr = generate_descr(
 399             nb_train_samples, "train", pruner=self.pruner_train
 400         )
 401         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 402
 403         # Build the tokenizer
 404         tokens = {"<nul>", "<img>"}
 405         for d in [self.train_descr, self.test_descr]:
 406             for s in d:
 407                 for t in s.strip().split(" "):
 408                     tokens.add(t)
 409         # make this set a sorted list to get the same tensors given
 410         # the same descr
 411         tokens = list(tokens)
 412         tokens.sort()
 413         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 414         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 415         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 416
 417         # Tokenize the train and test sets
 418         self.train_input = self.tensorize(self.train_descr)
 419         self.test_input = self.tensorize(self.test_descr)
 420
 421     def batches(self, split="train"):
 422         assert split in {"train", "test"}
 423         input = self.train_input if split == "train" else self.test_input
 424         for batch in tqdm.tqdm(
 425             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 426         ):
 427             yield self.trim(batch)
 428
 429     def vocabulary_size(self):
 430         return len(self.token2id)
 431
 432     def compute_missing_properties(
 433         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 434     ):
 435         acc_nb_requested_properties = []
 436         acc_nb_missing_properties = []
 437         acc_nb_results = 0
 438
 439         for input in tqdm.tqdm(
 440             self.test_input.split(self.batch_size),
 441             dynamic_ncols=True,
 442             desc=f"test-properties",
 443         ):
 444             result = input.clone()
 445             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 446             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 447             masked_inplace_autoregression(
 448                 model,
 449                 self.batch_size,
 450                 result,
 451                 ar_mask,
 452                 deterministic_synthesis,
 453                 progress_bar_desc=None,
 454                 device=self.device,
 455             )
 456
 457             result_descr = self.detensorize(result)
 458             np = picoclvr.nb_properties(
 459                 result_descr,
 460                 height=self.height,
 461                 width=self.width,
 462                 pruner=pruner,
 463             )
 464             nb_requested_properties, _, nb_missing_properties = zip(*np)
 465             acc_nb_requested_properties += nb_requested_properties
 466             acc_nb_missing_properties += nb_missing_properties
 467             acc_nb_results += len(result_descr)
 468
 469         nb_requested_properties = sum(acc_nb_requested_properties)
 470         nb_missing_properties = sum(acc_nb_missing_properties)
 471
 472         prefix = "" if pruner is None else "pruned_"
 473         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 474         logger(
 475             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 476         )
 477         logger(
 478             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 479         )
 480
 481     ######################################################################
 482
 483     def produce_results(
 484         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 485     ):
 486         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 487
 488         if self.pruner_eval is not None:
 489             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 490
 491         nb_tokens_to_generate = self.height * self.width + 3
 492         result_descr = []
 493         nb_per_primer = 8
 494         primer = []
 495
 496         for primer_descr in [
 497             "red above green <sep> green top <sep> blue right of red",
 498             "there is red <sep> there is yellow <sep> there is blue",
 499             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 500             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 501         ]:
 502             primer += [primer_descr + " <img>"] * nb_per_primer
 503
 504         result = self.tensorize(primer)
 505         fill = result.new_full(
 506             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 507         )
 508         result = torch.cat((result, fill), 1)
 509         ar_mask = (result == self.t_nul).long()
 510         masked_inplace_autoregression(
 511             model,
 512             self.batch_size,
 513             result,
 514             ar_mask,
 515             deterministic_synthesis,
 516             device=self.device,
 517         )
 518         result_descr = self.detensorize(result)
 519
 520         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 521
 522         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 523         acc_nb_results = len(result_descr)
 524
 525         nb_requested_properties = sum(acc_nb_requested_properties)
 526         nb_missing_properties = sum(acc_nb_missing_properties)
 527
 528         prefix = "demo_"
 529         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 530         logger(
 531             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 532         )
 533         logger(
 534             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 535         )
 536
