tasks.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math, os, tqdm
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 from mygpt import BracketedSequence
  16
  17 try:
  18     from graph import save_attention_image
  19 except ImportError:
  20     save_attention_image = None
  21
  22 ######################################################################
  23
  24
  25 def masked_inplace_autoregression(
  26     model,
  27     batch_size,
  28     input,
  29     ar_mask,
  30     deterministic_synthesis,
  31     forbidden_tokens=None,
  32     progress_bar_desc="autoregression",
  33     device=torch.device("cpu"),
  34 ):
  35     assert input.size() == ar_mask.size()
  36
  37     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
  38
  39     if progress_bar_desc is not None:
  40         batches = tqdm.tqdm(
  41             batches,
  42             dynamic_ncols=True,
  43             desc=progress_bar_desc,
  44             total=(input.size(0) + batch_size - 1) // batch_size,
  45         )
  46
  47     with torch.autograd.no_grad():
  48         t = model.training
  49         model.eval()
  50
  51         for input, ar_mask in batches:
  52             model.masked_inplace_autoregression(
  53                 input, ar_mask, forbidden_tokens, deterministic_synthesis
  54             )
  55
  56         model.train(t)
  57
  58
  59 ######################################################################
  60
  61
  62 class Task:
  63     def batches(self, split="train"):
  64         pass
  65
  66     def vocabulary_size(self):
  67         pass
  68
  69     def produce_results(
  70         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  71     ):
  72         pass
  73
  74
  75 ######################################################################
  76
  77
  78 class Problem:
  79     def generate_sequences(self, nb):
  80         pass
  81
  82     def seq2str(self, seq):
  83         return "[NOT IMPLEMENTED]"
  84
  85
  86 ####################
  87
  88
  89 class ProblemLevel0(Problem):
  90     def __init__(self, nb_sentences=100, len_prompt=5, len_result=5):
  91         self.seq = torch.randint(10, (nb_sentences, len_prompt + 1 + len_result))
  92         self.seq[:, len_prompt] = 10
  93
  94     def generate_sequences(self, nb):
  95         sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
  96         ar_mask = (sequences == 10).long()
  97         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
  98         return sequences, ar_mask
  99
 100
 101 class ProblemLevel1(Problem):
 102     def __init__(self, nb_operators=100, len_source=5, len_result=8):
 103         self.len_source = len_source
 104         self.len_result = len_result
 105         self.len_nb_operator = int(math.log(nb_operators) / math.log(10)) + 1
 106         self.operators = F.one_hot(
 107             torch.rand(nb_operators, len_result, len_source).argmax(-1),
 108             num_classes=len_source,
 109         )
 110
 111     def generate_sequences(self, nb):
 112         nb_operators = torch.randint(self.operators.size(0), (nb,))
 113         operators = self.operators[nb_operators]
 114         nb_operators = (
 115             nb_operators[:, None]
 116             // 10 ** torch.arange(self.len_nb_operator - 1, -1, -1)
 117         ) % 10
 118         marker1 = torch.full((nb, 1), 10)
 119         # source = torch.randint(10, (nb, self.len_source))
 120         source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
 121         marker2 = torch.full((nb, 1), 11)
 122         result = operators.bmm(source[:, :, None]).squeeze(-1)
 123         sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
 124         ar_mask = (sequences == 11).long()
 125         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
 126         return sequences, ar_mask
 127
 128     def seq2str(self, seq):
 129         return "".join("0123456789|>"[x.item()] for x in seq)
 130
 131
 132 class ProblemLevel2(Problem):
 133     def __init__(self, len_source=5, len_result=8):
 134         self.len_source = len_source
 135         self.len_result = len_result
 136
 137     def generate_sequences(self, nb):
 138         operators = F.one_hot(
 139             torch.rand(nb, self.len_result, self.len_source).argmax(-1),
 140             num_classes=self.len_source,
 141         )
 142         source1 = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
 143         marker1 = torch.full((nb, 1), 10)
 144         result1 = operators.bmm(source1[:, :, None]).squeeze(-1)
 145         marker2 = torch.full((nb, 1), 11)
 146         source2 = torch.randint(10, (nb, self.len_source))
 147         marker3 = torch.full((nb, 1), 12)
 148         result2 = operators.bmm(source2[:, :, None]).squeeze(-1)
 149
 150         sequences = torch.cat(
 151             (source1, marker1, result1, marker2, source2, marker3, result2), 1
 152         )
 153         ar_mask = (sequences == 12).long()
 154         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
 155         return sequences, ar_mask
 156
 157     def seq2str(self, seq):
 158         return "".join("0123456789>|~"[x.item()] for x in seq)
 159
 160
 161 ####################
 162
 163
 164 class ProblemAddition(Problem):
 165     def __init__(self, nb_digits=10, zero_padded=False, inverted_result=False):
 166         self.nb_digits = nb_digits
 167         self.zero_padded = zero_padded
 168         self.inverted_result = inverted_result
 169         self.char2id = dict([(c, n) for n, c in enumerate("0123456789+=$")])
 170         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 171
 172     def tensorize(self, strings):
 173         len_max = max([len(x) for x in strings])
 174         return torch.cat(
 175             [
 176                 torch.tensor(
 177                     [
 178                         [self.char2id[c] for c in s + "$" * (len_max - len(s))]
 179                         for s in strings
 180                     ]
 181                 )
 182             ],
 183             0,
 184         )
 185
 186     def generate_sequences(self, nb):
 187         sequences = []
 188         for k in range(nb):
 189             a, b = torch.randint(10**self.nb_digits, (2,))
 190             c = a + b
 191             a, b, c = str(a.item()), str(b.item()), str(c.item())
 192             if self.zero_padded:
 193                 a = "0" * (self.nb_digits - len(a)) + a
 194                 b = "0" * (self.nb_digits - len(b)) + b
 195                 c = "0" * (self.nb_digits + 1 - len(c)) + c
 196             if self.inverted_result:
 197                 c = c[::-1]
 198             sequences.append(f"{a}+{b}={c}$")
 199
 200         sequences = self.tensorize(sequences)
 201         ar_mask = (sequences == self.char2id["="]).long()
 202         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
 203         return sequences, ar_mask
 204
 205     def seq2str(self, seq):
 206         return "".join(self.id2char[x.item()] for x in seq)
 207
 208
 209 # class ProblemUnion(Problem):
 210 # problems = [ProblemByheart()]
 211 # nb_common_codes = 100
 212
 213 # def generate_sequences(nb_samples):
 214 # problem_indexes = torch.randint(len(problems), (nb_samples,))
 215 # nb_samples_per_problem = torch.one_hot(problem_indexes).sum(0)
 216 # print(f"{nb_samples_per_problem}")
 217 # all_seq = []
 218 # for nb, p in zip(nb_samples_per_problem, problems):
 219 # all_seq.append(p.generate_sequences(nb_samples_per_problem[nb]))
