tasks.py

   1 #!/usr/bin/env python
   2
   3 import math, os, tqdm
   4
   5 import torch, torchvision
   6
   7 from torch import nn
   8 from torch.nn import functional as F
   9
  10 ######################################################################
  11
  12
  13 def masked_inplace_autoregression(
  14     model,
  15     batch_size,
  16     input,
  17     ar_mask,
  18     deterministic_synthesis,
  19     forbidden_tokens=None,
  20     progress_bar_desc="autoregression",
  21     device=torch.device("cpu"),
  22 ):
  23     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
  24
  25     if progress_bar_desc is not None:
  26         batches = tqdm.tqdm(
  27             batches,
  28             dynamic_ncols=True,
  29             desc=progress_bar_desc,
  30             total=input.size(0) // batch_size,
  31         )
  32
  33     for input, ar_mask in batches:
  34         model.masked_inplace_autoregression(
  35             input, ar_mask, forbidden_tokens, deterministic_synthesis
  36         )
  37
  38
  39 class Task:
  40     def batches(self, split="train"):
  41         pass
  42
  43     def vocabulary_size(self):
  44         pass
  45
  46     def produce_results(
  47         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  48     ):
  49         pass
  50
  51
  52 ######################################################################
  53
  54 import picoclvr
  55
  56
  57 class PicoCLVR(Task):
  58     # Make a tensor from a list of strings
  59     def tensorize(self, descr):
  60         token_descr = [s.strip().split(" ") for s in descr]
  61         l = max([len(s) for s in token_descr])
  62         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
  63         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
  64         return torch.tensor(id_descr, device=self.device)
  65
  66     # Make a list of strings from a tensor
  67     def detensorize(self, x):
  68         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
  69
  70     # trim all the tensors in the tuple z to remove as much token from
  71     # left and right in the first tensor. If z is a tuple, all its
  72     # elements are trimed according to the triming for the first
  73     def trim(self, z, token="<nul>"):
  74         n = self.token2id[token]
  75         if type(z) == tuple:
  76             x = z[0]
  77             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  78             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  79             return tuple([t[:, a:b] for t in z])
  80         else:
  81             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
  82             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
  83             return z[:, a:b]
  84
  85     ######################
  86     # Not the cleanest part of the code
  87
  88     # Extract the last image of each sequence, from the last <img>
  89     # included, and set to <nul> all the tokens from the beginning of
  90     # that image to the end
  91     def excise_last_image(self, input):
  92         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
  93         nb_img_tokens = self.height * self.width + 1
  94
  95         input = input.clone()
  96         t = (input == t_img).long()
  97         tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
  98         i = (t * tail_masks).nonzero(as_tuple=True)
  99         j = (
 100             i[0][:, None],
 101             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 102         )
 103         images = self.trim(input[j])
 104         input[j] = t_nul
 105         loss_masks = 1 - tail_masks
 106         input, loss_masks = self.trim((input, loss_masks))
 107         return input, loss_masks, images
 108
 109     def add_true_image(self, input, images, loss_masks):
 110         t_nul = self.token2id["<nul>"]
 111         nb_img_tokens = self.height * self.width + 1
 112         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 113         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 114         t = (input == t_nul).long()
 115         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 116         j = (
 117             i[0][:, None],
 118             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 119         )
 120         input[j] = images
 121         loss_masks[j] = 1
 122         input, loss_masks = self.trim((input, loss_masks))
 123         return input, loss_masks
 124
 125     def add_generated_image(self, input, loss_masks, model, deterministic_synthesis):
 126         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 127         nb_img_tokens = self.height * self.width + 1
 128
 129         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 130         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 131         t = (input == t_nul).long()
 132         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 133         input[i] = t_img
 134
 135         j = (
 136             i[0][:, None],
 137             i[1][:, None]
 138             + 1
 139             + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
 140         )
 141         ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
 142         ar_masks[j] = 1
 143         forbidden_tokens = (
 144             torch.arange(self.vocabulary_size(), device=input.device) == t_nul
 145         )
 146         with torch.autograd.no_grad():
 147             t = model.training