 537         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 538
 539         if img.dim() == 5:
 540             if img.size(1) == 1:
 541                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 542             else:
 543                 img = torch.cat(
 544                     [
 545                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 546                         for x in img
 547                     ],
 548                     0,
 549                 )
 550
 551         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 552         torchvision.utils.save_image(
 553             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 554         )
 555         logger(f"wrote {image_name}")
 556
 557
 558 ######################################################################
 559
 560
 561 class MNIST(Task):
 562     def __init__(
 563         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 564     ):
 565         super().__init__()
 566
 567         self.nb_train_samples = (nb_train_samples,)
 568         self.nb_test_samples = (nb_test_samples,)
 569         self.batch_size = batch_size
 570         self.device = device
 571         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 572         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 573         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 574         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 575
 576     def batches(self, split="train", nb_to_use=-1, desc=None):
 577         assert split in {"train", "test"}
 578         input = self.train_input if split == "train" else self.test_input
 579         if nb_to_use > 0:
 580             input = input[:nb_to_use]
 581         if desc is None:
 582             desc = f"epoch-{split}"
 583         for batch in tqdm.tqdm(
 584             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 585         ):
 586             yield batch
 587
 588     def vocabulary_size(self):
 589         return 256
 590
 591     def produce_results(
 592         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 593     ):
 594         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 595         ar_mask = torch.full_like(results, 1)
 596         masked_inplace_autoregression(
 597             model,
 598             self.batch_size,
 599             results,
 600             ar_mask,
 601             deterministic_synthesis,
 602             device=self.device,
 603         )
 604         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 605         torchvision.utils.save_image(
 606             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 607             image_name,
 608             nrow=16,
 609             pad_value=0.8,
 610         )
 611         logger(f"wrote {image_name}")
 612
 613
 614 ######################################################################
 615
 616 import maze
 617
 618
 619 class Maze(Task):
 620     def map2seq(self, *m):
 621         return torch.cat([x.flatten(1) for x in m], 1)
 622
 623     def seq2map(self, s):
 624         s = s.reshape(s.size(0), -1, self.height, self.width)
 625         return (s[:, k] for k in range(s.size(1)))
 626
 627     def __init__(
 628         self,
 629         nb_train_samples,
 630         nb_test_samples,
 631         batch_size,
 632         height,
 633         width,
 634         nb_walls,
 635         device=torch.device("cpu"),
 636     ):
 637         super().__init__()
 638
 639         self.batch_size = batch_size
 640         self.height = height
 641         self.width = width
 642         self.device = device
 643
 644         train_mazes, train_paths, _ = maze.create_maze_data(
 645             nb_train_samples,
 646             height=height,
 647             width=width,
 648             nb_walls=nb_walls,
 649             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 650         )
 651         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 652
 653         test_mazes, test_paths, _ = maze.create_maze_data(
 654             nb_test_samples,
 655             height=height,
 656             width=width,
 657             nb_walls=nb_walls,
 658             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 659         )
 660         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 661
 662         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 663
 664     def batches(self, split="train", nb_to_use=-1, desc=None):
 665         assert split in {"train", "test"}
 666         input = self.train_input if split == "train" else self.test_input
 667         if nb_to_use > 0:
 668             input = input[:nb_to_use]
 669         if desc is None:
 670             desc = f"epoch-{split}"
 671         for batch in tqdm.tqdm(
 672             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 673         ):
 674             yield batch
 675
 676     def vocabulary_size(self):
 677         return self.nb_codes
 678
 679     def compute_error(
 680         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 681     ):
 682         nb_total, nb_correct = 0, 0
 683         count = torch.zeros(
 684             self.width * self.height,
 685             self.width * self.height,
 686             device=self.device,
 687             dtype=torch.int64,
 688         )
 689
 690         for input in self.batches(split, nb_to_use):
 691             result = input.clone()
 692             ar_mask = result.new_zeros(result.size())
 693             ar_mask[:, self.height * self.width :] = 1
 694             result *= 1 - ar_mask
