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
 114         # A bit of paranoia never hurts
 115         assert (
 116             self.nb_codes <= max_nb_codes
 117             and self.train_input.min() >= 0
 118             and self.test_input.min() >= 0
 119             and tuple(x.item() for x in self.train_ar_mask.unique()) in { (0,), (1,), (0,1) }
 120             and tuple(x.item() for x in self.test_ar_mask.unique()) in { (0,), (1,), (0,1) }
 121         )
 122
 123     def batches(self, split="train", nb_to_use=-1, desc=None):
 124         assert split in {"train", "test"}
 125         input = self.train_input if split == "train" else self.test_input
 126         if nb_to_use > 0:
 127             input = input[:nb_to_use]
 128         if desc is None:
 129             desc = f"epoch-{split}"
 130         for batch in tqdm.tqdm(
 131             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 132         ):
 133             yield batch
 134
 135     def vocabulary_size(self):
 136         return self.nb_codes
 137
 138     def produce_results(
 139         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 140     ):
 141         def compute_accuracy(input, ar_mask, logger=None):
 142             input, ar_mask = input[:nmax], ar_mask[:nmax]
 143             result = input.clone() * (1 - ar_mask)
 144
 145             masked_inplace_autoregression(
 146                 model,
 147                 self.batch_size,
 148                 result,
 149                 ar_mask,
 150                 deterministic_synthesis,
 151                 progress_bar_desc=None,
 152                 device=self.device,
 153             )
 154
 155             if logger is not None:
 156                 for sp, st in zip(result[:10], input[:10]):
 157                     logger(
 158                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 159                     )
 160                     logger(
 161                         f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 162                     )
 163
 164             nb_total, nb_correct = self.problem.compute_nb_correct(input, ar_mask, result)
 165
 166             # nb_total = ar_mask.sum().item()
 167             # nb_correct = ((result == input).long() * ar_mask).sum().item()
 168
 169             return nb_total, nb_correct
 170
 171         train_nb_total, train_nb_correct = compute_accuracy(
 172             self.train_input, self.train_ar_mask
 173         )
 174
 175         logger(
 176             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}%"
 177         )
 178
 179         test_nb_total, test_nb_correct = compute_accuracy(
 180             self.test_input, self.test_ar_mask, logger
 181         )
 182
 183         logger(
 184             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}%"
 185         )
 186
 187         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 188
 189         if save_attention_image is None:
 190             logger("no save_attention_image (is pycairo installed?)")
 191         else:
 192             for k in range(10):
 193                 ns = torch.randint(self.test_input.size(0), (1,)).item()
 194                 input = self.test_input[ns : ns + 1].clone()
 195
 196                 with torch.autograd.no_grad():
 197                     t = model.training
 198                     model.eval()
 199                     model.record_attention(True)
 200                     model(BracketedSequence(input))
 201                     model.train(t)
 202                     ram = model.retrieve_attention()
 203                     model.record_attention(False)
 204
 205                 tokens_output = [c for c in self.problem.seq2str(input[0])]
 206                 tokens_input = ["n/a"] + tokens_output[:-1]
 207                 for n_head in range(ram[0].size(1)):
 208                     filename = os.path.join(
 209                         result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
 210                     )
 211                     attention_matrices = [m[0, n_head] for m in ram]
 212                     save_attention_image(
 213                         filename,
 214                         tokens_input,
 215                         tokens_output,
 216                         attention_matrices,
 217                         k_top=10,
 218                         # min_total_attention=0.9,
 219                         token_gap=12,
 220                         layer_gap=50,
 221                     )
 222                     logger(f"wrote {filename}")
 223
 224
 225 ######################################################################
 226
 227 import picoclvr
 228
 229
 230 class PicoCLVR(Task):
 231     # Make a tensor from a list of strings
 232     def tensorize(self, descr):
 233         token_descr = [s.strip().split(" ") for s in descr]
 234         l = max([len(s) for s in token_descr])
 235         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 236         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 237         return torch.tensor(id_descr, device=self.device)
 238
 239     # Make a list of strings from a tensor
 240     def detensorize(self, x):
 241         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 242
 243     # trim all the tensors in the tuple z to remove as much token from
 244     # left and right in the first tensor. If z is a tuple, all its
 245     # elements are trimed according to the triming for the first
 246     def trim(self, z, token="<nul>"):
 247         n = self.token2id[token]
 248         if type(z) == tuple:
 249             x = z[0]
 250             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 251             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 252             return tuple([t[:, a:b] for t in z])
 253         else:
 254             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 255             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 256             return z[:, a:b]
 257
 258     ######################
 259
 260     def __init__(
 261         self,
 262         nb_train_samples,
 263         nb_test_samples,
 264         batch_size,
 265         height,
 266         width,
 267         nb_colors=5,
 268         logger=None,
 269         device=torch.device("cpu"),
 270         pruner_train=None,
 271         pruner_eval=None,
 272     ):
 273         super().__init__()
 274
 275         def generate_descr(nb, cache_suffix, pruner):
 276             return picoclvr.generate(
 277                 nb,
 278                 height=self.height,
 279                 width=self.width,
 280                 nb_colors=nb_colors,
 281                 pruner=pruner,
 282             )
 283
 284         self.height = height
 285         self.width = width
 286         self.batch_size = batch_size
 287         self.device = device
 288         self.pruner_train = pruner_train
 289         self.pruner_eval = pruner_eval
 290
 291         if logger is not None:
 292             logger(
 293                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 294             )
 295
 296         self.train_descr = generate_descr(
 297             nb_train_samples, "train", pruner=self.pruner_train
