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 import problems
  78
  79
  80 class SandBox(Task):
  81     def __init__(
  82         self,
  83         problem,
  84         nb_train_samples,
  85         nb_test_samples,
  86         batch_size,
  87         logger=None,
  88         device=torch.device("cpu"),
  89         max_nb_codes=1024,
  90     ):
  91         super().__init__()
  92
  93         self.batch_size = batch_size
  94         self.device = device
  95         self.problem = problem
  96
  97         self.train_input, self.train_ar_mask = self.problem.generate_sequences(
  98             nb_train_samples
  99         )
 100         self.test_input, self.test_ar_mask = self.problem.generate_sequences(
 101             nb_test_samples
 102         )
 103
 104         self.train_input, self.train_ar_mask = self.train_input.to(
 105             device
 106         ), self.train_ar_mask.to(device)
 107         self.test_input, self.test_ar_mask = self.test_input.to(
 108             device
 109         ), self.test_ar_mask.to(device)
 110
 111         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 112
 113         # A bit of paranoia never hurts
 114         assert (
 115             self.nb_codes <= max_nb_codes
 116             and self.train_input.min() >= 0
 117             and self.test_input.min() >= 0
 118             and tuple(self.train_ar_mask.unique()) == (0, 1)
 119             and tuple(self.test_ar_mask.unique()) == (0, 1)
 120         )
 121
 122     def batches(self, split="train", nb_to_use=-1, desc=None):
 123         assert split in {"train", "test"}
 124         input = self.train_input if split == "train" else self.test_input
 125         if nb_to_use > 0:
 126             input = input[:nb_to_use]
 127         if desc is None:
 128             desc = f"epoch-{split}"
 129         for batch in tqdm.tqdm(
 130             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 131         ):
 132             yield batch
 133
 134     def vocabulary_size(self):
 135         return self.nb_codes
 136
 137     def produce_results(
 138         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 139     ):
 140         def compute_accuracy(input, ar_mask, logger=None):
 141             input, ar_mask = input[:nmax], ar_mask[:nmax]
 142             result = input.clone() * (1 - ar_mask)
 143
 144             masked_inplace_autoregression(
 145                 model,
 146                 self.batch_size,
 147                 result,
 148                 ar_mask,
 149                 deterministic_synthesis,
 150                 progress_bar_desc=None,
 151                 device=self.device,
 152             )
 153
 154             if logger is not None:
 155                 for sp, st in zip(result[:10], input[:10]):
 156                     logger(
 157                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 158                     )
 159                     logger(
 160                         f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 161                     )
 162
 163             nb_total = ar_mask.sum().item()
 164             nb_correct = ((result == input).long() * ar_mask).sum().item()
 165
 166             return nb_total, nb_correct
 167
 168         train_nb_total, train_nb_correct = compute_accuracy(
 169             self.train_input, self.train_ar_mask
 170         )
 171
 172         logger(
 173             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}%"
 174         )
 175
 176         test_nb_total, test_nb_correct = compute_accuracy(
 177             self.test_input, self.test_ar_mask, logger
 178         )
 179
 180         logger(
 181             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}%"
 182         )
 183
 184         if save_attention_image is None:
 185             logger("no save_attention_image (is pycairo installed?)")
 186         else:
 187             for k in range(10):
 188                 ns = torch.randint(self.test_input.size(0), (1,)).item()
 189                 input = self.test_input[ns : ns + 1].clone()
 190
 191                 with torch.autograd.no_grad():
 192                     t = model.training
 193                     model.eval()
 194                     model.record_attention(True)
 195                     model(BracketedSequence(input))
 196                     model.train(t)
 197                     ram = model.retrieve_attention()
 198                     model.record_attention(False)
 199
 200                 tokens_output = [c for c in self.problem.seq2str(input[0])]
 201                 tokens_input = ["n/a"] + tokens_output[:-1]
 202                 for n_head in range(ram[0].size(1)):
 203                     filename = os.path.join(
 204                         result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
 205                     )
 206                     attention_matrices = [m[0, n_head] for m in ram]
 207                     save_attention_image(
 208                         filename,
 209                         tokens_input,
 210                         tokens_output,
 211                         attention_matrices,
 212                         k_top=10,
 213                         # min_total_attention=0.9,
 214                         token_gap=12,
 215                         layer_gap=50,
 216                     )
 217                     logger(f"wrote {filename}")
 218
 219
 220 ######################################################################
 221
 222 import picoclvr
 223
 224
 225 class PicoCLVR(Task):
 226     # Make a tensor from a list of strings
 227     def tensorize(self, descr):
 228         token_descr = [s.strip().split(" ") for s in descr]
 229         l = max([len(s) for s in token_descr])
 230         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 231         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 232         return torch.tensor(id_descr, device=self.device)
 233
 234     # Make a list of strings from a tensor
 235     def detensorize(self, x):
 236         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 237
 238     # trim all the tensors in the tuple z to remove as much token from
 239     # left and right in the first tensor. If z is a tuple, all its
 240     # elements are trimed according to the triming for the first
 241     def trim(self, z, token="<nul>"):
 242         n = self.token2id[token]
 243         if type(z) == tuple:
 244             x = z[0]
 245             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 246             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 247             return tuple([t[:, a:b] for t in z])
 248         else:
 249             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 250             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 251             return z[:, a:b]
 252
 253     ######################
 254
 255     def __init__(
 256         self,
 257         nb_train_samples,
 258         nb_test_samples,
 259         batch_size,
 260         height,
 261         width,
 262         nb_colors=5,
 263         logger=None,
 264         device=torch.device("cpu"),
 265         pruner_train=None,
 266         pruner_eval=None,
 267     ):
 268         super().__init__()
 269
 270         def generate_descr(nb, cache_suffix, pruner):
 271             return picoclvr.generate(
 272                 nb,
 273                 height=self.height,
 274                 width=self.width,
 275                 nb_colors=nb_colors,
 276                 pruner=pruner,
 277             )
 278
 279         self.height = height
 280         self.width = width
 281         self.batch_size = batch_size
 282         self.device = device
 283         self.pruner_train = pruner_train
 284         self.pruner_eval = pruner_eval
 285
 286         if logger is not None:
 287             logger(
 288                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 289             )
 290
 291         self.train_descr = generate_descr(
 292             nb_train_samples, "train", pruner=self.pruner_train
 293         )
 294         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 295
 296         # Build the tokenizer
 297         tokens = {"<nul>", "<img>"}
 298         for d in [self.train_descr, self.test_descr]:
 299             for s in d:
 300                 for t in s.strip().split(" "):
 301                     tokens.add(t)
 302         # make this set a sorted list to get the same tensors given
 303         # the same descr
 304         tokens = list(tokens)
 305         tokens.sort()
 306         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 307         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 308         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 309
 310         # Tokenize the train and test sets
 311         self.train_input = self.tensorize(self.train_descr)
 312         self.test_input = self.tensorize(self.test_descr)
 313
 314     def batches(self, split="train"):
 315         assert split in {"train", "test"}
 316         input = self.train_input if split == "train" else self.test_input
 317         for batch in tqdm.tqdm(
 318             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 319         ):
 320             yield self.trim(batch)
 321
 322     def vocabulary_size(self):
 323         return len(self.token2id)
 324
 325     def compute_missing_properties(
 326         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 327     ):
 328         acc_nb_requested_properties = []
 329         acc_nb_missing_properties = []
 330         acc_nb_results = 0
 331
 332         for input in tqdm.tqdm(
 333             self.test_input.split(self.batch_size),
 334             dynamic_ncols=True,
 335             desc=f"test-properties",
 336         ):
 337             result = input.clone()
 338             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 339             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 340             masked_inplace_autoregression(
 341                 model,
 342                 self.batch_size,
 343                 result,
 344                 ar_mask,
 345                 deterministic_synthesis,
 346                 progress_bar_desc=None,
 347                 device=self.device,
 348             )
 349
 350             result_descr = self.detensorize(result)
 351             np = picoclvr.nb_properties(
 352                 result_descr,
 353                 height=self.height,
 354                 width=self.width,
 355                 pruner=pruner,
 356             )
 357             nb_requested_properties, _, nb_missing_properties = zip(*np)
 358             acc_nb_requested_properties += nb_requested_properties
 359             acc_nb_missing_properties += nb_missing_properties
 360             acc_nb_results += len(result_descr)
 361
 362         nb_requested_properties = sum(acc_nb_requested_properties)
 363         nb_missing_properties = sum(acc_nb_missing_properties)
 364
 365         prefix = "" if pruner is None else "pruned_"
 366         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 367         logger(
 368             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 369         )
 370         logger(
 371             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 372         )
 373
 374     ######################################################################
 375
 376     def produce_results(
 377         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 378     ):
 379         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 380
 381         if self.pruner_eval is not None:
 382             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 383
 384         nb_tokens_to_generate = self.height * self.width + 3
 385         result_descr = []
 386         nb_per_primer = 8
 387         primer = []
 388
 389         for primer_descr in [
 390             "red above green <sep> green top <sep> blue right of red",
 391             "there is red <sep> there is yellow <sep> there is blue",
 392             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 393             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 394         ]:
 395             primer += [primer_descr + " <img>"] * nb_per_primer
 396
 397         result = self.tensorize(primer)
 398         fill = result.new_full(
 399             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 400         )
 401         result = torch.cat((result, fill), 1)
 402         ar_mask = (result == self.t_nul).long()
 403         masked_inplace_autoregression(
 404             model,
 405             self.batch_size,
 406             result,
 407             ar_mask,
 408             deterministic_synthesis,
 409             device=self.device,
 410         )
 411         result_descr = self.detensorize(result)
 412
 413         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 414
 415         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 416         acc_nb_results = len(result_descr)
 417
 418         nb_requested_properties = sum(acc_nb_requested_properties)
 419         nb_missing_properties = sum(acc_nb_missing_properties)
 420
 421         prefix = "demo_"
 422         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 423         logger(