 695             masked_inplace_autoregression(
 696                 model,
 697                 self.batch_size,
 698                 result,
 699                 ar_mask,
 700                 deterministic_synthesis,
 701                 progress_bar_desc=None,
 702                 device=self.device,
 703             )
 704             mazes, paths = self.seq2map(result)
 705             path_correctness = maze.path_correctness(mazes, paths)
 706             nb_correct += path_correctness.long().sum()
 707             nb_total += mazes.size(0)
 708
 709             optimal_path_lengths = (
 710                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 711             )
 712             predicted_path_lengths = (
 713                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 714             )
 715             optimal_path_lengths = optimal_path_lengths[path_correctness]
 716             predicted_path_lengths = predicted_path_lengths[path_correctness]
 717             count[optimal_path_lengths, predicted_path_lengths] += 1
 718
 719         if count.max() == 0:
 720             count = None
 721         else:
 722             count = count[
 723                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 724             ]
 725
 726         return nb_total, nb_correct, count
 727
 728     def produce_results(
 729         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 730     ):
 731         train_nb_total, train_nb_correct, count = self.compute_error(
 732             model,
 733             "train",
 734             nb_to_use=1000,
 735             deterministic_synthesis=deterministic_synthesis,
 736         )
 737         logger(
 738             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}%"
 739         )
 740
 741         test_nb_total, test_nb_correct, count = self.compute_error(
 742             model,
 743             "test",
 744             nb_to_use=1000,
 745             deterministic_synthesis=deterministic_synthesis,
 746         )
 747         logger(
 748             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}%"
 749         )
 750
 751         if count is not None:
 752             proportion_optimal = count.diagonal().sum().float() / count.sum()
 753             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 754             with open(
 755                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 756             ) as f:
 757                 for i in range(count.size(0)):
 758                     for j in range(count.size(1)):
 759                         eol = " " if j < count.size(1) - 1 else "\n"
 760                         f.write(f"{count[i,j]}{eol}")
 761
 762         input = self.test_input[:48]
 763         result = input.clone()
 764         ar_mask = result.new_zeros(result.size())
 765         ar_mask[:, self.height * self.width :] = 1
 766         result *= 1 - ar_mask
 767         masked_inplace_autoregression(
 768             model,
 769             self.batch_size,
 770             result,
 771             ar_mask,
 772             deterministic_synthesis,
 773             device=self.device,
 774         )
 775
 776         mazes, paths = self.seq2map(input)
 777         _, predicted_paths = self.seq2map(result)
 778
 779         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 780         maze.save_image(
 781             filename,
 782             mazes=mazes,
 783             target_paths=paths,
 784             predicted_paths=predicted_paths,
 785             path_correct=maze.path_correctness(mazes, predicted_paths),
 786             path_optimal=maze.path_optimality(paths, predicted_paths),
 787         )
 788         logger(f"wrote {filename}")
 789
 790
 791 ######################################################################
 792
 793
 794 import snake
 795
 796
 797 class Snake(Task):
 798     def __init__(
 799         self,
 800         nb_train_samples,
 801         nb_test_samples,
 802         batch_size,
 803         height,
 804         width,
 805         nb_colors,
 806         length,
 807         prompt_length,
 808         device=torch.device("cpu"),
 809     ):
 810         super().__init__()
 811
 812         self.batch_size = batch_size
 813         self.height = height
 814         self.width = width
 815         self.device = device
 816         self.prompt_length = prompt_length
 817
 818         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 819             nb_train_samples,
 820             height,
 821             width,
 822             nb_colors,
 823             length,
 824             prompt_length,
 825             self.device,
 826         )
 827         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 828             nb_test_samples,
 829             height,
 830             width,
 831             nb_colors,
 832             length,
 833             prompt_length,
 834             self.device,
 835         )
 836
 837         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 838
 839     def batches(self, split="train", nb_to_use=-1, desc=None):
 840         assert split in {"train", "test"}
 841         input = self.train_input if split == "train" else self.test_input
 842         if nb_to_use > 0:
 843             input = input[:nb_to_use]
 844         if desc is None:
 845             desc = f"epoch-{split}"
 846         for batch in tqdm.tqdm(
 847             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 848         ):
 849             yield batch
 850
 851     def vocabulary_size(self):
 852         return self.nb_codes