 298         )
 299         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 300
 301         # Build the tokenizer
 302         tokens = {"<nul>", "<img>"}
 303         for d in [self.train_descr, self.test_descr]:
 304             for s in d:
 305                 for t in s.strip().split(" "):
 306                     tokens.add(t)
 307         # make this set a sorted list to get the same tensors given
 308         # the same descr
 309         tokens = list(tokens)
 310         tokens.sort()
 311         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 312         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 313         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 314
 315         # Tokenize the train and test sets
 316         self.train_input = self.tensorize(self.train_descr)
 317         self.test_input = self.tensorize(self.test_descr)
 318
 319     def batches(self, split="train"):
 320         assert split in {"train", "test"}
 321         input = self.train_input if split == "train" else self.test_input
 322         for batch in tqdm.tqdm(
 323             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 324         ):
 325             yield self.trim(batch)
 326
 327     def vocabulary_size(self):
 328         return len(self.token2id)
 329
 330     def compute_missing_properties(
 331         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 332     ):
 333         acc_nb_requested_properties = []
 334         acc_nb_missing_properties = []
 335         acc_nb_results = 0
 336
 337         for input in tqdm.tqdm(
 338             self.test_input.split(self.batch_size),
 339             dynamic_ncols=True,
 340             desc=f"test-properties",
 341         ):
 342             result = input.clone()
 343             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 344             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 345             masked_inplace_autoregression(
 346                 model,
 347                 self.batch_size,
 348                 result,
 349                 ar_mask,
 350                 deterministic_synthesis,
 351                 progress_bar_desc=None,
 352                 device=self.device,
 353             )
 354
 355             result_descr = self.detensorize(result)
 356             np = picoclvr.nb_properties(
 357                 result_descr,
 358                 height=self.height,
 359                 width=self.width,
 360                 pruner=pruner,
 361             )
 362             nb_requested_properties, _, nb_missing_properties = zip(*np)
 363             acc_nb_requested_properties += nb_requested_properties
 364             acc_nb_missing_properties += nb_missing_properties
 365             acc_nb_results += len(result_descr)
 366
 367         nb_requested_properties = sum(acc_nb_requested_properties)
 368         nb_missing_properties = sum(acc_nb_missing_properties)
 369
 370         prefix = "" if pruner is None else "pruned_"
 371         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 372         logger(
 373             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 374         )
 375         logger(
 376             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 377         )
 378
 379         logger(
 380             f"main_test_accuracy {n_epoch} {1-nb_missing_properties/nb_requested_properties}"
 381         )
 382
 383     ######################################################################
 384
 385     def produce_results(
 386         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 387     ):
 388         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 389
 390         if self.pruner_eval is not None:
 391             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 392
 393         nb_tokens_to_generate = self.height * self.width + 3
 394         result_descr = []
 395         nb_per_primer = 8
 396         primer = []
 397
 398         for primer_descr in [
 399             "red above green <sep> green top <sep> blue right of red",
 400             "there is red <sep> there is yellow <sep> there is blue",
 401             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 402             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 403         ]:
 404             primer += [primer_descr + " <img>"] * nb_per_primer
 405
 406         result = self.tensorize(primer)
 407         fill = result.new_full(
 408             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 409         )
 410         result = torch.cat((result, fill), 1)
 411         ar_mask = (result == self.t_nul).long()
 412         masked_inplace_autoregression(
 413             model,
 414             self.batch_size,
 415             result,
 416             ar_mask,
 417             deterministic_synthesis,
 418             device=self.device,
 419         )
 420         result_descr = self.detensorize(result)
 421
 422         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 423
 424         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 425         acc_nb_results = len(result_descr)
 426
 427         nb_requested_properties = sum(acc_nb_requested_properties)
 428         nb_missing_properties = sum(acc_nb_missing_properties)
 429
 430         prefix = "demo_"
 431         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 432         logger(
 433             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 434         )
 435         logger(
 436             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 437         )
 438
 439         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 440
 441         if img.dim() == 5:
 442             if img.size(1) == 1:
 443                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 444             else:
 445                 img = torch.cat(
 446                     [
 447                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 448                         for x in img
 449                     ],
 450                     0,
 451                 )
 452
 453         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 454         torchvision.utils.save_image(
 455             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 456         )
 457         logger(f"wrote {image_name}")
 458
 459
 460 ######################################################################
 461
 462
 463 class MNIST(Task):
 464     def __init__(
 465         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 466     ):
 467         super().__init__()
 468
 469         self.nb_train_samples = (nb_train_samples,)
 470         self.nb_test_samples = (nb_test_samples,)
 471         self.batch_size = batch_size
 472         self.device = device
 473         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 474         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 475         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 476         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 477
 478     def batches(self, split="train", nb_to_use=-1, desc=None):
 479         assert split in {"train", "test"}
 480         input = self.train_input if split == "train" else self.test_input
 481         if nb_to_use > 0:
 482             input = input[:nb_to_use]
 483         if desc is None:
 484             desc = f"epoch-{split}"
 485         for batch in tqdm.tqdm(
 486             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 487         ):
 488             yield batch
 489
 490     def vocabulary_size(self):
 491         return 256
 492
 493     def produce_results(
 494         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 495     ):
 496         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 497         ar_mask = torch.full_like(results, 1)
 498         masked_inplace_autoregression(
 499             model,
 500             self.batch_size,
 501             results,
 502             ar_mask,
 503             deterministic_synthesis,
 504             device=self.device,
 505         )
 506         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 507         torchvision.utils.save_image(
 508             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 509             image_name,
 510             nrow=16,
 511             pad_value=0.8,
 512         )
 513         logger(f"wrote {image_name}")
 514
 515
 516 ######################################################################
 517
 518 import maze
 519
 520
 521 class Maze(Task):
 522     def map2seq(self, *m):
 523         return torch.cat([x.flatten(1) for x in m], 1)
 524
 525     def seq2map(self, s):
 526         s = s.reshape(s.size(0), -1, self.height, self.width)
 527         return (s[:, k] for k in range(s.size(1)))
 528
 529     def __init__(
 530         self,
 531         nb_train_samples,
 532         nb_test_samples,
 533         batch_size,
 534         height,
 535         width,
 536         nb_walls,
 537         device=torch.device("cpu"),
 538     ):
 539         super().__init__()
 540
 541         self.batch_size = batch_size
 542         self.height = height
 543         self.width = width
 544         self.device = device
 545
 546         train_mazes, train_paths, _ = maze.create_maze_data(
 547             nb_train_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-train"),
 552         )
 553         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 554
 555         test_mazes, test_paths, _ = maze.create_maze_data(
 556             nb_test_samples,
 557             height=height,
 558             width=width,
 559             nb_walls=nb_walls,
 560             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 561         )
 562         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 563
 564         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 565
 566     def batches(self, split="train", nb_to_use=-1, desc=None):
 567         assert split in {"train", "test"}
 568         input = self.train_input if split == "train" else self.test_input
 569         if nb_to_use > 0:
 570             input = input[:nb_to_use]
 571         if desc is None:
 572             desc = f"epoch-{split}"
 573         for batch in tqdm.tqdm(
 574             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 575         ):
 576             yield batch
 577
 578     def vocabulary_size(self):
 579         return self.nb_codes
 580
 581     def compute_error(
 582         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 583     ):
 584         nb_total, nb_correct = 0, 0
 585         count = torch.zeros(
 586             self.width * self.height,
 587             self.width * self.height,
 588             device=self.device,
 589             dtype=torch.int64,
 590         )
 591
 592         for input in self.batches(split, nb_to_use):
 593             result = input.clone()
 594             ar_mask = result.new_zeros(result.size())
 595             ar_mask[:, self.height * self.width :] = 1
 596             result *= 1 - ar_mask
 597             masked_inplace_autoregression(
 598                 model,
 599                 self.batch_size,
 600                 result,
 601                 ar_mask,
 602                 deterministic_synthesis,
 603                 progress_bar_desc=None,
 604                 device=self.device,
 605             )
 606             mazes, paths = self.seq2map(result)
 607             path_correctness = maze.path_correctness(mazes, paths)
 608             nb_correct += path_correctness.long().sum()
 609             nb_total += mazes.size(0)
 610
 611             optimal_path_lengths = (
 612                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 613             )
 614             predicted_path_lengths = (
 615                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 616             )
 617             optimal_path_lengths = optimal_path_lengths[path_correctness]
 618             predicted_path_lengths = predicted_path_lengths[path_correctness]
 619             count[optimal_path_lengths, predicted_path_lengths] += 1
 620
 621         if count.max() == 0:
 622             count = None
 623         else:
 624             count = count[
 625                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 626             ]
 627
 628         return nb_total, nb_correct, count
 629
 630     def produce_results(
 631         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 632     ):
 633         train_nb_total, train_nb_correct, count = self.compute_error(
 634             model,
 635             "train",
 636             nb_to_use=1000,
 637             deterministic_synthesis=deterministic_synthesis,
 638         )
 639         logger(
 640             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}%"
 641         )
 642
 643         test_nb_total, test_nb_correct, count = self.compute_error(
 644             model,
 645             "test",
 646             nb_to_use=1000,
 647             deterministic_synthesis=deterministic_synthesis,
 648         )
 649         logger(
 650             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}%"
 651         )
 652
 653         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 654
 655         if count is not None:
 656             proportion_optimal = count.diagonal().sum().float() / count.sum()
 657             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 658             with open(
 659                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 660             ) as f:
 661                 for i in range(count.size(0)):
 662                     for j in range(count.size(1)):
 663                         eol = " " if j < count.size(1) - 1 else "\n"
 664                         f.write(f"{count[i,j]}{eol}")
 665
 666         input = self.test_input[:48]
 667         result = input.clone()
 668         ar_mask = result.new_zeros(result.size())
 669         ar_mask[:, self.height * self.width :] = 1
 670         result *= 1 - ar_mask
 671         masked_inplace_autoregression(
 672             model,
 673             self.batch_size,
 674             result,
 675             ar_mask,
 676             deterministic_synthesis,
 677             device=self.device,
 678         )
 679
 680         mazes, paths = self.seq2map(input)
 681         _, predicted_paths = self.seq2map(result)
 682
 683         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 684         maze.save_image(
 685             filename,