 148             model.eval()
 149             masked_inplace_autoregression(
 150                 model,
 151                 self.batch_size,
 152                 input,
 153                 ar_masks,
 154                 deterministic_synthesis,
 155                 forbidden_tokens,
 156                 progress_bar_desc=None,
 157                 device=self.device,
 158             )
 159             model.train(t)
 160
 161         input, loss_masks = self.trim((input, loss_masks))
 162
 163         return input, loss_masks
 164
 165     ######################
 166
 167     def __init__(
 168         self,
 169         nb_train_samples,
 170         nb_test_samples,
 171         batch_size,
 172         height,
 173         width,
 174         nb_colors=5,
 175         logger=None,
 176         device=torch.device("cpu"),
 177         pruner_train=None,
 178         pruner_eval=None,
 179     ):
 180         def generate_descr(nb, cache_suffix, pruner):
 181             return picoclvr.generate(
 182                 nb,
 183                 height=self.height,
 184                 width=self.width,
 185                 nb_colors=nb_colors,
 186                 pruner=pruner,
 187             )
 188
 189         self.height = height
 190         self.width = width
 191         self.batch_size = batch_size
 192         self.device = device
 193         self.pruner_train = pruner_train
 194         self.pruner_eval = pruner_eval
 195
 196         param = {
 197             "nb_train_samples": nb_train_samples,
 198             "nb_test_samples": nb_test_samples,
 199             "height": height,
 200             "width": width,
 201             "nb_colors": nb_colors,
 202             "batch_size": batch_size,
 203             "rng_state": list(torch.get_rng_state()),
 204         }
 205
 206         if logger is not None:
 207             logger(
 208                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 209             )
 210
 211         self.train_descr = generate_descr(
 212             nb_train_samples, "train", pruner=self.pruner_train
 213         )
 214         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 215
 216         # Build the tokenizer
 217         tokens = {"<nul>", "<img>"}
 218         for d in [self.train_descr, self.test_descr]:
 219             for s in d:
 220                 for t in s.strip().split(" "):
 221                     tokens.add(t)
 222         # make this set a sorted list to get the same tensors given
 223         # the same descr
 224         tokens = list(tokens)
 225         tokens.sort()
 226         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 227         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 228
 229         # Tokenize the train and test sets
 230         self.train_input = self.tensorize(self.train_descr)
 231         self.test_input = self.tensorize(self.test_descr)
 232
 233     def batches(self, split="train"):
 234         assert split in {"train", "test"}
 235         input = self.train_input if split == "train" else self.test_input
 236         for batch in tqdm.tqdm(
 237             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 238         ):
 239             yield self.trim(batch)
 240
 241     def vocabulary_size(self):
 242         return len(self.token2id)
 243
 244     def compute_missing_properties(
 245         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 246     ):
 247         acc_nb_requested_properties = []
 248         acc_nb_missing_properties = []
 249         acc_nb_results = 0
 250
 251         for input in tqdm.tqdm(
 252             self.test_input.split(self.batch_size),
 253             dynamic_ncols=True,
 254             desc=f"test-properties",
 255         ):
 256             tape, loss_masks, _ = self.excise_last_image(input)
 257             tape, loss_masks = self.add_generated_image(
 258                 tape, loss_masks, model, deterministic_synthesis
 259             )
 260             result_descr = self.detensorize(tape)
 261             np = picoclvr.nb_properties(
 262                 result_descr,
 263                 height=self.height,
 264                 width=self.width,
 265                 pruner=pruner,
 266             )
 267             nb_requested_properties, _, nb_missing_properties = zip(*np)
 268             acc_nb_requested_properties += nb_requested_properties
 269             acc_nb_missing_properties += nb_missing_properties
 270             acc_nb_results += len(result_descr)
 271
 272         nb_requested_properties = sum(acc_nb_requested_properties)
 273         nb_missing_properties = sum(acc_nb_missing_properties)
 274
 275         prefix = "" if pruner is None else "pruned_"
 276         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 277         logger(
 278             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 279         )
 280         logger(
 281             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 282         )
 283
 284     ######################################################################
 285
 286     def produce_results(
 287         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 288     ):
 289         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 290
 291         if self.pruner_eval is not None:
 292             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 293
 294         nb_tokens_to_generate = self.height * self.width + 3