 853
 854     def produce_results(
 855         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 856     ):
 857         def compute_nb_correct(input, prior_visits):
 858             result = input.clone()
 859             i = torch.arange(result.size(1), device=result.device)[None, :]
 860             ar_mask = (
 861                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 862                 .long()
 863                 .expand_as(result)
 864             )
 865             result *= 1 - ar_mask
 866
 867             masked_inplace_autoregression(
 868                 model,
 869                 self.batch_size,
 870                 result,
 871                 ar_mask,
 872                 deterministic_synthesis,
 873                 device=self.device,
 874             )
 875
 876             nb_total = ((prior_visits > 0) * ar_mask).sum()
 877
 878             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 879
 880             return nb_total, nb_correct
 881
 882         test_nb_total, test_nb_correct = compute_nb_correct(
 883             self.test_input[:1000], self.test_prior_visits[:1000]
 884         )
 885
 886         logger(
 887             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}%"
 888         )
 889
 890
 891 ######################################################################
 892
 893
 894 import stack
 895
 896
 897 class Stack(Task):
 898     def __init__(
 899         self,
 900         nb_train_samples,
 901         nb_test_samples,
 902         batch_size,
 903         logger,
 904         nb_steps,
 905         nb_stacks,
 906         nb_digits,
 907         fraction_values_for_train=None,
 908         device=torch.device("cpu"),
 909     ):
 910         super().__init__()
 911
 912         self.batch_size = batch_size
 913         self.nb_steps = nb_steps
 914         self.nb_stacks = nb_stacks
 915         self.nb_digits = nb_digits
 916         self.device = device
 917
 918         if fraction_values_for_train is None:
 919             values_for_train = None
 920             values_for_test = None
 921         else:
 922             all = torch.randperm(10**nb_digits)
 923             nb_for_train = int(all.size(0) * fraction_values_for_train)
 924             values_for_train = all[:nb_for_train]
 925             values_for_test = all[nb_for_train:]
 926
 927         self.train_input, self.train_stack_counts = stack.generate_sequences(
 928             nb_train_samples,
 929             nb_steps,
 930             nb_stacks,
 931             nb_digits,
 932             values_for_train,
 933             self.device,
 934         )
 935
 936         self.test_input, self.test_stack_counts = stack.generate_sequences(
 937             nb_test_samples,
 938             nb_steps,
 939             nb_stacks,
 940             nb_digits,
 941             values_for_test,
 942             self.device,
 943         )
 944
 945         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 946         counts = self.test_stack_counts.flatten()[i.flatten()]
 947         counts = F.one_hot(counts).sum(0)
 948         logger(f"test_pop_stack_counts {counts}")
 949
 950         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 951
 952     def batches(self, split="train", nb_to_use=-1, desc=None):
 953         assert split in {"train", "test"}
 954         input = self.train_input if split == "train" else self.test_input
 955         if nb_to_use > 0:
 956             input = input[:nb_to_use]
 957         if desc is None:
 958             desc = f"epoch-{split}"
 959         for batch in tqdm.tqdm(
 960             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 961         ):
 962             yield batch
 963
 964     def vocabulary_size(self):
 965         return self.nb_codes
 966
 967     def produce_results(
 968         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 969     ):
 970         def compute_nb_correct(input):
 971             result = input.clone()
 972             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 973             ar_mask = (result != input).long()
 974             masked_inplace_autoregression(
 975                 model,
 976                 self.batch_size,
 977                 result,
 978                 ar_mask,
 979                 deterministic_synthesis,
 980                 device=self.device,
 981             )
 982
 983             errors = ((result != input).long() * ar_mask).reshape(
 984                 -1, 1 + self.nb_digits
 985             )
 986             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 987
 988             nb_total = ar_mask.max(1).values.sum()
 989             nb_correct = nb_total - errors.max(1).values.sum()
 990
 991             return nb_total, nb_correct
 992
 993         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 994
 995         logger(
 996             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}%"
 997         )
 998
 999         ##############################################################
1000         # Log a few generated sequences
1001         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
1002         result = input.clone()
1003         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1004         ar_mask = (result != input).long()
1005
1006         # for n in range(result.size(0)):
1007         # logger(
1008         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1009         # )
1010
1011         masked_inplace_autoregression(
1012             model,
1013             self.batch_size,