 686             mazes=mazes,
 687             target_paths=paths,
 688             predicted_paths=predicted_paths,
 689             path_correct=maze.path_correctness(mazes, predicted_paths),
 690             path_optimal=maze.path_optimality(paths, predicted_paths),
 691         )
 692         logger(f"wrote {filename}")
 693
 694
 695 ######################################################################
 696
 697
 698 import snake
 699
 700
 701 class Snake(Task):
 702     def __init__(
 703         self,
 704         nb_train_samples,
 705         nb_test_samples,
 706         batch_size,
 707         height,
 708         width,
 709         nb_colors,
 710         length,
 711         prompt_length,
 712         device=torch.device("cpu"),
 713     ):
 714         super().__init__()
 715
 716         self.batch_size = batch_size
 717         self.height = height
 718         self.width = width
 719         self.device = device
 720         self.prompt_length = prompt_length
 721
 722         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 723             nb_train_samples,
 724             height,
 725             width,
 726             nb_colors,
 727             length,
 728             prompt_length,
 729             self.device,
 730         )
 731         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 732             nb_test_samples,
 733             height,
 734             width,
 735             nb_colors,
 736             length,
 737             prompt_length,
 738             self.device,
 739         )
 740
 741         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 742
 743     def batches(self, split="train", nb_to_use=-1, desc=None):
 744         assert split in {"train", "test"}
 745         input = self.train_input if split == "train" else self.test_input
 746         if nb_to_use > 0:
 747             input = input[:nb_to_use]
 748         if desc is None:
 749             desc = f"epoch-{split}"
 750         for batch in tqdm.tqdm(
 751             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 752         ):
 753             yield batch
 754
 755     def vocabulary_size(self):
 756         return self.nb_codes
 757
 758     def produce_results(
 759         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 760     ):
 761         def compute_nb_correct(input, prior_visits):
 762             result = input.clone()
 763             i = torch.arange(result.size(1), device=result.device)[None, :]
 764             ar_mask = (
 765                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 766                 .long()
 767                 .expand_as(result)
 768             )
 769             result *= 1 - ar_mask
 770
 771             masked_inplace_autoregression(
 772                 model,
 773                 self.batch_size,
 774                 result,
 775                 ar_mask,
 776                 deterministic_synthesis,
 777                 device=self.device,
 778             )
 779
 780             nb_total = ((prior_visits > 0) * ar_mask).sum()
 781
 782             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 783
 784             return nb_total, nb_correct
 785
 786         test_nb_total, test_nb_correct = compute_nb_correct(
 787             self.test_input[:1000], self.test_prior_visits[:1000]
 788         )
 789
 790         logger(
 791             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}%"
 792         )
 793
 794         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 795
 796
 797 ######################################################################
 798
 799
 800 import stack
 801
 802
 803 class Stack(Task):
 804     def __init__(
 805         self,
 806         nb_train_samples,
 807         nb_test_samples,
 808         batch_size,
 809         logger,
 810         nb_steps,
 811         nb_stacks,
 812         nb_digits,
 813         fraction_values_for_train=None,
 814         device=torch.device("cpu"),
 815     ):
 816         super().__init__()
 817
 818         self.batch_size = batch_size
 819         self.nb_steps = nb_steps
 820         self.nb_stacks = nb_stacks
 821         self.nb_digits = nb_digits
 822         self.device = device
 823
 824         if fraction_values_for_train is None:
 825             values_for_train = None
 826             values_for_test = None
 827         else:
 828             all = torch.randperm(10**nb_digits)
 829             nb_for_train = int(all.size(0) * fraction_values_for_train)
 830             values_for_train = all[:nb_for_train]
 831             values_for_test = all[nb_for_train:]
 832
 833         self.train_input, self.train_stack_counts = stack.generate_sequences(
 834             nb_train_samples,
 835             nb_steps,
 836             nb_stacks,
 837             nb_digits,
 838             values_for_train,
 839             self.device,
 840         )
 841
 842         self.test_input, self.test_stack_counts = stack.generate_sequences(
 843             nb_test_samples,
 844             nb_steps,
 845             nb_stacks,
 846             nb_digits,
 847             values_for_test,
 848             self.device,
 849         )
 850
 851         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 852         counts = self.test_stack_counts.flatten()[i.flatten()]
 853         counts = F.one_hot(counts).sum(0)
 854         logger(f"test_pop_stack_counts {counts}")
 855
 856         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 857
 858     def batches(self, split="train", nb_to_use=-1, desc=None):
 859         assert split in {"train", "test"}
 860         input = self.train_input if split == "train" else self.test_input
 861         if nb_to_use > 0:
 862             input = input[:nb_to_use]
 863         if desc is None:
 864             desc = f"epoch-{split}"
 865         for batch in tqdm.tqdm(
 866             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 867         ):
 868             yield batch
 869
 870     def vocabulary_size(self):
 871         return self.nb_codes
 872
 873     def produce_results(
 874         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 875     ):
 876         def compute_nb_correct(input):
 877             result = input.clone()
 878             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 879             ar_mask = (result != input).long()
 880             masked_inplace_autoregression(
 881                 model,
 882                 self.batch_size,
 883                 result,
 884                 ar_mask,
 885                 deterministic_synthesis,
 886                 device=self.device,
 887             )
 888
 889             errors = ((result != input).long() * ar_mask).reshape(
 890                 -1, 1 + self.nb_digits
 891             )
 892             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 893
 894             nb_total = ar_mask.max(1).values.sum()
 895             nb_correct = nb_total - errors.max(1).values.sum()
 896
 897             return nb_total, nb_correct
 898
 899         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 900
 901         logger(
 902             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}%"
 903         )
 904