 295         result_descr = []
 296         nb_per_primer = 8
 297         primer = []
 298
 299         for primer_descr in [
 300             "red above green <sep> green top <sep> blue right of red",
 301             "there is red <sep> there is yellow <sep> there is blue",
 302             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 303             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 304         ]:
 305             primer += [primer_descr] * nb_per_primer
 306
 307         tape = self.tensorize(primer)
 308         loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
 309         tape, loss_masks = self.add_generated_image(
 310             tape, loss_masks, model, deterministic_synthesis
 311         )
 312         result_descr = self.detensorize(tape)
 313
 314         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 315
 316         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 317         acc_nb_results = len(result_descr)
 318
 319         nb_requested_properties = sum(acc_nb_requested_properties)
 320         nb_missing_properties = sum(acc_nb_missing_properties)
 321
 322         prefix = "demo_"
 323         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 324         logger(
 325             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 326         )
 327         logger(
 328             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 329         )
 330
 331         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 332
 333         if img.dim() == 5:
 334             if img.size(1) == 1:
 335                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 336             else:
 337                 img = torch.cat(
 338                     [
 339                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 340                         for x in img
 341                     ],
 342                     0,
 343                 )
 344
 345         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 346         torchvision.utils.save_image(
 347             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 348         )
 349         logger(f"wrote {image_name}")
 350
 351
 352 ######################################################################
 353
 354
 355 class MNIST(Task):
 356     def __init__(
 357         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 358     ):
 359         self.nb_train_samples = (nb_train_samples,)
 360         self.nb_test_samples = (nb_test_samples,)
 361         self.batch_size = batch_size
 362         self.device = device
 363         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 364         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 365         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 366         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 367
 368     def batches(self, split="train", nb_to_use=-1, desc=None):
 369         assert split in {"train", "test"}
 370         input = self.train_input if split == "train" else self.test_input
 371         if nb_to_use > 0:
 372             input = input[:nb_to_use]
 373         if desc is None:
 374             desc = f"epoch-{split}"
 375         for batch in tqdm.tqdm(
 376             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 377         ):
 378             yield batch
 379
 380     def vocabulary_size(self):
 381         return 256
 382
 383     def produce_results(
 384         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 385     ):
 386         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 387         ar_mask = torch.full_like(results, 1)
 388         masked_inplace_autoregression(
 389             model,
 390             self.batch_size,
 391             results,
 392             ar_mask,
 393             deterministic_synthesis,
 394             device=self.device,
 395         )
 396         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 397         torchvision.utils.save_image(
 398             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 399             image_name,
 400             nrow=16,
 401             pad_value=0.8,
 402         )
 403         logger(f"wrote {image_name}")
 404
 405
 406 ######################################################################
 407
 408 import maze
 409
 410
 411 class Maze(Task):
 412     def map2seq(self, *m):
 413         return torch.cat([x.flatten(1) for x in m], 1)
 414
 415     def seq2map(self, s):
 416         s = s.reshape(s.size(0), -1, self.height, self.width)
 417         return (s[:, k] for k in range(s.size(1)))
 418
 419     def __init__(
 420         self,
 421         nb_train_samples,
 422         nb_test_samples,
 423         batch_size,
 424         height,
 425         width,
 426         nb_walls,
 427         device=torch.device("cpu"),
 428     ):
 429         self.batch_size = batch_size
 430         self.height = height
 431         self.width = width
 432         self.device = device
 433
 434         train_mazes, train_paths, _ = maze.create_maze_data(
 435             nb_train_samples,
 436             height=height,
 437             width=width,
 438             nb_walls=nb_walls,
 439             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 440         )
 441         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 442
 443         test_mazes, test_paths, _ = maze.create_maze_data(
 444             nb_test_samples,