1014             result,
1015             ar_mask,
1016             deterministic_synthesis,
1017             device=self.device,
1018         )
1019
1020         for n in range(result.size(0)):
1021             logger(
1022                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1023             )
1024         ##############################################################
1025
1026
1027 ######################################################################
1028
1029 import rpl
1030
1031
1032 class RPL(Task):
1033     def tensorize(self, sequences):
1034         len_max = max([len(x) for x in sequences])
1035         return torch.cat(
1036             [
1037                 torch.tensor(
1038                     [
1039                         [
1040                             self.token2id[str(c)]
1041                             for c in s + ["<nul>"] * (len_max - len(s))
1042                         ]
1043                         for s in sequences
1044                     ]
1045                 )
1046             ],
1047             0,
1048         )
1049
1050     def seq2str(self, seq):
1051         return " ".join([self.id2token[i] for i in seq])
1052
1053     def __init__(
1054         self,
1055         nb_train_samples,
1056         nb_test_samples,
1057         batch_size,
1058         nb_starting_values=3,
1059         max_input=9,
1060         prog_len=6,
1061         nb_runs=5,
1062         no_prog=False,
1063         logger=None,
1064         device=torch.device("cpu"),
1065     ):
1066         super().__init__()
1067
1068         self.batch_size = batch_size
1069         self.device = device
1070         self.no_prog = no_prog
1071
1072         train_sequences = [
1073             rpl.generate(
1074                 nb_starting_values=nb_starting_values,
1075                 nb_result_values_max=4 * nb_starting_values,
1076                 max_input=max_input,
1077                 prog_len=prog_len,
1078                 nb_runs=nb_runs,
1079             )
1080             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
1081         ]
1082
1083         test_sequences = [
1084             rpl.generate(
1085                 nb_starting_values=nb_starting_values,
1086                 nb_result_values_max=4 * nb_starting_values,
1087                 max_input=max_input,
1088                 prog_len=prog_len,
1089                 nb_runs=nb_runs,
1090             )
1091             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
1092         ]
1093
1094         symbols = list(
1095             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
1096         )
1097         val_max = max([x if type(x) is int else 0 for x in symbols])
1098         symbols = list(filter(lambda x: type(x) is str, symbols))
1099         symbols.sort()
1100         symbols += [str(n) for n in range(val_max + 1)]
1101         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
1102         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
1103
1104         self.t_nul = self.token2id["<nul>"]
1105         self.t_input = self.token2id["<input>"]
1106         self.t_output = self.token2id["<output>"]
1107         self.t_prog = self.token2id["<prog>"]
1108         self.t_end = self.token2id["<end>"]
1109
1110         self.train_input = self.tensorize(train_sequences)
1111         self.test_input = self.tensorize(test_sequences)
1112
1113         if no_prog:
1114             # Excise the program from every train and test example
1115             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
1116                 None, :
1117             ]
1118             p = (
1119                 ((self.train_input == self.t_prog).long() * k)
1120                 .max(1, keepdim=True)
1121                 .values
1122             )
1123             self.train_input = (
1124                 self.train_input * (k <= p).long()
1125                 + self.t_end * (k == p + 1).long()
1126                 + self.t_nul * (k > p + 1).long()
1127             )
1128             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
1129                 None, :
1130             ]
1131             p = (
1132                 ((self.test_input == self.t_prog).long() * k)
1133                 .max(1, keepdim=True)
1134                 .values
1135             )
1136             self.test_input = (
1137                 self.test_input * (k <= p).long()
1138                 + self.t_end * (k == p + 1).long()
1139                 + self.t_nul * (k > p + 1).long()
1140             )
1141
1142         if logger is not None:
1143             logger(f"value_max {val_max}")
1144             for x in self.train_input[:25]:
1145                 end = (x != self.t_nul).nonzero().max().item() + 1
1146                 seq = [self.id2token[i.item()] for i in x[:end]]
1147                 s = " ".join(seq)
1148                 logger(f"example_seq {s}")
1149
1150         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1151
1152     def batches(self, split="train", nb_to_use=-1, desc=None):
1153         assert split in {"train", "test"}
1154         input = self.train_input if split == "train" else self.test_input
1155         if nb_to_use > 0:
1156             input = input[:nb_to_use]
1157         if desc is None:
1158             desc = f"epoch-{split}"
1159         for batch in tqdm.tqdm(
1160             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1161         ):
1162             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1163             batch = batch[:, :last].to(self.device)