 424             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 425         )
 426         logger(
 427             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 428         )
 429
 430         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 431
 432         if img.dim() == 5:
 433             if img.size(1) == 1:
 434                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 435             else:
 436                 img = torch.cat(
 437                     [
 438                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 439                         for x in img
 440                     ],
 441                     0,
 442                 )
 443
 444         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 445         torchvision.utils.save_image(
 446             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 447         )
 448         logger(f"wrote {image_name}")
 449
 450
 451 ######################################################################
 452
 453
 454 class MNIST(Task):
 455     def __init__(
 456         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 457     ):
 458         super().__init__()
 459
 460         self.nb_train_samples = (nb_train_samples,)
 461         self.nb_test_samples = (nb_test_samples,)
 462         self.batch_size = batch_size
 463         self.device = device
 464         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 465         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 466         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 467         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 468
 469     def batches(self, split="train", nb_to_use=-1, desc=None):
 470         assert split in {"train", "test"}
 471         input = self.train_input if split == "train" else self.test_input
 472         if nb_to_use > 0:
 473             input = input[:nb_to_use]
 474         if desc is None:
 475             desc = f"epoch-{split}"
 476         for batch in tqdm.tqdm(
 477             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 478         ):
 479             yield batch
 480
 481     def vocabulary_size(self):
 482         return 256
 483
 484     def produce_results(
 485         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 486     ):
 487         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 488         ar_mask = torch.full_like(results, 1)
 489         masked_inplace_autoregression(
 490             model,
 491             self.batch_size,
 492             results,
 493             ar_mask,
 494             deterministic_synthesis,
 495             device=self.device,
 496         )
 497         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 498         torchvision.utils.save_image(
 499             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 500             image_name,
 501             nrow=16,
 502             pad_value=0.8,
 503         )
 504         logger(f"wrote {image_name}")
 505
 506
 507 ######################################################################
 508
 509 import maze
 510
 511
 512 class Maze(Task):
 513     def map2seq(self, *m):
 514         return torch.cat([x.flatten(1) for x in m], 1)
 515
 516     def seq2map(self, s):
 517         s = s.reshape(s.size(0), -1, self.height, self.width)
 518         return (s[:, k] for k in range(s.size(1)))
 519
 520     def __init__(
 521         self,
 522         nb_train_samples,
 523         nb_test_samples,
 524         batch_size,
 525         height,
 526         width,
 527         nb_walls,
 528         device=torch.device("cpu"),
 529     ):
 530         super().__init__()
 531
 532         self.batch_size = batch_size
 533         self.height = height
 534         self.width = width
 535         self.device = device
 536
 537         train_mazes, train_paths, _ = maze.create_maze_data(
 538             nb_train_samples,
 539             height=height,
 540             width=width,
 541             nb_walls=nb_walls,
 542             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 543         )
 544         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 545
 546         test_mazes, test_paths, _ = maze.create_maze_data(
 547             nb_test_samples,
 548             height=height,
 549             width=width,
 550             nb_walls=nb_walls,
 551             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 552         )
 553         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 554
 555         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 556
 557     def batches(self, split="train", nb_to_use=-1, desc=None):
 558         assert split in {"train", "test"}
 559         input = self.train_input if split == "train" else self.test_input
 560         if nb_to_use > 0:
 561             input = input[:nb_to_use]
 562         if desc is None:
 563             desc = f"epoch-{split}"
 564         for batch in tqdm.tqdm(
 565             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 566         ):
 567             yield batch
 568
 569     def vocabulary_size(self):
 570         return self.nb_codes
 571
 572     def compute_error(
 573         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 574     ):
 575         nb_total, nb_correct = 0, 0
 576         count = torch.zeros(
 577             self.width * self.height,
 578             self.width * self.height,
 579             device=self.device,
 580             dtype=torch.int64,
 581         )
 582
 583         for input in self.batches(split, nb_to_use):
 584             result = input.clone()
 585             ar_mask = result.new_zeros(result.size())
 586             ar_mask[:, self.height * self.width :] = 1
 587             result *= 1 - ar_mask
 588             masked_inplace_autoregression(
 589                 model,
 590                 self.batch_size,
 591                 result,
 592                 ar_mask,
 593                 deterministic_synthesis,
 594                 progress_bar_desc=None,
 595                 device=self.device,
 596             )
 597             mazes, paths = self.seq2map(result)
 598             path_correctness = maze.path_correctness(mazes, paths)
 599             nb_correct += path_correctness.long().sum()
 600             nb_total += mazes.size(0)
 601
 602             optimal_path_lengths = (
 603                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 604             )
 605             predicted_path_lengths = (
 606                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 607             )
 608             optimal_path_lengths = optimal_path_lengths[path_correctness]