 445             height=height,
 446             width=width,
 447             nb_walls=nb_walls,
 448             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 449         )
 450         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 451
 452         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 453
 454     def batches(self, split="train", nb_to_use=-1, desc=None):
 455         assert split in {"train", "test"}
 456         input = self.train_input if split == "train" else self.test_input
 457         if nb_to_use > 0:
 458             input = input[:nb_to_use]
 459         if desc is None:
 460             desc = f"epoch-{split}"
 461         for batch in tqdm.tqdm(
 462             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 463         ):
 464             yield batch
 465
 466     def vocabulary_size(self):
 467         return self.nb_codes
 468
 469     def compute_error(
 470         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 471     ):
 472         nb_total, nb_correct = 0, 0
 473         count = torch.zeros(
 474             self.width * self.height,
 475             self.width * self.height,
 476             device=self.device,
 477             dtype=torch.int64,
 478         )
 479
 480         for input in self.batches(split, nb_to_use):
 481             result = input.clone()
 482             ar_mask = result.new_zeros(result.size())
 483             ar_mask[:, self.height * self.width :] = 1
 484             result *= 1 - ar_mask
 485             masked_inplace_autoregression(
 486                 model,
 487                 self.batch_size,
 488                 result,
 489                 ar_mask,
 490                 deterministic_synthesis,
 491                 progress_bar_desc=None,
 492                 device=self.device,
 493             )
 494             mazes, paths = self.seq2map(result)
 495             path_correctness = maze.path_correctness(mazes, paths)
 496             nb_correct += path_correctness.long().sum()
 497             nb_total += mazes.size(0)
 498
 499             optimal_path_lengths = (
 500                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 501             )
 502             predicted_path_lengths = (
 503                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 504             )
 505             optimal_path_lengths = optimal_path_lengths[path_correctness]
 506             predicted_path_lengths = predicted_path_lengths[path_correctness]
 507             count[optimal_path_lengths, predicted_path_lengths] += 1
 508
 509         if count.max() == 0:
 510             count = None
 511         else:
 512             count = count[
 513                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 514             ]
 515
 516         return nb_total, nb_correct, count
 517
 518     def produce_results(
 519         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 520     ):
 521         with torch.autograd.no_grad():
 522             t = model.training
 523             model.eval()
 524
 525             train_nb_total, train_nb_correct, count = self.compute_error(
 526                 model,
 527                 "train",
 528                 nb_to_use=1000,
 529                 deterministic_synthesis=deterministic_synthesis,
 530             )
 531             logger(
 532                 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}%"
 533             )
 534
 535             test_nb_total, test_nb_correct, count = self.compute_error(
 536                 model,
 537                 "test",
 538                 nb_to_use=1000,
 539                 deterministic_synthesis=deterministic_synthesis,
 540             )
 541             logger(
 542                 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}%"
 543             )
 544
 545             if count is not None:
 546                 proportion_optimal = count.diagonal().sum().float() / count.sum()
 547                 logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 548                 with open(
 549                     os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 550                 ) as f:
 551                     for i in range(count.size(0)):
 552                         for j in range(count.size(1)):
 553                             eol = " " if j < count.size(1) - 1 else "\n"
 554                             f.write(f"{count[i,j]}{eol}")
 555
 556             input = self.test_input[:48]
 557             result = input.clone()
 558             ar_mask = result.new_zeros(result.size())
 559             ar_mask[:, self.height * self.width :] = 1
 560             result *= 1 - ar_mask
 561             masked_inplace_autoregression(
 562                 model,
 563                 self.batch_size,
 564                 result,
 565                 ar_mask,
 566                 deterministic_synthesis,
 567                 device=self.device,
 568             )
 569
 570             mazes, paths = self.seq2map(input)
 571             _, predicted_paths = self.seq2map(result)
 572
 573             filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 574             maze.save_image(
 575                 filename,
 576                 mazes=mazes,
 577                 target_paths=paths,
 578                 predicted_paths=predicted_paths,
 579                 path_correct=maze.path_correctness(mazes, predicted_paths),