 905         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 906
 907         ##############################################################
 908         # Log a few generated sequences
 909         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
 910         result = input.clone()
 911         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 912         ar_mask = (result != input).long()
 913
 914         # for n in range(result.size(0)):
 915         # logger(
 916         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 917         # )
 918
 919         masked_inplace_autoregression(
 920             model,
 921             self.batch_size,
 922             result,
 923             ar_mask,
 924             deterministic_synthesis,
 925             device=self.device,
 926         )
 927
 928         for n in range(result.size(0)):
 929             logger(
 930                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 931             )
 932         ##############################################################
 933
 934
 935 ######################################################################
 936
 937 import rpl
 938
 939
 940 class RPL(Task):
 941     def tensorize(self, sequences):
 942         len_max = max([len(x) for x in sequences])
 943         return torch.cat(
 944             [
 945                 torch.tensor(
 946                     [
 947                         [
 948                             self.token2id[str(c)]
 949                             for c in s + ["<nul>"] * (len_max - len(s))
 950                         ]
 951                         for s in sequences
 952                     ]
 953                 )
 954             ],
 955             0,
 956         )
 957
 958     def seq2str(self, seq):
 959         return " ".join([self.id2token[i] for i in seq])
 960
 961     def __init__(
 962         self,
 963         nb_train_samples,
 964         nb_test_samples,
 965         batch_size,
 966         nb_starting_values=3,
 967         max_input=9,
 968         prog_len=6,
 969         nb_runs=5,
 970         no_prog=False,
 971         logger=None,
 972         device=torch.device("cpu"),
 973     ):
 974         super().__init__()
 975
 976         self.batch_size = batch_size
 977         self.device = device
 978         self.no_prog = no_prog
 979
 980         train_sequences = [
 981             rpl.generate(
 982                 nb_starting_values=nb_starting_values,
 983                 nb_result_values_max=4 * nb_starting_values,
 984                 max_input=max_input,
 985                 prog_len=prog_len,
 986                 nb_runs=nb_runs,
 987             )
 988             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
 989         ]
 990
 991         test_sequences = [
 992             rpl.generate(
 993                 nb_starting_values=nb_starting_values,
 994                 nb_result_values_max=4 * nb_starting_values,
 995                 max_input=max_input,
 996                 prog_len=prog_len,
 997                 nb_runs=nb_runs,
 998             )
 999             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
1000         ]
1001
1002         symbols = list(
1003             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
1004         )
1005         val_max = max([x if type(x) is int else 0 for x in symbols])
1006         symbols = list(filter(lambda x: type(x) is str, symbols))
1007         symbols.sort()
1008         symbols += [str(n) for n in range(val_max + 1)]
1009         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
1010         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
1011
1012         self.t_nul = self.token2id["<nul>"]
1013         self.t_input = self.token2id["<in>"]
1014         self.t_output = self.token2id["<out>"]
1015         self.t_prog = self.token2id["<prg>"]
1016         self.t_end = self.token2id["<end>"]
1017
1018         self.train_input = self.tensorize(train_sequences)
1019         self.test_input = self.tensorize(test_sequences)
1020
1021         if no_prog:
1022             # Excise the program from every train and test example
1023             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
1024                 None, :
1025             ]
1026             p = (
1027                 ((self.train_input == self.t_prog).long() * k)
1028                 .max(1, keepdim=True)
1029                 .values
1030             )
1031             self.train_input = (
1032                 self.train_input * (k <= p).long()
1033                 + self.t_end * (k == p + 1).long()
1034                 + self.t_nul * (k > p + 1).long()
1035             )
1036             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
1037                 None, :
1038             ]
1039             p = (
1040                 ((self.test_input == self.t_prog).long() * k)
1041                 .max(1, keepdim=True)
1042                 .values
1043             )
1044             self.test_input = (
1045                 self.test_input * (k <= p).long()
1046                 + self.t_end * (k == p + 1).long()
1047                 + self.t_nul * (k > p + 1).long()
1048             )
1049
1050         if logger is not None:
1051             logger(f"value_max {val_max}")
1052             for x in self.train_input[:25]:
1053                 end = (x != self.t_nul).nonzero().max().item() + 1
1054                 seq = [self.id2token[i.item()] for i in x[:end]]
1055                 s = " ".join(seq)
1056                 logger(f"example_seq {s}")
1057
1058         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1059
1060     def batches(self, split="train", nb_to_use=-1, desc=None):
1061         assert split in {"train", "test"}
1062         input = self.train_input if split == "train" else self.test_input
1063         if nb_to_use > 0:
1064             input = input[:nb_to_use]
1065         if desc is None:
1066             desc = f"epoch-{split}"
1067         for batch in tqdm.tqdm(
1068             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1069         ):
1070             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1071             batch = batch[:, :last].to(self.device)
1072             yield batch
1073
1074     def vocabulary_size(self):
1075         return self.nb_codes
1076
1077     def produce_results(
1078         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1079     ):
1080         # --------------------------------------------------------------------
1081         def compute_nb_errors_prog(input, nb_to_log=0):
1082             result = input.clone()
1083             s = (result == self.t_prog).long()
1084             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1085             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1086
1087             masked_inplace_autoregression(
1088                 model,
1089                 self.batch_size,
1090                 result,
1091                 ar_mask,
1092                 deterministic_synthesis,
1093                 device=self.device,
1094             )
1095
1096             sum_nb_total, sum_nb_errors = 0, 0
1097             for one_input, one_result in zip(input, result):
1098                 seq = [self.id2token[i.item()] for i in one_result]