 220 # return all_seq
 221
 222 # for strain, stest in zip(train_seq, test_seq):
 223 # s = torch.cat((strain, stest), 0)
 224
 225 ####################
 226
 227
 228 class SandBox(Task):
 229     def __init__(
 230         self,
 231         problem,
 232         nb_train_samples,
 233         nb_test_samples,
 234         batch_size,
 235         logger=None,
 236         device=torch.device("cpu"),
 237         max_nb_codes=1024,
 238     ):
 239         super().__init__()
 240
 241         self.batch_size = batch_size
 242         self.device = device
 243         self.problem = problem
 244
 245         self.train_input, self.train_ar_mask = self.problem.generate_sequences(
 246             nb_train_samples
 247         )
 248         self.test_input, self.test_ar_mask = self.problem.generate_sequences(
 249             nb_test_samples
 250         )
 251
 252         self.train_input, self.train_ar_mask = self.train_input.to(
 253             device
 254         ), self.train_ar_mask.to(device)
 255         self.test_input, self.test_ar_mask = self.test_input.to(
 256             device
 257         ), self.test_ar_mask.to(device)
 258
 259         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 260
 261         # A bit of paranoia never hurts
 262         assert (
 263             self.nb_codes <= max_nb_codes
 264             and self.train_input.min() >= 0
 265             and self.test_input.min() >= 0
 266             and tuple(self.train_ar_mask.unique()) == (0, 1)
 267             and tuple(self.test_ar_mask.unique()) == (0, 1)
 268         )
 269
 270     def batches(self, split="train", nb_to_use=-1, desc=None):
 271         assert split in {"train", "test"}
 272         input = self.train_input if split == "train" else self.test_input
 273         if nb_to_use > 0:
 274             input = input[:nb_to_use]
 275         if desc is None:
 276             desc = f"epoch-{split}"
 277         for batch in tqdm.tqdm(
 278             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 279         ):
 280             yield batch
 281
 282     def vocabulary_size(self):
 283         return self.nb_codes
 284
 285     def produce_results(
 286         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 287     ):
 288         def compute_accuracy(input, ar_mask, logger=None):
 289             input, ar_mask = input[:nmax], ar_mask[:nmax]
 290             result = input.clone() * (1 - ar_mask)
 291
 292             masked_inplace_autoregression(
 293                 model,
 294                 self.batch_size,
 295                 result,
 296                 ar_mask,
 297                 deterministic_synthesis,
 298                 progress_bar_desc=None,
 299                 device=self.device,
 300             )
 301
 302             if logger is not None:
 303                 for sp, st in zip(result[:10], input[:10]):
 304                     logger(
 305                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 306                     )
 307                     logger(
 308                         f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 309                     )
 310
 311             nb_total = ar_mask.sum().item()
 312             nb_correct = ((result == input).long() * ar_mask).sum().item()
 313
 314             return nb_total, nb_correct
 315
 316         train_nb_total, train_nb_correct = compute_accuracy(
 317             self.train_input, self.train_ar_mask
 318         )
 319
 320         logger(
 321             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}%"
 322         )
 323
 324         test_nb_total, test_nb_correct = compute_accuracy(
 325             self.test_input, self.test_ar_mask, logger
 326         )
 327
 328         logger(
 329             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}%"
 330         )
 331
 332
 333 ######################################################################
 334
 335 import picoclvr
 336
 337
 338 class PicoCLVR(Task):
 339     # Make a tensor from a list of strings
 340     def tensorize(self, descr):
 341         token_descr = [s.strip().split(" ") for s in descr]
 342         l = max([len(s) for s in token_descr])
 343         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 344         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 345         return torch.tensor(id_descr, device=self.device)
 346
 347     # Make a list of strings from a tensor
 348     def detensorize(self, x):
 349         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 350
 351     # trim all the tensors in the tuple z to remove as much token from
 352     # left and right in the first tensor. If z is a tuple, all its
 353     # elements are trimed according to the triming for the first
 354     def trim(self, z, token="<nul>"):
 355         n = self.token2id[token]
 356         if type(z) == tuple:
 357             x = z[0]
 358             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 359             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 360             return tuple([t[:, a:b] for t in z])
 361         else:
 362             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 363             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 364             return z[:, a:b]
 365
 366     ######################
 367
 368     def __init__(
 369         self,
 370         nb_train_samples,
 371         nb_test_samples,
 372         batch_size,
 373         height,
 374         width,
 375         nb_colors=5,
 376         logger=None,
 377         device=torch.device("cpu"),
 378         pruner_train=None,
 379         pruner_eval=None,
 380     ):
 381         super().__init__()
 382
 383         def generate_descr(nb, cache_suffix, pruner):
 384             return picoclvr.generate(
 385                 nb,
 386                 height=self.height,
 387                 width=self.width,
 388                 nb_colors=nb_colors,
 389                 pruner=pruner,
 390             )
 391
 392         self.height = height
 393         self.width = width
 394         self.batch_size = batch_size
 395         self.device = device
 396         self.pruner_train = pruner_train
 397         self.pruner_eval = pruner_eval
 398
 399         if logger is not None:
 400             logger(
 401                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 402             )
 403
 404         self.train_descr = generate_descr(
 405             nb_train_samples, "train", pruner=self.pruner_train
 406         )
 407         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 408
 409         # Build the tokenizer
 410         tokens = {"<nul>", "<img>"}
 411         for d in [self.train_descr, self.test_descr]:
 412             for s in d:
 413                 for t in s.strip().split(" "):
 414                     tokens.add(t)
 415         # make this set a sorted list to get the same tensors given
 416         # the same descr
 417         tokens = list(tokens)
 418         tokens.sort()
 419         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 420         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 421         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 422
 423         # Tokenize the train and test sets
 424         self.train_input = self.tensorize(self.train_descr)
 425         self.test_input = self.tensorize(self.test_descr)
 426
 427     def batches(self, split="train"):