 580                 path_optimal=maze.path_optimality(paths, predicted_paths),
 581             )
 582             logger(f"wrote {filename}")
 583
 584             model.train(t)
 585
 586
 587 ######################################################################
 588
 589
 590 import snake
 591
 592
 593 class Snake(Task):
 594     def __init__(
 595         self,
 596         nb_train_samples,
 597         nb_test_samples,
 598         batch_size,
 599         height,
 600         width,
 601         nb_colors,
 602         length,
 603         prompt_length,
 604         device=torch.device("cpu"),
 605     ):
 606         self.batch_size = batch_size
 607         self.height = height
 608         self.width = width
 609         self.device = device
 610         self.prompt_length = prompt_length
 611
 612         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 613             nb_train_samples,
 614             height,
 615             width,
 616             nb_colors,
 617             length,
 618             prompt_length,
 619             self.device,
 620         )
 621         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 622             nb_test_samples,
 623             height,
 624             width,
 625             nb_colors,
 626             length,
 627             prompt_length,
 628             self.device,
 629         )
 630
 631         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 632
 633     def batches(self, split="train", nb_to_use=-1, desc=None):
 634         assert split in {"train", "test"}
 635         input = self.train_input if split == "train" else self.test_input
 636         if nb_to_use > 0:
 637             input = input[:nb_to_use]
 638         if desc is None:
 639             desc = f"epoch-{split}"
 640         for batch in tqdm.tqdm(
 641             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 642         ):
 643             yield batch
 644
 645     def vocabulary_size(self):
 646         return self.nb_codes
 647
 648     def produce_results(
 649         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 650     ):
 651         with torch.autograd.no_grad():
 652             t = model.training
 653             model.eval()
 654
 655             def compute_nb_correct(input, prior_visits):
 656                 result = input.clone()
 657                 i = torch.arange(result.size(1), device=result.device)[None, :]
 658                 ar_mask = (
 659                     torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 660                     .long()
 661                     .expand_as(result)
 662                 )
 663                 result *= 1 - ar_mask
 664
 665                 # snake.solver(result,ar_mask)
 666
 667                 masked_inplace_autoregression(
 668                     model,
 669                     self.batch_size,
 670                     result,
 671                     ar_mask,
 672                     deterministic_synthesis,
 673                     device=self.device,
 674                 )
 675
 676                 nb_total = ((prior_visits > 0) * ar_mask).sum()
 677
 678                 nb_correct = (
 679                     (result == input).long() * (prior_visits > 0) * ar_mask
 680                 ).sum()
 681
 682                 # nb_total = result.size(0)
 683                 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 684
 685                 return nb_total, nb_correct
 686
 687             # train_nb_total, train_nb_correct = compute_nb_correct(
 688             # self.train_input, self.train_prior_visits
 689             # )
 690
 691             # logger(
 692             # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 693             # )
 694
 695             test_nb_total, test_nb_correct = compute_nb_correct(
 696                 self.test_input[:1000], self.test_prior_visits[:1000]
 697             )
 698
 699             logger(
 700                 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}%"
 701             )
 702
 703             model.train(t)
 704
 705
 706 ######################################################################
 707
 708
 709 import stack
 710
 711
 712 class Stack(Task):
 713     def __init__(
 714         self,
 715         nb_train_samples,
 716         nb_test_samples,
 717         batch_size,
 718         logger,
 719         nb_steps,
 720         nb_stacks,
 721         nb_digits,
 722         fraction_values_for_train=None,
 723         device=torch.device("cpu"),
 724     ):
 725         self.batch_size = batch_size
 726         self.nb_steps = nb_steps
 727         self.nb_stacks = nb_stacks
 728         self.nb_digits = nb_digits
 729         self.device = device
 730
 731         if fraction_values_for_train is None:
 732             values_for_train = None
 733             values_for_test = None
 734         else:
 735             all = torch.randperm(10**nb_digits)
 736             nb_for_train = int(all.size(0) * fraction_values_for_train)
 737             values_for_train = all[:nb_for_train]
 738             values_for_test = all[nb_for_train:]
 739
 740         self.train_input, self.train_stack_counts = stack.generate_sequences(
 741             nb_train_samples,
 742             nb_steps,
 743             nb_stacks,
 744             nb_digits,
 745             values_for_train,
 746             self.device,
 747         )
 748
 749         self.test_input, self.test_stack_counts = stack.generate_sequences(