1099                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1100                 sum_nb_total += 1
1101                 sum_nb_errors += 0 if nb_errors == 0 else 1
1102                 if nb_to_log > 0:
1103                     gt_seq = [self.id2token[i.item()] for i in one_input]
1104                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1105                     gt_prog = " ".join([str(x) for x in gt_prog])
1106                     prog = " ".join([str(x) for x in prog])
1107                     comment = "*" if nb_errors == 0 else "-"
1108                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1109                     for start_stack, target_stack, result_stack, correct in stacks:
1110                         comment = "*" if correct else "-"
1111                         start_stack = " ".join([str(x) for x in start_stack])
1112                         target_stack = " ".join([str(x) for x in target_stack])
1113                         result_stack = " ".join([str(x) for x in result_stack])
1114                         logger(
1115                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1116                         )
1117                     nb_to_log -= 1
1118
1119             return sum_nb_total, sum_nb_errors
1120
1121         # --------------------------------------------------------------------
1122         def compute_nb_errors_output(input, nb_to_log=0):
1123             result = input.clone()
1124             k = torch.arange(result.size(1), device=result.device)[None, :]
1125             last_output_idx = (
1126                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1127             )
1128             first_prog_idx = (
1129                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1130             )
1131             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1132             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1133
1134             masked_inplace_autoregression(
1135                 model,
1136                 self.batch_size,
1137                 result,
1138                 ar_mask,
1139                 deterministic_synthesis,
1140                 device=self.device,
1141             )
1142
1143             sum_nb_total, sum_nb_errors = 0, 0
1144             for one_input, one_result, i, j in zip(
1145                 input, result, last_output_idx, first_prog_idx
1146             ):
1147                 seq = [self.id2token[i.item()] for i in one_result]
1148                 sum_nb_total += 1
1149                 correct = (one_input - one_result).abs().max() == 0
1150                 sum_nb_errors += 0 if correct else 1
1151                 if nb_to_log > 0:
1152                     result_stack = [
1153                         self.id2token[i.item()] for i in one_result[i : j + 1]
1154                     ]
1155                     target_stack = [
1156                         self.id2token[i.item()] for i in one_input[i : j + 1]
1157                     ]
1158                     comment = "*" if correct else "-"
1159                     result_stack = " ".join([str(x) for x in result_stack])
1160                     target_stack = " ".join([str(x) for x in target_stack])
1161                     logger(
1162                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1163                     )
1164                     nb_to_log -= 1
1165
1166             return sum_nb_total, sum_nb_errors
1167
1168         # --------------------------------------------------------------------
1169
1170         if not self.no_prog:
1171             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1172                 self.test_input[:1000].to(self.device), nb_to_log=10
1173             )
1174
1175             logger(
1176                 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}%"
1177             )
1178
1179             logger(f"main_test_accuracy {n_epoch} {1-test_nb_errors/test_nb_total}")
1180
1181         test_nb_total, test_nb_errors = compute_nb_errors_output(
1182             self.test_input[:1000].to(self.device), nb_to_log=10
1183         )
1184
1185         logger(
1186             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}%"
1187         )
1188
1189         if save_attention_image is None:
1190             logger("no save_attention_image (is pycairo installed?)")
1191         else:
1192             ns = torch.randint(self.test_input.size(0), (1,)).item()
1193             input = self.test_input[ns : ns + 1].clone()
1194             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1195             input = input[:, :last].to(self.device)
1196
1197             with torch.autograd.no_grad():
1198                 t = model.training
1199                 model.eval()
1200                 model.record_attention(True)
1201                 model(BracketedSequence(input))
1202                 model.train(t)
1203                 ram = model.retrieve_attention()
1204                 model.record_attention(False)
1205
1206             tokens_output = [self.id2token[i.item()] for i in input[0]]
1207             tokens_input = ["n/a"] + tokens_output[:-1]
1208             for n_head in range(ram[0].size(1)):
1209                 filename = os.path.join(
1210                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1211                 )
1212                 attention_matrices = [m[0, n_head] for m in ram]
1213                 save_attention_image(
1214                     filename,
1215                     tokens_input,
1216                     tokens_output,
1217                     attention_matrices,
1218                     k_top=10,
1219                     # min_total_attention=0.9,
1220                     token_gap=12,
1221                     layer_gap=50,
1222                 )
1223                 logger(f"wrote {filename}")
1224
1225
1226 ######################################################################
1227
1228
1229 import expr
1230
1231
1232 class Expr(Task):
1233     def tensorize(self, sequences):
1234         len_max = max([len(x) for x in sequences])
1235         return torch.cat(
1236             [
1237                 torch.tensor(
1238                     [
1239                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1240                         for s in sequences
1241                     ]
1242                 )
1243             ],
1244             0,
1245         ).to(self.device)
1246
1247     def __init__(
1248         self,
1249         nb_train_samples,
1250         nb_test_samples,
1251         nb_variables,
1252         sequence_length,
1253         operand_max,
1254         result_max,
1255         batch_size,
1256         device=torch.device("cpu"),
1257     ):
1258         super().__init__()
1259
1260         self.batch_size = batch_size
1261         self.device = device
1262
1263         train_sequences = expr.generate_sequences(
1264             nb_train_samples,
1265             nb_variables=nb_variables,
1266             length=sequence_length,
1267             operand_max=operand_max,
1268             result_max=result_max,
1269         )
1270
1271         test_sequences = expr.generate_sequences(
1272             nb_test_samples,
