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