 428         assert split in {"train", "test"}
 429         input = self.train_input if split == "train" else self.test_input
 430         for batch in tqdm.tqdm(
 431             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 432         ):
 433             yield self.trim(batch)
 434
 435     def vocabulary_size(self):
 436         return len(self.token2id)
 437
 438     def compute_missing_properties(
 439         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 440     ):
 441         acc_nb_requested_properties = []
 442         acc_nb_missing_properties = []
 443         acc_nb_results = 0
 444
 445         for input in tqdm.tqdm(
 446             self.test_input.split(self.batch_size),
 447             dynamic_ncols=True,
 448             desc=f"test-properties",
 449         ):
 450             result = input.clone()
 451             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 452             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 453             masked_inplace_autoregression(
 454                 model,
 455                 self.batch_size,
 456                 result,
 457                 ar_mask,
 458                 deterministic_synthesis,
 459                 progress_bar_desc=None,
 460                 device=self.device,
 461             )
 462
 463             result_descr = self.detensorize(result)
 464             np = picoclvr.nb_properties(
 465                 result_descr,
 466                 height=self.height,
 467                 width=self.width,
 468                 pruner=pruner,
 469             )
 470             nb_requested_properties, _, nb_missing_properties = zip(*np)
 471             acc_nb_requested_properties += nb_requested_properties
 472             acc_nb_missing_properties += nb_missing_properties
 473             acc_nb_results += len(result_descr)
 474
 475         nb_requested_properties = sum(acc_nb_requested_properties)
 476         nb_missing_properties = sum(acc_nb_missing_properties)
 477
 478         prefix = "" if pruner is None else "pruned_"
 479         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 480         logger(
 481             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 482         )
 483         logger(
 484             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 485         )
 486
 487     ######################################################################
 488
 489     def produce_results(
 490         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 491     ):
 492         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 493
 494         if self.pruner_eval is not None:
 495             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 496
 497         nb_tokens_to_generate = self.height * self.width + 3
 498         result_descr = []
 499         nb_per_primer = 8
 500         primer = []
 501
 502         for primer_descr in [
 503             "red above green <sep> green top <sep> blue right of red",
 504             "there is red <sep> there is yellow <sep> there is blue",
 505             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 506             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 507         ]:
 508             primer += [primer_descr + " <img>"] * nb_per_primer
 509
 510         result = self.tensorize(primer)
 511         fill = result.new_full(
 512             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 513         )
 514         result = torch.cat((result, fill), 1)
 515         ar_mask = (result == self.t_nul).long()
 516         masked_inplace_autoregression(
 517             model,
 518             self.batch_size,
 519             result,
 520             ar_mask,
 521             deterministic_synthesis,
 522             device=self.device,
 523         )
 524         result_descr = self.detensorize(result)
 525
 526         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 527
 528         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 529         acc_nb_results = len(result_descr)
 530
 531         nb_requested_properties = sum(acc_nb_requested_properties)
 532         nb_missing_properties = sum(acc_nb_missing_properties)
 533
 534         prefix = "demo_"
 535         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 536         logger(
 537             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 538         )
 539         logger(
 540             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 541         )
 542
 543         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 544
 545         if img.dim() == 5:
 546             if img.size(1) == 1:
 547                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 548             else:
 549                 img = torch.cat(
 550                     [
 551                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 552                         for x in img
 553                     ],
 554                     0,
 555                 )
 556
 557         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 558         torchvision.utils.save_image(
 559             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 560         )
 561         logger(f"wrote {image_name}")
 562
 563
 564 ######################################################################
 565
 566
 567 class MNIST(Task):
 568     def __init__(
 569         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 570     ):
 571         super().__init__()
 572
 573         self.nb_train_samples = (nb_train_samples,)
 574         self.nb_test_samples = (nb_test_samples,)
 575         self.batch_size = batch_size
 576         self.device = device
 577         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 578         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 579         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 580         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 581
 582     def batches(self, split="train", nb_to_use=-1, desc=None):
 583         assert split in {"train", "test"}
 584         input = self.train_input if split == "train" else self.test_input
 585         if nb_to_use > 0:
 586             input = input[:nb_to_use]
 587         if desc is None:
 588             desc = f"epoch-{split}"
 589         for batch in tqdm.tqdm(
 590             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 591         ):
 592             yield batch
 593
 594     def vocabulary_size(self):
 595         return 256
 596
 597     def produce_results(
 598         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 599     ):
 600         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 601         ar_mask = torch.full_like(results, 1)
 602         masked_inplace_autoregression(
 603             model,
 604             self.batch_size,
 605             results,
 606             ar_mask,
 607             deterministic_synthesis,
 608             device=self.device,
 609         )
 610         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 611         torchvision.utils.save_image(
 612             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 613             image_name,
 614             nrow=16,
 615             pad_value=0.8,
 616         )
 617         logger(f"wrote {image_name}")
 618
 619
 620 ######################################################################
 621
 622 import maze
 623
 624
 625 class Maze(Task):