1164             yield batch
1165
1166     def vocabulary_size(self):
1167         return self.nb_codes
1168
1169     def produce_results(
1170         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1171     ):
1172         # --------------------------------------------------------------------
1173         def compute_nb_errors_prog(input, nb_to_log=0):
1174             result = input.clone()
1175             s = (result == self.t_prog).long()
1176             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1177             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1178
1179             masked_inplace_autoregression(
1180                 model,
1181                 self.batch_size,
1182                 result,
1183                 ar_mask,
1184                 deterministic_synthesis,
1185                 device=self.device,
1186             )
1187
1188             sum_nb_total, sum_nb_errors = 0, 0
1189             for one_input, one_result in zip(input, result):
1190                 seq = [self.id2token[i.item()] for i in one_result]
1191                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1192                 sum_nb_total += 1
1193                 sum_nb_errors += 0 if nb_errors == 0 else 1
1194                 if nb_to_log > 0:
1195                     gt_seq = [self.id2token[i.item()] for i in one_input]
1196                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1197                     gt_prog = " ".join([str(x) for x in gt_prog])
1198                     prog = " ".join([str(x) for x in prog])
1199                     comment = "*" if nb_errors == 0 else "-"
1200                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1201                     for start_stack, target_stack, result_stack, correct in stacks:
1202                         comment = "*" if correct else "-"
1203                         start_stack = " ".join([str(x) for x in start_stack])
1204                         target_stack = " ".join([str(x) for x in target_stack])
1205                         result_stack = " ".join([str(x) for x in result_stack])
1206                         logger(
1207                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1208                         )
1209                     nb_to_log -= 1
1210
1211             return sum_nb_total, sum_nb_errors
1212
1213         # --------------------------------------------------------------------
1214         def compute_nb_errors_output(input, nb_to_log=0):
1215             result = input.clone()
1216             k = torch.arange(result.size(1), device=result.device)[None, :]
1217             last_output_idx = (
1218                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1219             )
1220             first_prog_idx = (
1221                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1222             )
1223             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1224             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1225
1226             masked_inplace_autoregression(
1227                 model,
1228                 self.batch_size,
1229                 result,
1230                 ar_mask,
1231                 deterministic_synthesis,
1232                 device=self.device,
1233             )
1234
1235             sum_nb_total, sum_nb_errors = 0, 0
1236             for one_input, one_result, i, j in zip(
1237                 input, result, last_output_idx, first_prog_idx
1238             ):
1239                 seq = [self.id2token[i.item()] for i in one_result]
1240                 sum_nb_total += 1
1241                 correct = (one_input - one_result).abs().max() == 0
1242                 sum_nb_errors += 0 if correct else 1
1243                 if nb_to_log > 0:
1244                     result_stack = [
1245                         self.id2token[i.item()] for i in one_result[i : j + 1]
1246                     ]
1247                     target_stack = [
1248                         self.id2token[i.item()] for i in one_input[i : j + 1]
1249                     ]
1250                     comment = "*" if correct else "-"
1251                     result_stack = " ".join([str(x) for x in result_stack])
1252                     target_stack = " ".join([str(x) for x in target_stack])
1253                     logger(
1254                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1255                     )
1256                     nb_to_log -= 1
1257
1258             return sum_nb_total, sum_nb_errors
1259
1260         # --------------------------------------------------------------------
1261
1262         if not self.no_prog:
1263             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1264                 self.test_input[:1000].to(self.device), nb_to_log=10
1265             )
1266
1267             logger(
1268                 f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1269             )
1270
1271         test_nb_total, test_nb_errors = compute_nb_errors_output(
1272             self.test_input[:1000].to(self.device), nb_to_log=10
1273         )
1274
1275         logger(
1276             f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1277         )
1278
1279
1280 ######################################################################
1281
1282
1283 import expr
1284
1285
1286 class Expr(Task):
1287     def tensorize(self, sequences):
1288         len_max = max([len(x) for x in sequences])
1289         return torch.cat(
1290             [
1291                 torch.tensor(
1292                     [