 609             predicted_path_lengths = predicted_path_lengths[path_correctness]
 610             count[optimal_path_lengths, predicted_path_lengths] += 1
 611
 612         if count.max() == 0:
 613             count = None
 614         else:
 615             count = count[
 616                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 617             ]
 618
 619         return nb_total, nb_correct, count
 620
 621     def produce_results(
 622         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 623     ):
 624         train_nb_total, train_nb_correct, count = self.compute_error(
 625             model,
 626             "train",
 627             nb_to_use=1000,
 628             deterministic_synthesis=deterministic_synthesis,
 629         )
 630         logger(
 631             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}%"
 632         )
 633
 634         test_nb_total, test_nb_correct, count = self.compute_error(
 635             model,
 636             "test",
 637             nb_to_use=1000,
 638             deterministic_synthesis=deterministic_synthesis,
 639         )
 640         logger(
 641             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}%"
 642         )
 643
 644         if count is not None:
 645             proportion_optimal = count.diagonal().sum().float() / count.sum()
 646             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 647             with open(
 648                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 649             ) as f:
 650                 for i in range(count.size(0)):
 651                     for j in range(count.size(1)):
 652                         eol = " " if j < count.size(1) - 1 else "\n"
 653                         f.write(f"{count[i,j]}{eol}")
 654
 655         input = self.test_input[:48]
 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             device=self.device,
 667         )
 668
 669         mazes, paths = self.seq2map(input)
 670         _, predicted_paths = self.seq2map(result)
 671
 672         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 673         maze.save_image(
 674             filename,
 675             mazes=mazes,
 676             target_paths=paths,
 677             predicted_paths=predicted_paths,
 678             path_correct=maze.path_correctness(mazes, predicted_paths),
 679             path_optimal=maze.path_optimality(paths, predicted_paths),
 680         )
 681         logger(f"wrote {filename}")
 682
 683
 684 ######################################################################
 685
 686
 687 import snake
 688
 689
 690 class Snake(Task):
 691     def __init__(
 692         self,
 693         nb_train_samples,
 694         nb_test_samples,
 695         batch_size,
 696         height,
 697         width,
 698         nb_colors,
 699         length,
 700         prompt_length,
 701         device=torch.device("cpu"),
 702     ):
 703         super().__init__()
 704
 705         self.batch_size = batch_size
 706         self.height = height
 707         self.width = width
 708         self.device = device
 709         self.prompt_length = prompt_length
 710
 711         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 712             nb_train_samples,
 713             height,
 714             width,
 715             nb_colors,
 716             length,
 717             prompt_length,
 718             self.device,
 719         )
 720         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 721             nb_test_samples,
 722             height,
 723             width,
 724             nb_colors,
 725             length,
 726             prompt_length,
 727             self.device,
 728         )
 729
 730         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 731
 732     def batches(self, split="train", nb_to_use=-1, desc=None):
 733         assert split in {"train", "test"}
 734         input = self.train_input if split == "train" else self.test_input
 735         if nb_to_use > 0:
 736             input = input[:nb_to_use]
 737         if desc is None:
 738             desc = f"epoch-{split}"
 739         for batch in tqdm.tqdm(
 740             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 741         ):
 742             yield batch
 743
 744     def vocabulary_size(self):
 745         return self.nb_codes
 746
 747     def produce_results(
 748         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 749     ):
 750         def compute_nb_correct(input, prior_visits):
 751             result = input.clone()
 752             i = torch.arange(result.size(1), device=result.device)[None, :]
 753             ar_mask = (
 754                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 755                 .long()
 756                 .expand_as(result)
 757             )
 758             result *= 1 - ar_mask
 759
 760             masked_inplace_autoregression(
 761                 model,
 762                 self.batch_size,
 763                 result,
 764                 ar_mask,
 765                 deterministic_synthesis,
 766                 device=self.device,
 767             )
 768
 769             nb_total = ((prior_visits > 0) * ar_mask).sum()
 770
 771             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 772
 773             return nb_total, nb_correct
 774
 775         test_nb_total, test_nb_correct = compute_nb_correct(
 776             self.test_input[:1000], self.test_prior_visits[:1000]
 777         )
 778
 779         logger(
 780             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}%"
 781         )
 782
 783
 784 ######################################################################
 785
 786
 787 import stack
 788
 789
 790 class Stack(Task):
 791     def __init__(
 792         self,
 793         nb_train_samples,
 794         nb_test_samples,
 795         batch_size,
 796         logger,
 797         nb_steps,
 798         nb_stacks,
 799         nb_digits,
 800         fraction_values_for_train=None,
 801         device=torch.device("cpu"),
 802     ):
 803         super().__init__()
 804
 805         self.batch_size = batch_size
 806         self.nb_steps = nb_steps
 807         self.nb_stacks = nb_stacks
 808         self.nb_digits = nb_digits
 809         self.device = device
 810
 811         if fraction_values_for_train is None:
 812             values_for_train = None
 813             values_for_test = None
 814         else:
 815             all = torch.randperm(10**nb_digits)
 816             nb_for_train = int(all.size(0) * fraction_values_for_train)
 817             values_for_train = all[:nb_for_train]
 818             values_for_test = all[nb_for_train:]
 819