1273             nb_variables=nb_variables,
1274             length=sequence_length,
1275             operand_max=operand_max,
1276             result_max=result_max,
1277         )
1278
1279         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1280         symbols.sort()
1281
1282         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1283         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1284
1285         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1286
1287         self.train_input = self.tensorize(train_sequences)
1288         self.test_input = self.tensorize(test_sequences)
1289
1290         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1291
1292     def batches(self, split="train", nb_to_use=-1, desc=None):
1293         assert split in {"train", "test"}
1294         input = self.train_input if split == "train" else self.test_input
1295         if nb_to_use > 0:
1296             input = input[:nb_to_use]
1297         if desc is None:
1298             desc = f"epoch-{split}"
1299         for batch in tqdm.tqdm(
1300             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1301         ):
1302             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1303             batch = batch[:, :last]
1304             yield batch
1305
1306     def vocabulary_size(self):
1307         return self.nb_codes
1308
1309     def seq2str(self, s):
1310         return "".join([self.id2char[k.item()] for k in s])
1311
1312     def produce_results(
1313         self,
1314         n_epoch,
1315         model,
1316         result_dir,
1317         logger,
1318         deterministic_synthesis,
1319         input_file=None,
1320     ):
1321         def compute_nb_correct(input):
1322             result = input.clone()
1323             s = (result == self.space).long()
1324             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1325             result = (1 - ar_mask) * result + ar_mask * self.filler
1326             masked_inplace_autoregression(
1327                 model,
1328                 self.batch_size,
1329                 result,
1330                 ar_mask,
1331                 deterministic_synthesis,
1332                 device=self.device,
1333             )
1334
1335             nb_total = input.size(0)
1336             nb_correct = (input == result).long().min(1).values.sum()
1337
1338             #######################################################################
1339             # Comput predicted vs. true variable values
1340
1341             nb_delta = torch.zeros(5, dtype=torch.int64)
1342             nb_missed = 0
1343
1344             values_input = expr.extract_results([self.seq2str(s) for s in input])
1345             values_result = expr.extract_results([self.seq2str(s) for s in result])
1346
1347             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1348
1349             with open(filename, "w") as f:
1350                 for i, r in zip(values_input, values_result):
1351                     for n, vi in i.items():
1352                         vr = r.get(n)
1353                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1354
1355                         if vr is None or vr < 0:
1356                             nb_missed += 1
1357                         else:
1358                             d = abs(vr - vi)
1359                             if d >= nb_delta.size(0):
1360                                 nb_missed += 1
1361                             else:
1362                                 nb_delta[d] += 1
1363
1364             ######################################################################
1365
1366             return nb_total, nb_correct, nb_delta, nb_missed
1367
1368         (
1369             test_nb_total,
1370             test_nb_correct,
1371             test_nb_delta,
1372             test_nb_missed,
1373         ) = compute_nb_correct(self.test_input[:10000])
1374
1375         logger(
1376             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}%"
1377         )
1378
1379         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
1380
1381         nb_total = test_nb_delta.sum() + test_nb_missed
1382         for d in range(test_nb_delta.size(0)):
1383             logger(
1384                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1385             )
1386         logger(
1387             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1388         )
1389
1390         ##############################################################
1391         # Log a few generated sequences
1392         if input_file is None:
1393             input = self.test_input[:10]
1394         else:
1395             with open(input_file, "r") as f:
1396                 sequences = [e.strip() for e in f.readlines()]
1397                 sequences = [s + " " + "#" * 50 for s in sequences]
1398                 input = self.tensorize(sequences)
1399
1400         result = input.clone()
1401         s = (result == self.space).long()
1402         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1403         result = (1 - ar_mask) * result + ar_mask * self.filler
1404
1405         for n in range(result.size(0)):
1406             logger(f"test_before {self.seq2str(result[n])}")
1407
1408         masked_inplace_autoregression(
1409             model,
1410             self.batch_size,
1411             result,
1412             ar_mask,
1413             deterministic_synthesis,
1414             device=self.device,
1415         )
1416
1417         correct = (1 - ar_mask) * self.space + ar_mask * input
1418         for n in range(result.size(0)):
1419             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1420             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1421             logger(f"truth       {self.seq2str(correct[n])}")
1422         ##############################################################
1423
1424
1425 ######################################################################
1426
1427 import grid
1428
1429
1430 class Grid(Task):
1431     # Make a tensor from a list of strings
1432     def str2tensor(self, descr):
1433         token_descr = [s.strip().split(" ") for s in descr]
1434         l = max([len(s) for s in token_descr])
1435         token_descr = [s + ["#"] * (l - len(s)) for s in token_descr]
1436         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
1437         return torch.tensor(id_descr, device=self.device)
1438
1439     # Make a list of strings from a tensor
1440     def tensor2str(self, x):
1441         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
1442
1443     # trim all the tensors in the tuple z to remove as much token from
1444     # left and right in the first tensor. If z is a tuple, all its
1445     # elements are trimed according to the triming for the first
1446     def trim(self, z, token="#"):
1447         n = self.token2id[token]
1448         if type(z) == tuple:
1449             x = z[0]
1450             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
1451             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
1452             return tuple([t[:, a:b] for t in z])
1453         else:
1454             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
