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