 820         self.train_input, self.train_stack_counts = stack.generate_sequences(
 821             nb_train_samples,
 822             nb_steps,
 823             nb_stacks,
 824             nb_digits,
 825             values_for_train,
 826             self.device,
 827         )
 828
 829         self.test_input, self.test_stack_counts = stack.generate_sequences(
 830             nb_test_samples,
 831             nb_steps,
 832             nb_stacks,
 833             nb_digits,
 834             values_for_test,
 835             self.device,
 836         )
 837
 838         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 839         counts = self.test_stack_counts.flatten()[i.flatten()]
 840         counts = F.one_hot(counts).sum(0)
 841         logger(f"test_pop_stack_counts {counts}")
 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):
 864             result = input.clone()
 865             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 866             ar_mask = (result != input).long()
 867             masked_inplace_autoregression(
 868                 model,
 869                 self.batch_size,
 870                 result,
 871                 ar_mask,
 872                 deterministic_synthesis,
 873                 device=self.device,
 874             )
 875
 876             errors = ((result != input).long() * ar_mask).reshape(
 877                 -1, 1 + self.nb_digits
 878             )
 879             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 880
 881             nb_total = ar_mask.max(1).values.sum()
 882             nb_correct = nb_total - errors.max(1).values.sum()
 883
 884             return nb_total, nb_correct
 885
 886         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 887
 888         logger(
 889             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}%"
 890         )
 891
 892         ##############################################################
 893         # Log a few generated sequences
 894         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 895         result = input.clone()
 896         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 897         ar_mask = (result != input).long()
 898
 899         # for n in range(result.size(0)):
 900         # logger(
 901         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 902         # )
 903
 904         masked_inplace_autoregression(
 905             model,
 906             self.batch_size,
 907             result,
 908             ar_mask,
 909             deterministic_synthesis,
 910             device=self.device,
 911         )
 912
 913         for n in range(result.size(0)):
 914             logger(
 915                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 916             )
 917         ##############################################################
 918
 919
 920 ######################################################################
 921
 922 import rpl
 923
 924
 925 class RPL(Task):
 926     def tensorize(self, sequences):
 927         len_max = max([len(x) for x in sequences])
 928         return torch.cat(
 929             [
 930                 torch.tensor(
 931                     [
 932                         [
 933                             self.token2id[str(c)]
 934                             for c in s + ["<nul>"] * (len_max - len(s))
 935                         ]
 936                         for s in sequences
 937                     ]
 938                 )
 939             ],
 940             0,
 941         )
 942
 943     def seq2str(self, seq):
 944         return " ".join([self.id2token[i] for i in seq])
 945
 946     def __init__(
 947         self,
 948         nb_train_samples,
 949         nb_test_samples,
 950         batch_size,
 951         nb_starting_values=3,
 952         max_input=9,
 953         prog_len=6,
 954         nb_runs=5,
 955         no_prog=False,
 956         logger=None,
 957         device=torch.device("cpu"),
 958     ):
 959         super().__init__()
 960
 961         self.batch_size = batch_size
 962         self.device = device
 963         self.no_prog = no_prog
 964
 965         train_sequences = [
 966             rpl.generate(
 967                 nb_starting_values=nb_starting_values,
 968                 nb_result_values_max=4 * nb_starting_values,
 969                 max_input=max_input,
 970                 prog_len=prog_len,
 971                 nb_runs=nb_runs,
 972             )
 973             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
 974         ]
 975
 976         test_sequences = [
 977             rpl.generate(
 978                 nb_starting_values=nb_starting_values,
 979                 nb_result_values_max=4 * nb_starting_values,
 980                 max_input=max_input,
 981                 prog_len=prog_len,
 982                 nb_runs=nb_runs,
 983             )
 984             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
 985         ]
 986
 987         symbols = list(
 988             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
 989         )
 990         val_max = max([x if type(x) is int else 0 for x in symbols])
 991         symbols = list(filter(lambda x: type(x) is str, symbols))
 992         symbols.sort()
 993         symbols += [str(n) for n in range(val_max + 1)]
 994         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
 995         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
 996
 997         self.t_nul = self.token2id["<nul>"]
 998         self.t_input = self.token2id["<in>"]
 999         self.t_output = self.token2id["<out>"]
1000         self.t_prog = self.token2id["<prg>"]
1001         self.t_end = self.token2id["<end>"]
1002
1003         self.train_input = self.tensorize(train_sequences)
1004         self.test_input = self.tensorize(test_sequences)
1005
1006         if no_prog:
1007             # Excise the program from every train and test example
1008             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
1009                 None, :
1010             ]
1011             p = (
1012                 ((self.train_input == self.t_prog).long() * k)
1013                 .max(1, keepdim=True)
1014                 .values
1015             )
1016             self.train_input = (
1017                 self.train_input * (k <= p).long()
1018                 + self.t_end * (k == p + 1).long()
1019                 + self.t_nul * (k > p + 1).long()
1020             )
1021             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
1022                 None, :
1023             ]
1024             p = (
1025                 ((self.test_input == self.t_prog).long() * k)