1455             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
1456             return z[:, a:b]
1457
1458     ######################
1459
1460     def __init__(
1461         self,
1462         nb_train_samples,
1463         nb_test_samples,
1464         batch_size,
1465         size,
1466         logger=None,
1467         device=torch.device("cpu"),
1468     ):
1469         super().__init__()
1470
1471         self.device = device
1472         self.batch_size = batch_size
1473         self.grid_factory = grid.GridFactory(size=size)
1474
1475         if logger is not None:
1476             logger(
1477                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
1478             )
1479
1480         self.train_descr = self.grid_factory.generate_samples(
1481             nb_train_samples, lambda r: tqdm.tqdm(r)
1482         )
1483         self.test_descr = self.grid_factory.generate_samples(
1484             nb_test_samples, lambda r: tqdm.tqdm(r)
1485         )
1486
1487         # Build the tokenizer
1488         tokens = set()
1489         for d in [self.train_descr, self.test_descr]:
1490             for s in d:
1491                 for t in s.strip().split(" "):
1492                     tokens.add(t)
1493         # make this set a sorted list to get the same tensors given
1494         # the same descr
1495         tokens = list(tokens)
1496         tokens.sort()
1497         tokens = ["#"] + tokens
1498         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
1499         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
1500         self.t_nul = self.token2id["#"]
1501         self.t_true = self.token2id["true"]
1502         self.t_false = self.token2id["false"]
1503
1504         # Tokenize the train and test sets
1505         self.train_input = self.str2tensor(self.train_descr)
1506         self.test_input = self.str2tensor(self.test_descr)
1507
1508     def batches(self, split="train"):
1509         assert split in {"train", "test"}
1510         input = self.train_input if split == "train" else self.test_input
1511         for batch in tqdm.tqdm(
1512             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
1513         ):
1514             yield self.trim(batch)
1515
1516     def vocabulary_size(self):
1517         return len(self.token2id)
1518
1519     def produce_results(
1520         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1521     ):
1522         correct = self.test_input[:1000]
1523         result = correct.clone()
1524         ar_mask = torch.logical_or(result == self.t_true, result == self.t_false).long()
1525         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
1526
1527         logger(f"----------------------------------------------------------")
1528
1529         for e in self.tensor2str(result[:10]):
1530             logger(f"test_before {e}")
1531
1532         masked_inplace_autoregression(
1533             model,
1534             self.batch_size,
1535             result,
1536             ar_mask,
1537             deterministic_synthesis,
1538             device=self.device,
1539         )
1540
1541         logger(f"----------------------------------------------------------")
1542
1543         for e in self.tensor2str(result[:10]):
1544             logger(f"test_after  {e}")
1545
1546         logger(f"----------------------------------------------------------")
1547
1548         nb_total = ar_mask.sum().item()
1549         nb_correct = ((correct == result).long() * ar_mask).sum().item()
1550
1551         logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
1552         logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
1553
1554
1555 ######################################################################
1556
1557 import qmlp
1558
1559
1560 class QMLP(Task):
1561     ######################
1562
1563     def __init__(
1564         self,
1565         nb_train_samples,
1566         nb_test_samples,
1567         batch_size,
1568         result_dir,
1569         logger=None,
1570         device=torch.device("cpu"),
1571     ):
1572         super().__init__()
1573
1574         self.device = device
1575         self.batch_size = batch_size
1576         self.nb_samples_per_mlp = 256
1577
1578         if logger is not None:
1579             logger(
1580                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
1581             )
1582
1583         seq, q_test_set, test_error = qmlp.generate_sequence_and_test_set(
1584             nb_mlps=nb_train_samples + nb_test_samples,
1585             nb_samples=self.nb_samples_per_mlp,
1586             device=self.device,
1587             batch_size=64,
1588             nb_epochs=250,
1589             nb_mlps_per_batch=1024,
1590         )
1591
1592         self.train_input = seq[:nb_train_samples]
1593         self.train_q_test_set = q_test_set[:nb_train_samples]
1594         self.train_ref_test_errors = test_error[:nb_train_samples]
1595         self.test_input = seq[nb_train_samples:]
1596         self.test_q_test_set = q_test_set[nb_train_samples:]
1597         self.test_ref_test_errors = test_error[nb_train_samples:]
1598
1599         filename = os.path.join(result_dir, f"train_errors_ref.dat")
1600         with open(filename, "w") as f:
1601             for e in self.train_ref_test_errors:
1602                 f.write(f"{e}\n")
1603
1604         filename = os.path.join(result_dir, f"test_errors_ref.dat")
1605         with open(filename, "w") as f:
1606             for e in self.test_ref_test_errors:
1607                 f.write(f"{e}\n")
1608
1609         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1610
1611     def batches(self, split="train"):
1612         assert split in {"train", "test"}
1613         input = self.train_input if split == "train" else self.test_input
1614         for batch in tqdm.tqdm(
1615             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
1616         ):
1617             yield batch
1618
1619     def vocabulary_size(self):
1620         return self.nb_codes
1621
1622     def produce_results(
1623         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1624     ):
1625         correct = self.test_input[:1000]
1626         result = correct.clone()
1627         ar_mask = (
1628             torch.arange(result.size(1), device=result.device)
1629             > self.nb_samples_per_mlp * 3 + 1
1630         ).long()[None, :]
1631         ar_mask = ar_mask.expand_as(result)
1632         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
1633
1634         masked_inplace_autoregression(
1635             model,
1636             self.batch_size,
1637             result,
1638             ar_mask,
1639             deterministic_synthesis,
1640             device=self.device,
1641         )
1642
1643         q_train_set = result[:, : self.nb_samples_per_mlp * 3]
1644         q_params = result[:, self.nb_samples_per_mlp * 3 + 1 :]
1645         error_test = qmlp.evaluate_q_params(q_params, self.test_q_test_set)
1646
1647         filename = os.path.join(result_dir, f"test_errors_{n_epoch:04d}.dat")
1648         with open(filename, "w") as f:
1649             for e in error_test:
1650                 f.write(f"{e}\n")
1651
1652
1653 ######################################################################