 750             nb_test_samples,
 751             nb_steps,
 752             nb_stacks,
 753             nb_digits,
 754             values_for_test,
 755             self.device,
 756         )
 757
 758         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 759         counts = self.test_stack_counts.flatten()[i.flatten()]
 760         counts = F.one_hot(counts).sum(0)
 761         logger(f"test_pop_stack_counts {counts}")
 762
 763         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 764
 765     def batches(self, split="train", nb_to_use=-1, desc=None):
 766         assert split in {"train", "test"}
 767         input = self.train_input if split == "train" else self.test_input
 768         if nb_to_use > 0:
 769             input = input[:nb_to_use]
 770         if desc is None:
 771             desc = f"epoch-{split}"
 772         for batch in tqdm.tqdm(
 773             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 774         ):
 775             yield batch
 776
 777     def vocabulary_size(self):
 778         return self.nb_codes
 779
 780     def produce_results(
 781         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 782     ):
 783         with torch.autograd.no_grad():
 784             t = model.training
 785             model.eval()
 786
 787             def compute_nb_correct(input):
 788                 result = input.clone()
 789                 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 790                 ar_mask = (result != input).long()
 791                 masked_inplace_autoregression(
 792                     model,
 793                     self.batch_size,
 794                     result,
 795                     ar_mask,
 796                     deterministic_synthesis,
 797                     device=self.device,
 798                 )
 799
 800                 errors = ((result != input).long() * ar_mask).reshape(
 801                     -1, 1 + self.nb_digits
 802                 )
 803                 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 804
 805                 nb_total = ar_mask.max(1).values.sum()
 806                 nb_correct = nb_total - errors.max(1).values.sum()
 807
 808                 return nb_total, nb_correct
 809
 810             test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 811
 812             logger(
 813                 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}%"
 814             )
 815
 816             ##############################################################
 817             # Log a few generated sequences
 818             input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 819             result = input.clone()
 820             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 821             ar_mask = (result != input).long()
 822             for n in range(result.size(0)):
 823                 logger(
 824                     f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 825                 )
 826                 masked_inplace_autoregression(
 827                     model,
 828                     self.batch_size,
 829                     result,
 830                     ar_mask,
 831                     deterministic_synthesis,
 832                     device=self.device,
 833                 )
 834             for n in range(result.size(0)):
 835                 logger(
 836                     f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 837                 )
 838             ##############################################################
 839
 840             model.train(t)
 841
 842
 843 ######################################################################
 844
 845
 846 import expr
 847
 848
 849 class Expr(Task):
 850     def __init__(
 851         self,
 852         nb_train_samples,
 853         nb_test_samples,
 854         nb_variables,
 855         sequence_length,
 856         batch_size,
 857         device=torch.device("cpu"),
 858     ):
 859         self.batch_size = batch_size
 860         self.device = device
 861
 862         train_sequences = expr.generate_sequences(
 863             nb_train_samples,
 864             nb_variables=nb_variables,
 865             length=sequence_length,
 866             # length=2 * sequence_length,
 867             # randomize_length=True,
 868         )
 869         test_sequences = expr.generate_sequences(
 870             nb_test_samples,
 871             nb_variables=nb_variables,
 872             length=sequence_length,
 873         )
 874         self.char2id = dict(
 875             [
 876                 (c, n)
 877                 for n, c in enumerate(
 878                     set("#" + "".join(train_sequences + test_sequences))
 879                 )
 880             ]
 881         )
 882         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 883
 884         self.filler, self.space = self.char2id["#"], self.char2id[" "]
 885
 886         len_max = max([len(x) for x in train_sequences])
 887         self.train_input = torch.cat(
 888             [
 889                 torch.tensor(
 890                     [
 891                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
 892                         for s in train_sequences
 893                     ]
 894                 )
 895             ],
 896             0,
 897         ).to(device)
 898
 899         len_max = max([len(x) for x in test_sequences])