1026                 .max(1, keepdim=True)
1027                 .values
1028             )
1029             self.test_input = (
1030                 self.test_input * (k <= p).long()
1031                 + self.t_end * (k == p + 1).long()
1032                 + self.t_nul * (k > p + 1).long()
1033             )
1034
1035         if logger is not None:
1036             logger(f"value_max {val_max}")
1037             for x in self.train_input[:25]:
1038                 end = (x != self.t_nul).nonzero().max().item() + 1
1039                 seq = [self.id2token[i.item()] for i in x[:end]]
1040                 s = " ".join(seq)
1041                 logger(f"example_seq {s}")
1042
1043         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1044
1045     def batches(self, split="train", nb_to_use=-1, desc=None):
1046         assert split in {"train", "test"}
1047         input = self.train_input if split == "train" else self.test_input
1048         if nb_to_use > 0:
1049             input = input[:nb_to_use]
1050         if desc is None:
1051             desc = f"epoch-{split}"
1052         for batch in tqdm.tqdm(
1053             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1054         ):
1055             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1056             batch = batch[:, :last].to(self.device)
1057             yield batch
1058
1059     def vocabulary_size(self):
1060         return self.nb_codes
1061
1062     def produce_results(
1063         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1064     ):
1065         # --------------------------------------------------------------------
1066         def compute_nb_errors_prog(input, nb_to_log=0):
1067             result = input.clone()
1068             s = (result == self.t_prog).long()
1069             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1070             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1071
1072             masked_inplace_autoregression(
1073                 model,
1074                 self.batch_size,
1075                 result,
1076                 ar_mask,
1077                 deterministic_synthesis,
1078                 device=self.device,
1079             )
1080
1081             sum_nb_total, sum_nb_errors = 0, 0
1082             for one_input, one_result in zip(input, result):
1083                 seq = [self.id2token[i.item()] for i in one_result]
1084                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1085                 sum_nb_total += 1
1086                 sum_nb_errors += 0 if nb_errors == 0 else 1
1087                 if nb_to_log > 0:
1088                     gt_seq = [self.id2token[i.item()] for i in one_input]
1089                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1090                     gt_prog = " ".join([str(x) for x in gt_prog])
1091                     prog = " ".join([str(x) for x in prog])
1092                     comment = "*" if nb_errors == 0 else "-"
1093                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1094                     for start_stack, target_stack, result_stack, correct in stacks:
1095                         comment = "*" if correct else "-"
1096                         start_stack = " ".join([str(x) for x in start_stack])
1097                         target_stack = " ".join([str(x) for x in target_stack])
1098                         result_stack = " ".join([str(x) for x in result_stack])
1099                         logger(
1100                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1101                         )
1102                     nb_to_log -= 1
1103
1104             return sum_nb_total, sum_nb_errors
1105
1106         # --------------------------------------------------------------------
1107         def compute_nb_errors_output(input, nb_to_log=0):
1108             result = input.clone()
1109             k = torch.arange(result.size(1), device=result.device)[None, :]
1110             last_output_idx = (
1111                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1112             )
1113             first_prog_idx = (
1114                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1115             )
1116             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1117             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1118
1119             masked_inplace_autoregression(
1120                 model,
1121                 self.batch_size,
1122                 result,
1123                 ar_mask,
1124                 deterministic_synthesis,
1125                 device=self.device,
1126             )
1127
1128             sum_nb_total, sum_nb_errors = 0, 0
1129             for one_input, one_result, i, j in zip(
1130                 input, result, last_output_idx, first_prog_idx
1131             ):
1132                 seq = [self.id2token[i.item()] for i in one_result]
1133                 sum_nb_total += 1
1134                 correct = (one_input - one_result).abs().max() == 0
1135                 sum_nb_errors += 0 if correct else 1
1136                 if nb_to_log > 0:
1137                     result_stack = [
1138                         self.id2token[i.item()] for i in one_result[i : j + 1]
1139                     ]
1140                     target_stack = [
1141                         self.id2token[i.item()] for i in one_input[i : j + 1]
1142                     ]
1143                     comment = "*" if correct else "-"
1144                     result_stack = " ".join([str(x) for x in result_stack])
1145                     target_stack = " ".join([str(x) for x in target_stack])
1146                     logger(
1147                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1148                     )
1149                     nb_to_log -= 1
1150
1151             return sum_nb_total, sum_nb_errors
1152
1153         # --------------------------------------------------------------------
1154
1155         if not self.no_prog:
1156             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1157                 self.test_input[:1000].to(self.device), nb_to_log=10
1158             )
1159
1160             logger(
1161                 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}%"
1162             )
1163
1164         test_nb_total, test_nb_errors = compute_nb_errors_output(
1165             self.test_input[:1000].to(self.device), nb_to_log=10
1166         )
1167
1168         logger(
1169             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}%"
1170         )
1171
1172         if save_attention_image is None:
1173             logger("no save_attention_image (is pycairo installed?)")