 626     def map2seq(self, *m):
 627         return torch.cat([x.flatten(1) for x in m], 1)
 628
 629     def seq2map(self, s):
 630         s = s.reshape(s.size(0), -1, self.height, self.width)
 631         return (s[:, k] for k in range(s.size(1)))
 632
 633     def __init__(
 634         self,
 635         nb_train_samples,
 636         nb_test_samples,
 637         batch_size,
 638         height,
 639         width,
 640         nb_walls,
 641         device=torch.device("cpu"),
 642     ):
 643         super().__init__()
 644
 645         self.batch_size = batch_size
 646         self.height = height
 647         self.width = width
 648         self.device = device
 649
 650         train_mazes, train_paths, _ = maze.create_maze_data(
 651             nb_train_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-train"),
 656         )
 657         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 658
 659         test_mazes, test_paths, _ = maze.create_maze_data(
 660             nb_test_samples,
 661             height=height,
 662             width=width,
 663             nb_walls=nb_walls,
 664             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 665         )
 666         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 667
 668         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 669
 670     def batches(self, split="train", nb_to_use=-1, desc=None):
 671         assert split in {"train", "test"}
 672         input = self.train_input if split == "train" else self.test_input
 673         if nb_to_use > 0:
 674             input = input[:nb_to_use]
 675         if desc is None:
 676             desc = f"epoch-{split}"
 677         for batch in tqdm.tqdm(
 678             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 679         ):
 680             yield batch
 681
 682     def vocabulary_size(self):
 683         return self.nb_codes
 684
 685     def compute_error(
 686         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 687     ):
 688         nb_total, nb_correct = 0, 0
 689         count = torch.zeros(
 690             self.width * self.height,
 691             self.width * self.height,
 692             device=self.device,
 693             dtype=torch.int64,
 694         )
 695
 696         for input in self.batches(split, nb_to_use):
 697             result = input.clone()
 698             ar_mask = result.new_zeros(result.size())
 699             ar_mask[:, self.height * self.width :] = 1
 700             result *= 1 - ar_mask
 701             masked_inplace_autoregression(
 702                 model,
 703                 self.batch_size,
 704                 result,
 705                 ar_mask,
 706                 deterministic_synthesis,
 707                 progress_bar_desc=None,
 708                 device=self.device,
 709             )
 710             mazes, paths = self.seq2map(result)
 711             path_correctness = maze.path_correctness(mazes, paths)
 712             nb_correct += path_correctness.long().sum()
 713             nb_total += mazes.size(0)
 714
 715             optimal_path_lengths = (
 716                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 717             )
 718             predicted_path_lengths = (
 719                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 720             )
 721             optimal_path_lengths = optimal_path_lengths[path_correctness]
 722             predicted_path_lengths = predicted_path_lengths[path_correctness]
 723             count[optimal_path_lengths, predicted_path_lengths] += 1
 724
 725         if count.max() == 0:
 726             count = None
 727         else:
 728             count = count[
 729                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 730             ]
 731
 732         return nb_total, nb_correct, count
 733
 734     def produce_results(
 735         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 736     ):
 737         train_nb_total, train_nb_correct, count = self.compute_error(
 738             model,
 739             "train",
 740             nb_to_use=1000,
 741             deterministic_synthesis=deterministic_synthesis,
 742         )
 743         logger(
 744             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}%"
 745         )
 746
 747         test_nb_total, test_nb_correct, count = self.compute_error(
 748             model,
 749             "test",
 750             nb_to_use=1000,
 751             deterministic_synthesis=deterministic_synthesis,
 752         )
 753         logger(
 754             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}%"
 755         )
 756
 757         if count is not None:
 758             proportion_optimal = count.diagonal().sum().float() / count.sum()
 759             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 760             with open(
 761                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 762             ) as f:
 763                 for i in range(count.size(0)):
 764                     for j in range(count.size(1)):
 765                         eol = " " if j < count.size(1) - 1 else "\n"
 766                         f.write(f"{count[i,j]}{eol}")
 767
 768         input = self.test_input[:48]
 769         result = input.clone()
 770         ar_mask = result.new_zeros(result.size())
 771         ar_mask[:, self.height * self.width :] = 1
 772         result *= 1 - ar_mask
 773         masked_inplace_autoregression(
 774             model,
 775             self.batch_size,
 776             result,
 777             ar_mask,
 778             deterministic_synthesis,
 779             device=self.device,
 780         )
 781
 782         mazes, paths = self.seq2map(input)
 783         _, predicted_paths = self.seq2map(result)
 784
 785         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 786         maze.save_image(
 787             filename,
 788             mazes=mazes,
 789             target_paths=paths,
 790             predicted_paths=predicted_paths,
 791             path_correct=maze.path_correctness(mazes, predicted_paths),
 792             path_optimal=maze.path_optimality(paths, predicted_paths),
 793         )
 794         logger(f"wrote {filename}")
 795
 796
 797 ######################################################################
 798
 799
 800 import snake
 801
 802
 803 class Snake(Task):
 804     def __init__(
 805         self,
 806         nb_train_samples,
 807         nb_test_samples,
 808         batch_size,
 809         height,
 810         width,
 811         nb_colors,
 812         length,
 813         prompt_length,
 814         device=torch.device("cpu"),
 815     ):
 816         super().__init__()
 817
 818         self.batch_size = batch_size
 819         self.height = height
 820         self.width = width
 821         self.device = device
 822         self.prompt_length = prompt_length
 823
 824         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 825             nb_train_samples,
 826             height,
 827             width,
 828             nb_colors,
 829             length,
 830             prompt_length,
 831             self.device,
 832         )
 833         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 834             nb_test_samples,
 835             height,
 836             width,
 837             nb_colors,
 838             length,
 839             prompt_length,
 840             self.device,
 841         )