 900         self.test_input = torch.cat(
 901             [
 902                 torch.tensor(
 903                     [
 904                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
 905                         for s in test_sequences
 906                     ]
 907                 )
 908             ],
 909             0,
 910         ).to(device)
 911
 912         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 913
 914     def batches(self, split="train", nb_to_use=-1, desc=None):
 915         assert split in {"train", "test"}
 916         input = self.train_input if split == "train" else self.test_input
 917         if nb_to_use > 0:
 918             input = input[:nb_to_use]
 919         if desc is None:
 920             desc = f"epoch-{split}"
 921         for batch in tqdm.tqdm(
 922             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 923         ):
 924             if split == "train":
 925                 last = (batch != self.filler).max(0).values.nonzero().max() + 3
 926                 batch = batch[:, :last]
 927             yield batch
 928
 929     def vocabulary_size(self):
 930         return self.nb_codes
 931
 932     def seq2str(self, s):
 933         return "".join([self.id2char[k.item()] for k in s])
 934
 935     def produce_results(
 936         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 937     ):
 938         with torch.autograd.no_grad():
 939             t = model.training
 940             model.eval()
 941
 942             def compute_nb_correct(input):
 943                 result = input.clone()
 944                 ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
 945                 result = (1 - ar_mask) * result + ar_mask * self.filler
 946                 masked_inplace_autoregression(
 947                     model,
 948                     self.batch_size,
 949                     result,
 950                     ar_mask,
 951                     deterministic_synthesis,
 952                     device=self.device,
 953                 )
 954
 955                 nb_total = input.size(0)
 956                 nb_correct = (input == result).long().min(1).values.sum()
 957
 958                 #######################################################################
 959                 # Comput predicted vs. true variable values
 960
 961                 nb_delta = torch.zeros(5, dtype=torch.int64)
 962                 nb_missed = 0
 963
 964                 values_input = expr.extract_results([self.seq2str(s) for s in input])
 965                 values_result = expr.extract_results([self.seq2str(s) for s in result])
 966
 967                 for i, r in zip(values_input, values_result):
 968                     for n, vi in i.items():
 969                         vr = r.get(n)
 970                         if vr is None or vr < 0:
 971                             nb_missed += 1
 972                         else:
 973                             d = abs(vr - vi)
 974                             if d >= nb_delta.size(0):
 975                                 nb_missed += 1
 976                             else:
 977                                 nb_delta[d] += 1
 978
 979                 ######################################################################
 980
 981                 return nb_total, nb_correct, nb_delta, nb_missed
 982
 983             (
 984                 test_nb_total,
 985                 test_nb_correct,
 986                 test_nb_delta,
 987                 test_nb_missed,
 988             ) = compute_nb_correct(self.test_input[:1000])
 989
 990             logger(
 991                 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}%"
 992             )
 993
 994             nb_total = test_nb_delta.sum() + test_nb_missed
 995             for d in range(test_nb_delta.size(0)):
 996                 logger(
 997                     f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
 998                 )
 999             logger(
1000                 f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1001             )
1002
1003             ##############################################################
1004             # Log a few generated sequences
1005             input = self.test_input[:10]
1006             result = input.clone()
1007             ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
1008             result = (1 - ar_mask) * result + ar_mask * self.filler
1009             for n in range(result.size(0)):
1010                 logger(f"test_before {self.seq2str(result[n])}")
1011                 masked_inplace_autoregression(
1012                     model,
1013                     self.batch_size,
1014                     result,
1015                     ar_mask,
1016                     deterministic_synthesis,
1017                     device=self.device,
1018                 )
1019             correct = (1 - ar_mask) * self.space + ar_mask * input
1020             for n in range(result.size(0)):
1021                 comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1022                 logger(f"test_after  {self.seq2str(result[n])} {comment}")
1023                 logger(f"correct     {self.seq2str(correct[n])}")
1024             ##############################################################
1025
1026             model.train(t)
1027
1028
1029 ######################################################################