1174         else:
1175             ns = torch.randint(self.test_input.size(0), (1,)).item()
1176             input = self.test_input[ns : ns + 1].clone()
1177             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1178             input = input[:, :last].to(self.device)
1179
1180             with torch.autograd.no_grad():
1181                 t = model.training
1182                 model.eval()
1183                 model.record_attention(True)
1184                 model(BracketedSequence(input))
1185                 model.train(t)
1186                 ram = model.retrieve_attention()
1187                 model.record_attention(False)
1188
1189             tokens_output = [self.id2token[i.item()] for i in input[0]]
1190             tokens_input = ["n/a"] + tokens_output[:-1]
1191             for n_head in range(ram[0].size(1)):
1192                 filename = os.path.join(
1193                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1194                 )
1195                 attention_matrices = [m[0, n_head] for m in ram]
1196                 save_attention_image(
1197                     filename,
1198                     tokens_input,
1199                     tokens_output,
1200                     attention_matrices,
1201                     k_top=10,
1202                     # min_total_attention=0.9,
1203                     token_gap=12,
1204                     layer_gap=50,
1205                 )
1206                 logger(f"wrote {filename}")
1207
1208
1209 ######################################################################
1210
1211
1212 import expr
1213
1214
1215 class Expr(Task):
1216     def tensorize(self, sequences):
1217         len_max = max([len(x) for x in sequences])
1218         return torch.cat(
1219             [
1220                 torch.tensor(
1221                     [
1222                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1223                         for s in sequences
1224                     ]
1225                 )
1226             ],
1227             0,
1228         ).to(self.device)
1229
1230     def __init__(
1231         self,
1232         nb_train_samples,
1233         nb_test_samples,
1234         nb_variables,
1235         sequence_length,
1236         operand_max,
1237         result_max,
1238         batch_size,
1239         device=torch.device("cpu"),
1240     ):
1241         super().__init__()
1242
1243         self.batch_size = batch_size
1244         self.device = device
1245
1246         train_sequences = expr.generate_sequences(
1247             nb_train_samples,
1248             nb_variables=nb_variables,
1249             length=sequence_length,
1250             operand_max=operand_max,
1251             result_max=result_max,
1252         )
1253
1254         test_sequences = expr.generate_sequences(
1255             nb_test_samples,
1256             nb_variables=nb_variables,
1257             length=sequence_length,
1258             operand_max=operand_max,
1259             result_max=result_max,
1260         )
1261
1262         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1263         symbols.sort()
1264
1265         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1266         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1267
1268         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1269
1270         self.train_input = self.tensorize(train_sequences)
1271         self.test_input = self.tensorize(test_sequences)
1272
1273         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1274
1275     def batches(self, split="train", nb_to_use=-1, desc=None):
1276         assert split in {"train", "test"}
1277         input = self.train_input if split == "train" else self.test_input
1278         if nb_to_use > 0:
1279             input = input[:nb_to_use]
1280         if desc is None:
1281             desc = f"epoch-{split}"
1282         for batch in tqdm.tqdm(
1283             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1284         ):
1285             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1286             batch = batch[:, :last]
1287             yield batch
1288
1289     def vocabulary_size(self):
1290         return self.nb_codes
1291
1292     def seq2str(self, s):
1293         return "".join([self.id2char[k.item()] for k in s])
1294
1295     def produce_results(
1296         self,
1297         n_epoch,
1298         model,
1299         result_dir,
1300         logger,
1301         deterministic_synthesis,
1302         input_file=None,
1303     ):
1304         def compute_nb_correct(input):
1305             result = input.clone()
1306             s = (result == self.space).long()
1307             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1308             result = (1 - ar_mask) * result + ar_mask * self.filler
1309             masked_inplace_autoregression(
1310                 model,
1311                 self.batch_size,
1312                 result,
1313                 ar_mask,
1314                 deterministic_synthesis,
1315                 device=self.device,
1316             )
1317
1318             nb_total = input.size(0)
1319             nb_correct = (input == result).long().min(1).values.sum()
1320
1321             #######################################################################
1322             # Comput predicted vs. true variable values
1323
1324             nb_delta = torch.zeros(5, dtype=torch.int64)
1325             nb_missed = 0
1326
1327             values_input = expr.extract_results([self.seq2str(s) for s in input])
1328             values_result = expr.extract_results([self.seq2str(s) for s in result])
1329
1330             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1331
1332             with open(filename, "w") as f:
1333                 for i, r in zip(values_input, values_result):
1334                     for n, vi in i.items():
1335                         vr = r.get(n)
1336                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1337
1338                         if vr is None or vr < 0:
1339                             nb_missed += 1
1340                         else:
1341                             d = abs(vr - vi)
1342                             if d >= nb_delta.size(0):
1343                                 nb_missed += 1
1344                             else:
1345                                 nb_delta[d] += 1
1346
1347             ######################################################################