 842
 843         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 844
 845     def batches(self, split="train", nb_to_use=-1, desc=None):
 846         assert split in {"train", "test"}
 847         input = self.train_input if split == "train" else self.test_input
 848         if nb_to_use > 0:
 849             input = input[:nb_to_use]
 850         if desc is None:
 851             desc = f"epoch-{split}"
 852         for batch in tqdm.tqdm(
 853             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 854         ):
 855             yield batch
 856
 857     def vocabulary_size(self):
 858         return self.nb_codes
 859
 860     def produce_results(
 861         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 862     ):
 863         def compute_nb_correct(input, prior_visits):
 864             result = input.clone()
 865             i = torch.arange(result.size(1), device=result.device)[None, :]
 866             ar_mask = (
 867                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 868                 .long()
 869                 .expand_as(result)
 870             )
 871             result *= 1 - ar_mask
 872
 873             masked_inplace_autoregression(
 874                 model,
 875                 self.batch_size,
 876                 result,
 877                 ar_mask,
 878                 deterministic_synthesis,
 879                 device=self.device,
 880             )
 881
 882             nb_total = ((prior_visits > 0) * ar_mask).sum()
 883
 884             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 885
 886             return nb_total, nb_correct
 887
 888         test_nb_total, test_nb_correct = compute_nb_correct(
 889             self.test_input[:1000], self.test_prior_visits[:1000]
 890         )
 891
 892         logger(
 893             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}%"
 894         )
 895
 896
 897 ######################################################################
 898
 899
 900 import stack
 901
 902
 903 class Stack(Task):
 904     def __init__(
 905         self,
 906         nb_train_samples,
 907         nb_test_samples,
 908         batch_size,
 909         logger,
 910         nb_steps,
 911         nb_stacks,
 912         nb_digits,
 913         fraction_values_for_train=None,
 914         device=torch.device("cpu"),
 915     ):
 916         super().__init__()
 917
 918         self.batch_size = batch_size
 919         self.nb_steps = nb_steps
 920         self.nb_stacks = nb_stacks
 921         self.nb_digits = nb_digits
 922         self.device = device
 923
 924         if fraction_values_for_train is None:
 925             values_for_train = None
 926             values_for_test = None
 927         else:
 928             all = torch.randperm(10**nb_digits)
 929             nb_for_train = int(all.size(0) * fraction_values_for_train)
 930             values_for_train = all[:nb_for_train]
 931             values_for_test = all[nb_for_train:]
 932
 933         self.train_input, self.train_stack_counts = stack.generate_sequences(
 934             nb_train_samples,
 935             nb_steps,
 936             nb_stacks,
 937             nb_digits,
 938             values_for_train,
 939             self.device,
 940         )
 941
 942         self.test_input, self.test_stack_counts = stack.generate_sequences(
 943             nb_test_samples,
 944             nb_steps,
 945             nb_stacks,
 946             nb_digits,
 947             values_for_test,
 948             self.device,
 949         )
 950
 951         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 952         counts = self.test_stack_counts.flatten()[i.flatten()]
 953         counts = F.one_hot(counts).sum(0)
 954         logger(f"test_pop_stack_counts {counts}")
 955
 956         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 957
 958     def batches(self, split="train", nb_to_use=-1, desc=None):
 959         assert split in {"train", "test"}
 960         input = self.train_input if split == "train" else self.test_input
 961         if nb_to_use > 0:
 962             input = input[:nb_to_use]
 963         if desc is None:
 964             desc = f"epoch-{split}"
 965         for batch in tqdm.tqdm(
 966             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 967         ):
 968             yield batch
 969
 970     def vocabulary_size(self):
 971         return self.nb_codes
 972
 973     def produce_results(
 974         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 975     ):
 976         def compute_nb_correct(input):
 977             result = input.clone()
 978             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 979             ar_mask = (result != input).long()
 980             masked_inplace_autoregression(
 981                 model,
 982                 self.batch_size,
 983                 result,
 984                 ar_mask,
 985                 deterministic_synthesis,
 986                 device=self.device,
 987             )
 988
 989             errors = ((result != input).long() * ar_mask).reshape(
 990                 -1, 1 + self.nb_digits
 991             )
 992             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 993
 994             nb_total = ar_mask.max(1).values.sum()
 995             nb_correct = nb_total - errors.max(1).values.sum()
 996
 997             return nb_total, nb_correct
 998
 999         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
1000
1001         logger(
1002             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}%"
1003         )
1004
1005         ##############################################################
1006         # Log a few generated sequences
1007         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
1008         result = input.clone()
1009         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1010         ar_mask = (result != input).long()
1011
1012         # for n in range(result.size(0)):
1013         # logger(
1014         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1015         # )
1016
1017         masked_inplace_autoregression(
1018             model,
1019             self.batch_size,
1020             result,
1021             ar_mask,
1022             deterministic_synthesis,
1023             device=self.device,
1024         )
1025
1026         for n in range(result.size(0)):
1027             logger(
1028                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1029             )
1030         ##############################################################
1031
1032
1033 ######################################################################
1034
1035 import rpl
1036
1037
1038 class RPL(Task):
1039     def tensorize(self, sequences):
1040         len_max = max([len(x) for x in sequences])
1041         return torch.cat(
1042             [
1043                 torch.tensor(
1044                     [
1045                         [
1046                             self.token2id[str(c)]
1047                             for c in s + ["<nul>"] * (len_max - len(s))
1048                         ]
1049                         for s in sequences
1050                     ]
1051                 )
1052             ],
1053             0,
1054         )
1055
1056     def seq2str(self, seq):
1057         return " ".join([self.id2token[i] for i in seq])
1058
1059     def __init__(
1060         self,
1061         nb_train_samples,