1293                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1294                         for s in sequences
1295                     ]
1296                 )
1297             ],
1298             0,
1299         ).to(self.device)
1300
1301     def __init__(
1302         self,
1303         nb_train_samples,
1304         nb_test_samples,
1305         nb_variables,
1306         sequence_length,
1307         operand_max,
1308         result_max,
1309         batch_size,
1310         device=torch.device("cpu"),
1311     ):
1312         super().__init__()
1313
1314         self.batch_size = batch_size
1315         self.device = device
1316
1317         train_sequences = expr.generate_sequences(
1318             nb_train_samples,
1319             nb_variables=nb_variables,
1320             length=sequence_length,
1321             operand_max=operand_max,
1322             result_max=result_max,
1323         )
1324
1325         test_sequences = expr.generate_sequences(
1326             nb_test_samples,
1327             nb_variables=nb_variables,
1328             length=sequence_length,
1329             operand_max=operand_max,
1330             result_max=result_max,
1331         )
1332
1333         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1334         symbols.sort()
1335
1336         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1337         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1338
1339         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1340
1341         self.train_input = self.tensorize(train_sequences)
1342         self.test_input = self.tensorize(test_sequences)
1343
1344         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1345
1346     def batches(self, split="train", nb_to_use=-1, desc=None):
1347         assert split in {"train", "test"}
1348         input = self.train_input if split == "train" else self.test_input
1349         if nb_to_use > 0:
1350             input = input[:nb_to_use]
1351         if desc is None:
1352             desc = f"epoch-{split}"
1353         for batch in tqdm.tqdm(
1354             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1355         ):
1356             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1357             batch = batch[:, :last]
1358             yield batch
1359
1360     def vocabulary_size(self):
1361         return self.nb_codes
1362
1363     def seq2str(self, s):
1364         return "".join([self.id2char[k.item()] for k in s])
1365
1366     def produce_results(
1367         self,
1368         n_epoch,
1369         model,
1370         result_dir,
1371         logger,
1372         deterministic_synthesis,
1373         input_file=None,
1374     ):
1375         def compute_nb_correct(input):
1376             result = input.clone()
1377             s = (result == self.space).long()
1378             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1379             result = (1 - ar_mask) * result + ar_mask * self.filler
1380             masked_inplace_autoregression(
1381                 model,
1382                 self.batch_size,
1383                 result,
1384                 ar_mask,
1385                 deterministic_synthesis,
1386                 device=self.device,
1387             )
1388
1389             nb_total = input.size(0)
1390             nb_correct = (input == result).long().min(1).values.sum()
1391
1392             #######################################################################
1393             # Comput predicted vs. true variable values
1394
1395             nb_delta = torch.zeros(5, dtype=torch.int64)
1396             nb_missed = 0
1397
1398             values_input = expr.extract_results([self.seq2str(s) for s in input])
1399             values_result = expr.extract_results([self.seq2str(s) for s in result])
1400
1401             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1402
1403             with open(filename, "w") as f:
1404                 for i, r in zip(values_input, values_result):
1405                     for n, vi in i.items():
1406                         vr = r.get(n)
1407                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1408
1409                         if vr is None or vr < 0:
1410                             nb_missed += 1
1411                         else:
1412                             d = abs(vr - vi)
1413                             if d >= nb_delta.size(0):
1414                                 nb_missed += 1
1415                             else:
1416                                 nb_delta[d] += 1
1417
1418             ######################################################################
1419
1420             return nb_total, nb_correct, nb_delta, nb_missed
1421
1422         (
1423             test_nb_total,
1424             test_nb_correct,
1425             test_nb_delta,
1426             test_nb_missed,
1427         ) = compute_nb_correct(self.test_input[:10000])
1428
1429         logger(
1430             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}%"
1431         )
1432
1433         nb_total = test_nb_delta.sum() + test_nb_missed
1434         for d in range(test_nb_delta.size(0)):
1435             logger(
1436                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1437             )
1438         logger(
1439             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1440         )
1441
1442         ##############################################################
1443         # Log a few generated sequences
1444         if input_file is None:
1445             input = self.test_input[:10]
1446         else:
1447             with open(input_file, "r") as f:
1448                 sequences = [e.strip() for e in f.readlines()]
1449                 sequences = [s + " " + "#" * 50 for s in sequences]
1450                 input = self.tensorize(sequences)
1451
1452         result = input.clone()
1453         s = (result == self.space).long()
1454         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1455         result = (1 - ar_mask) * result + ar_mask * self.filler
1456
1457         for n in range(result.size(0)):
1458             logger(f"test_before {self.seq2str(result[n])}")
1459
1460         masked_inplace_autoregression(
1461             model,
1462             self.batch_size,
1463             result,
1464             ar_mask,
1465             deterministic_synthesis,
1466             device=self.device,
1467         )
1468
1469         correct = (1 - ar_mask) * self.space + ar_mask * input
1470         for n in range(result.size(0)):
1471             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1472             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1473             logger(f"truth       {self.seq2str(correct[n])}")
1474         ##############################################################
1475
1476
1477 ######################################################################
1478
1479 import world
1480
1481
1482 class World(Task):
1483     def __init__(
1484         self,
1485         nb_train_samples,
1486         nb_test_samples,
1487         batch_size,
1488         vqae_nb_epochs,
1489         logger=None,
1490         device=torch.device("cpu"),
1491         device_storage=torch.device("cpu"),
1492     ):
1493         super().__init__()
1494
1495         self.batch_size = batch_size
1496         self.device = device
1497
1498         (
1499             train_frames,
1500             train_action_seq,
1501             test_frames,
1502             test_action_seq,
1503             self.frame2seq,
1504             self.seq2frame,
1505         ) = world.create_data_and_processors(
1506             nb_train_samples,
1507             nb_test_samples,
1508             mode="first_last",
1509             nb_steps=30,
1510             nb_epochs=vqae_nb_epochs,
1511             logger=logger,
1512             device=device,
1513             device_storage=device_storage,
1514         )
1515
1516         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
1517         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
1518
1519         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
1520         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
1521
1522         self.len_frame_seq = train_frame_seq.size(1)
1523         self.len_action_seq = train_action_seq.size(1)
1524         self.nb_codes = nb_frame_codes + nb_action_codes
1525
1526         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
1527
1528         train_action_seq += nb_frame_codes
1529         self.train_input = torch.cat(
1530             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
1531         )
1532
1533         test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
1534         test_action_seq += nb_frame_codes
1535         self.test_input = torch.cat(
1536             (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
1537         )
1538
1539     def batches(self, split="train", nb_to_use=-1, desc=None):
1540         assert split in {"train", "test"}
1541         input = self.train_input if split == "train" else self.test_input
1542         if nb_to_use > 0:
1543             input = input[:nb_to_use]
1544         if desc is None:
1545             desc = f"epoch-{split}"
1546         for batch in tqdm.tqdm(
1547             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1548         ):
1549             yield batch.to(self.device)
1550
1551     def vocabulary_size(self):
1552         return self.nb_codes
1553
1554     def produce_results(
1555         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1556     ):
1557         k = torch.arange(
1558             2 * self.len_frame_seq + self.len_action_seq, device=self.device
1559         )[None, :]
1560
1561         input = self.test_input[:64].to(self.device)
1562         result = input.clone()
1563
1564         ar_mask = (
1565             (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
1566         )
1567         result *= 1 - ar_mask
1568
1569         masked_inplace_autoregression(
1570             model,
1571             self.batch_size,
1572             result,
1573             ar_mask,
1574             deterministic_synthesis,
1575             device=self.device,
1576         )
1577
1578         seq_start = input[:, : self.len_frame_seq]
1579         seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
1580         seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
1581
1582         result = torch.cat(
1583             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
1584         )
1585         result = result.reshape(-1, result.size(-1))
1586
1587         frames = self.seq2frame(result)
1588         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1589         torchvision.utils.save_image(
1590             frames.float() / (world.Box.nb_rgb_levels - 1),
1591             image_name,
1592             nrow=12,
1593             padding=1,
1594             pad_value=0.0,
1595         )
1596         logger(f"wrote {image_name}")
1597
1598
1599 ######################################################################