1348
1349             return nb_total, nb_correct, nb_delta, nb_missed
1350
1351         (
1352             test_nb_total,
1353             test_nb_correct,
1354             test_nb_delta,
1355             test_nb_missed,
1356         ) = compute_nb_correct(self.test_input[:10000])
1357
1358         logger(
1359             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}%"
1360         )
1361
1362         nb_total = test_nb_delta.sum() + test_nb_missed
1363         for d in range(test_nb_delta.size(0)):
1364             logger(
1365                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1366             )
1367         logger(
1368             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1369         )
1370
1371         ##############################################################
1372         # Log a few generated sequences
1373         if input_file is None:
1374             input = self.test_input[:10]
1375         else:
1376             with open(input_file, "r") as f:
1377                 sequences = [e.strip() for e in f.readlines()]
1378                 sequences = [s + " " + "#" * 50 for s in sequences]
1379                 input = self.tensorize(sequences)
1380
1381         result = input.clone()
1382         s = (result == self.space).long()
1383         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1384         result = (1 - ar_mask) * result + ar_mask * self.filler
1385
1386         for n in range(result.size(0)):
1387             logger(f"test_before {self.seq2str(result[n])}")
1388
1389         masked_inplace_autoregression(
1390             model,
1391             self.batch_size,
1392             result,
1393             ar_mask,
1394             deterministic_synthesis,
1395             device=self.device,
1396         )
1397
1398         correct = (1 - ar_mask) * self.space + ar_mask * input
1399         for n in range(result.size(0)):
1400             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1401             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1402             logger(f"truth       {self.seq2str(correct[n])}")
1403         ##############################################################
1404
1405
1406 ######################################################################
1407
1408 import world
1409
1410
1411 class World(Task):
1412     def __init__(
1413         self,
1414         nb_train_samples,
1415         nb_test_samples,
1416         batch_size,
1417         vqae_nb_epochs,
1418         logger=None,
1419         device=torch.device("cpu"),
1420         device_storage=torch.device("cpu"),
1421     ):
1422         super().__init__()
1423
1424         self.batch_size = batch_size
1425         self.device = device
1426
1427         (
1428             train_frames,
1429             train_action_seq,
1430             test_frames,
1431             test_action_seq,
1432             self.frame2seq,
1433             self.seq2frame,
1434         ) = world.create_data_and_processors(
1435             nb_train_samples,
1436             nb_test_samples,
1437             mode="first_last",
1438             nb_steps=30,
1439             nb_epochs=vqae_nb_epochs,
1440             logger=logger,
1441             device=device,
1442             device_storage=device_storage,
1443         )
1444
1445         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
1446         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
1447
1448         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
1449         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
1450
1451         self.len_frame_seq = train_frame_seq.size(1)
1452         self.len_action_seq = train_action_seq.size(1)
1453         self.nb_codes = nb_frame_codes + nb_action_codes
1454
1455         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
1456
1457         train_action_seq += nb_frame_codes
1458         self.train_input = torch.cat(
1459             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
1460         )
1461
1462         test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
1463         test_action_seq += nb_frame_codes
1464         self.test_input = torch.cat(
1465             (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
1466         )
1467
1468     def batches(self, split="train", nb_to_use=-1, desc=None):
1469         assert split in {"train", "test"}
1470         input = self.train_input if split == "train" else self.test_input
1471         if nb_to_use > 0:
1472             input = input[:nb_to_use]
1473         if desc is None:
1474             desc = f"epoch-{split}"
1475         for batch in tqdm.tqdm(
1476             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1477         ):
1478             yield batch.to(self.device)
1479
1480     def vocabulary_size(self):
1481         return self.nb_codes
1482
1483     def produce_results(
1484         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1485     ):
1486         k = torch.arange(
1487             2 * self.len_frame_seq + self.len_action_seq, device=self.device
1488         )[None, :]
1489
1490         input = self.test_input[:64].to(self.device)
1491         result = input.clone()
1492
1493         ar_mask = (
1494             (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
1495         )
1496         result *= 1 - ar_mask
1497
1498         masked_inplace_autoregression(
1499             model,
1500             self.batch_size,
1501             result,
1502             ar_mask,
1503             deterministic_synthesis,
1504             device=self.device,
1505         )
1506
1507         seq_start = input[:, : self.len_frame_seq]
1508         seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
1509         seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
1510
1511         result = torch.cat(
1512             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
1513         )
1514         result = result.reshape(-1, result.size(-1))
1515
1516         frames = self.seq2frame(result)
1517         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
1518         torchvision.utils.save_image(
1519             frames.float() / (world.Box.nb_rgb_levels - 1),
1520             image_name,
1521             nrow=12,
1522             padding=1,
1523             pad_value=0.0,
1524         )
1525         logger(f"wrote {image_name}")
1526
1527
1528 ######################################################################