1062         nb_test_samples,
1063         batch_size,
1064         nb_starting_values=3,
1065         max_input=9,
1066         prog_len=6,
1067         nb_runs=5,
1068         no_prog=False,
1069         logger=None,
1070         device=torch.device("cpu"),
1071     ):
1072         super().__init__()
1073
1074         self.batch_size = batch_size
1075         self.device = device
1076         self.no_prog = no_prog
1077
1078         train_sequences = [
1079             rpl.generate(
1080                 nb_starting_values=nb_starting_values,
1081                 nb_result_values_max=4 * nb_starting_values,
1082                 max_input=max_input,
1083                 prog_len=prog_len,
1084                 nb_runs=nb_runs,
1085             )
1086             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
1087         ]
1088
1089         test_sequences = [
1090             rpl.generate(
1091                 nb_starting_values=nb_starting_values,
1092                 nb_result_values_max=4 * nb_starting_values,
1093                 max_input=max_input,
1094                 prog_len=prog_len,
1095                 nb_runs=nb_runs,
1096             )
1097             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
1098         ]
1099
1100         symbols = list(
1101             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
1102         )
1103         val_max = max([x if type(x) is int else 0 for x in symbols])
1104         symbols = list(filter(lambda x: type(x) is str, symbols))
1105         symbols.sort()
1106         symbols += [str(n) for n in range(val_max + 1)]
1107         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
1108         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
1109
1110         self.t_nul = self.token2id["<nul>"]
1111         self.t_input = self.token2id["<in>"]
1112         self.t_output = self.token2id["<out>"]
1113         self.t_prog = self.token2id["<prg>"]
1114         self.t_end = self.token2id["<end>"]
1115
1116         self.train_input = self.tensorize(train_sequences)
1117         self.test_input = self.tensorize(test_sequences)
1118
1119         if no_prog:
1120             # Excise the program from every train and test example
1121             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
1122                 None, :
1123             ]
1124             p = (
1125                 ((self.train_input == self.t_prog).long() * k)
1126                 .max(1, keepdim=True)
1127                 .values
1128             )
1129             self.train_input = (
1130                 self.train_input * (k <= p).long()
1131                 + self.t_end * (k == p + 1).long()
1132                 + self.t_nul * (k > p + 1).long()
1133             )
1134             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
1135                 None, :
1136             ]
1137             p = (
1138                 ((self.test_input == self.t_prog).long() * k)
1139                 .max(1, keepdim=True)
1140                 .values
1141             )
1142             self.test_input = (
1143                 self.test_input * (k <= p).long()
1144                 + self.t_end * (k == p + 1).long()
1145                 + self.t_nul * (k > p + 1).long()
1146             )
1147
1148         if logger is not None:
1149             logger(f"value_max {val_max}")
1150             for x in self.train_input[:25]:
1151                 end = (x != self.t_nul).nonzero().max().item() + 1
1152                 seq = [self.id2token[i.item()] for i in x[:end]]
1153                 s = " ".join(seq)
1154                 logger(f"example_seq {s}")
1155
1156         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1157
1158     def batches(self, split="train", nb_to_use=-1, desc=None):
1159         assert split in {"train", "test"}
1160         input = self.train_input if split == "train" else self.test_input
1161         if nb_to_use > 0:
1162             input = input[:nb_to_use]
1163         if desc is None:
1164             desc = f"epoch-{split}"
1165         for batch in tqdm.tqdm(
1166             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1167         ):
1168             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1169             batch = batch[:, :last].to(self.device)
1170             yield batch
1171
1172     def vocabulary_size(self):
1173         return self.nb_codes
1174
1175     def produce_results(
1176         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1177     ):
1178         # --------------------------------------------------------------------
1179         def compute_nb_errors_prog(input, nb_to_log=0):
1180             result = input.clone()
1181             s = (result == self.t_prog).long()
1182             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1183             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1184
1185             masked_inplace_autoregression(
1186                 model,
1187                 self.batch_size,
1188                 result,
1189                 ar_mask,
1190                 deterministic_synthesis,
1191                 device=self.device,
1192             )
1193
1194             sum_nb_total, sum_nb_errors = 0, 0
1195             for one_input, one_result in zip(input, result):
1196                 seq = [self.id2token[i.item()] for i in one_result]
1197                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1198                 sum_nb_total += 1
1199                 sum_nb_errors += 0 if nb_errors == 0 else 1
1200                 if nb_to_log > 0:
1201                     gt_seq = [self.id2token[i.item()] for i in one_input]
1202                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1203                     gt_prog = " ".join([str(x) for x in gt_prog])
1204                     prog = " ".join([str(x) for x in prog])
1205                     comment = "*" if nb_errors == 0 else "-"
1206                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1207                     for start_stack, target_stack, result_stack, correct in stacks:
1208                         comment = "*" if correct else "-"
1209                         start_stack = " ".join([str(x) for x in start_stack])
1210                         target_stack = " ".join([str(x) for x in target_stack])
1211                         result_stack = " ".join([str(x) for x in result_stack])
1212                         logger(
1213                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1214                         )
1215                     nb_to_log -= 1
1216
1217             return sum_nb_total, sum_nb_errors
1218
1219         # --------------------------------------------------------------------
1220         def compute_nb_errors_output(input, nb_to_log=0):
1221             result = input.clone()
1222             k = torch.arange(result.size(1), device=result.device)[None, :]
1223             last_output_idx = (
1224                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1225             )
1226             first_prog_idx = (
1227                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1228             )
1229             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1230             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1231
1232             masked_inplace_autoregression(
1233                 model,
1234                 self.batch_size,
1235                 result,
1236                 ar_mask,
1237                 deterministic_synthesis,
1238                 device=self.device,
1239             )
1240
1241             sum_nb_total, sum_nb_errors = 0, 0
1242             for one_input, one_result, i, j in zip(
1243                 input, result, last_output_idx, first_prog_idx
1244             ):
1245                 seq = [self.id2token[i.item()] for i in one_result]
1246                 sum_nb_total += 1
1247                 correct = (one_input - one_result).abs().max() == 0
1248                 sum_nb_errors += 0 if correct else 1
1249                 if nb_to_log > 0:
1250                     result_stack = [
1251                         self.id2token[i.item()] for i in one_result[i : j + 1]
1252                     ]
1253                     target_stack = [
1254                         self.id2token[i.item()] for i in one_input[i : j + 1]
1255                     ]
1256                     comment = "*" if correct else "-"
1257                     result_stack = " ".join([str(x) for x in result_stack])
1258                     target_stack = " ".join([str(x) for x in target_stack])
1259                     logger(
1260                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1261                     )
1262                     nb_to_log -= 1
1263
1264             return sum_nb_total, sum_nb_errors
1265
1266         # --------------------------------------------------------------------
1267
1268         if not self.no_prog:
1269             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1270                 self.test_input[:1000].to(self.device), nb_to_log=10
1271             )
1272
1273             logger(
1274                 f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1275             )
1276
1277         test_nb_total, test_nb_errors = compute_nb_errors_output(
1278             self.test_input[:1000].to(self.device), nb_to_log=10
1279         )
1280
1281         logger(
1282             f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1283         )
1284
1285         if save_attention_image is not None:
1286             input = self.test_input[:1].clone()
1287             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1288             input = input[:, :last].to(self.device)
1289
1290             with torch.autograd.no_grad():
1291                 t = model.training
1292                 model.eval()
1293                 model.record_attention(True)
1294                 model(BracketedSequence(input))
1295                 model.train(t)
1296                 ram = model.retrieve_attention()
1297                 model.record_attention(False)
1298
1299             tokens_output = [self.id2token[i.item()] for i in input[0]]
1300             tokens_input = ["n/a"] + tokens_output[:-1]
1301             for n_head in range(ram[0].size(1)):
1302                 filename = os.path.join(
1303                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1304                 )
1305                 attention_matrices = [m[0, n_head] for m in ram]
1306                 save_attention_image(
1307                     filename,
1308                     tokens_input,
1309                     tokens_output,
1310                     attention_matrices,
1311                     k_top=10,
1312                     # min_total_attention=0.9,
1313                     token_gap=12,
1314                     layer_gap=50,
1315                 )
1316                 logger(f"wrote {filename}")
1317
1318
1319 ######################################################################
1320
1321
1322 import expr
1323
1324
1325 class Expr(Task):
1326     def tensorize(self, sequences):
1327         len_max = max([len(x) for x in sequences])
1328         return torch.cat(
1329             [
1330                 torch.tensor(
1331                     [
1332                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1333                         for s in sequences
1334                     ]
1335                 )
1336             ],
1337             0,
1338         ).to(self.device)
1339
1340     def __init__(
1341         self,
1342         nb_train_samples,
1343         nb_test_samples,
1344         nb_variables,
1345         sequence_length,
1346         operand_max,
1347         result_max,
1348         batch_size,
1349         device=torch.device("cpu"),
1350     ):
1351         super().__init__()
1352
1353         self.batch_size = batch_size
1354         self.device = device
1355
1356         train_sequences = expr.generate_sequences(
1357             nb_train_samples,
1358             nb_variables=nb_variables,
1359             length=sequence_length,
1360             operand_max=operand_max,
1361             result_max=result_max,
1362         )
1363
1364         test_sequences = expr.generate_sequences(
1365             nb_test_samples,
1366             nb_variables=nb_variables,
1367             length=sequence_length,
1368             operand_max=operand_max,
1369             result_max=result_max,
1370         )
1371
1372         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1373         symbols.sort()
1374
1375         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1376         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1377
1378         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1379
1380         self.train_input = self.tensorize(train_sequences)
1381         self.test_input = self.tensorize(test_sequences)
1382
1383         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1384
1385     def batches(self, split="train", nb_to_use=-1, desc=None):
1386         assert split in {"train", "test"}
1387         input = self.train_input if split == "train" else self.test_input
1388         if nb_to_use > 0:
1389             input = input[:nb_to_use]
1390         if desc is None:
1391             desc = f"epoch-{split}"
1392         for batch in tqdm.tqdm(
1393             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1394         ):
1395             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1396             batch = batch[:, :last]
1397             yield batch
1398
1399     def vocabulary_size(self):
1400         return self.nb_codes
1401
1402     def seq2str(self, s):
1403         return "".join([self.id2char[k.item()] for k in s])
1404
1405     def produce_results(
1406         self,
1407         n_epoch,
1408         model,
1409         result_dir,
1410         logger,
1411         deterministic_synthesis,
1412         input_file=None,
1413     ):
1414         def compute_nb_correct(input):
1415             result = input.clone()
1416             s = (result == self.space).long()
1417             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1418             result = (1 - ar_mask) * result + ar_mask * self.filler
1419             masked_inplace_autoregression(
1420                 model,
1421                 self.batch_size,
1422                 result,
1423                 ar_mask,
1424                 deterministic_synthesis,
1425                 device=self.device,
1426             )
1427
1428             nb_total = input.size(0)
1429             nb_correct = (input == result).long().min(1).values.sum()
1430
1431             #######################################################################
1432             # Comput predicted vs. true variable values
1433
1434             nb_delta = torch.zeros(5, dtype=torch.int64)
1435             nb_missed = 0
1436
1437             values_input = expr.extract_results([self.seq2str(s) for s in input])
1438             values_result = expr.extract_results([self.seq2str(s) for s in result])
1439
1440             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1441
1442             with open(filename, "w") as f:
1443                 for i, r in zip(values_input, values_result):
1444                     for n, vi in i.items():
1445                         vr = r.get(n)
1446                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1447
1448                         if vr is None or vr < 0:
1449                             nb_missed += 1
1450                         else:
1451                             d = abs(vr - vi)
1452                             if d >= nb_delta.size(0):
1453                                 nb_missed += 1
1454                             else:
1455                                 nb_delta[d] += 1
1456
1457             ######################################################################
1458
1459             return nb_total, nb_correct, nb_delta, nb_missed
1460
1461         (
1462             test_nb_total,
1463             test_nb_correct,
1464             test_nb_delta,
1465             test_nb_missed,
1466         ) = compute_nb_correct(self.test_input[:10000])
1467
1468         logger(
1469             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}%"
1470         )
1471
1472         nb_total = test_nb_delta.sum() + test_nb_missed
1473         for d in range(test_nb_delta.size(0)):
1474             logger(
1475                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1476             )
1477         logger(
1478             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1479         )
1480
1481         ##############################################################
1482         # Log a few generated sequences
1483         if input_file is None:
1484             input = self.test_input[:10]
1485         else:
1486             with open(input_file, "r") as f:
1487                 sequences = [e.strip() for e in f.readlines()]
1488                 sequences = [s + " " + "#" * 50 for s in sequences]
1489                 input = self.tensorize(sequences)
1490
1491         result = input.clone()
1492         s = (result == self.space).long()
1493         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1494         result = (1 - ar_mask) * result + ar_mask * self.filler
1495
1496         for n in range(result.size(0)):
1497             logger(f"test_before {self.seq2str(result[n])}")
1498
1499         masked_inplace_autoregression(
1500             model,
1501             self.batch_size,
1502             result,
1503             ar_mask,
1504             deterministic_synthesis,
1505             device=self.device,
1506         )
1507
1508         correct = (1 - ar_mask) * self.space + ar_mask * input
1509         for n in range(result.size(0)):
1510             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1511             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1512             logger(f"truth       {self.seq2str(correct[n])}")
1513         ##############################################################
1514
1515
1516 ######################################################################
1517
1518 import world
1519
1520
1521 class World(Task):
1522     def __init__(
1523         self,
1524         nb_train_samples,
1525         nb_test_samples,
1526         batch_size,
1527         vqae_nb_epochs,
1528         logger=None,
1529         device=torch.device("cpu"),
1530         device_storage=torch.device("cpu"),
1531     ):
1532         super().__init__()
1533
1534         self.batch_size = batch_size
1535         self.device = device
1536
1537         (
1538             train_frames,
1539             train_action_seq,
1540             test_frames,
1541             test_action_seq,
1542             self.frame2seq,
1543             self.seq2frame,
1544         ) = world.create_data_and_processors(
1545             nb_train_samples,
1546             nb_test_samples,
1547             mode="first_last",
1548             nb_steps=30,
1549             nb_epochs=vqae_nb_epochs,
1550             logger=logger,
1551             device=device,
1552             device_storage=device_storage,
1553         )
1554
1555         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
1556         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
1557
1558         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
1559         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
1560
1561         self.len_frame_seq = train_frame_seq.size(1)
1562         self.len_action_seq = train_action_seq.size(1)
1563         self.nb_codes = nb_frame_codes + nb_action_codes
1564
1565         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
1566
1567         train_action_seq += nb_frame_codes
1568         self.train_input = torch.cat(
1569             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
1570         )
1571
1572         test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
1573         test_action_seq += nb_frame_codes
1574         self.test_input = torch.cat(
1575             (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
1576         )
1577
1578     def batches(self, split="train", nb_to_use=-1, desc=None):
1579         assert split in {"train", "test"}
1580         input = self.train_input if split == "train" else self.test_input
1581         if nb_to_use > 0:
1582             input = input[:nb_to_use]
1583         if desc is None:
1584             desc = f"epoch-{split}"
1585         for batch in tqdm.tqdm(
1586             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1587         ):
1588             yield batch.to(self.device)
1589
1590     def vocabulary_size(self):
1591         return self.nb_codes
1592
1593     def produce_results(
1594         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1595     ):
1596         k = torch.arange(
1597             2 * self.len_frame_seq + self.len_action_seq, device=self.device
1598         )[None, :]
1599
1600         input = self.test_input[:64].to(self.device)
1601         result = input.clone()
1602
1603         ar_mask = (
1604             (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
1605         )
1606         result *= 1 - ar_mask
1607
1608         masked_inplace_autoregression(
1609             model,
1610             self.batch_size,
1611             result,
1612             ar_mask,
1613             deterministic_synthesis,
1614             device=self.device,
1615         )
1616
1617         seq_start = input[:, : self.len_frame_seq]
1618         seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
1619         seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
1620
1621         result = torch.cat(
1622             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
1623         )
1624         result = result.reshape(-1, result.size(-1))
1625
1626         frames = self.seq2frame(result)
1627         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1628         torchvision.utils.save_image(
1629             frames.float() / (world.Box.nb_rgb_levels - 1),
1630             image_name,
1631             nrow=12,
1632             padding=1,
1633             pad_value=0.0,
1634         )
1635         logger(f"wrote {image_name}")
1636
1637
1638 ######################################################################