vignette_set.py

   1
   2 #  svrt is the ``Synthetic Visual Reasoning Test'', an image
   3 #  generator for evaluating classification performance of machine
   4 #  learning systems, humans and primates.
   5 #
   6 #  Copyright (c) 2017 Idiap Research Institute, http://www.idiap.ch/
   7 #  Written by Francois Fleuret <francois.fleuret@idiap.ch>
   8 #
   9 #  This file is part of svrt.
  10 #
  11 #  svrt is free software: you can redistribute it and/or modify it
  12 #  under the terms of the GNU General Public License version 3 as
  13 #  published by the Free Software Foundation.
  14 #
  15 #  svrt is distributed in the hope that it will be useful, but
  16 #  WITHOUT ANY WARRANTY; without even the implied warranty of
  17 #  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18 #  General Public License for more details.
  19 #
  20 #  You should have received a copy of the GNU General Public License
  21 #  along with selector.  If not, see <http://www.gnu.org/licenses/>.
  22
  23 import torch
  24 from math import sqrt
  25
  26 from torch import Tensor
  27 from torch.autograd import Variable
  28
  29 import svrt
  30
  31 ######################################################################
  32
  33 class VignetteSet:
  34     def __init__(self, problem_number, nb_batches, batch_size):
  35         self.batch_size = batch_size
  36         self.problem_number = problem_number
  37         self.nb_batches = nb_batches
  38         self.nb_samples = self.nb_batches * self.batch_size
  39         self.targets = []
  40         self.inputs = []
  41
  42         acc = 0.0
  43         acc_sq = 0.0
  44
  45         for b in range(0, self.nb_batches):
  46             target = torch.LongTensor(self.batch_size).bernoulli_(0.5)
  47             input = svrt.generate_vignettes(problem_number, target)
  48             input = input.float().view(input.size(0), 1, input.size(1), input.size(2))
  49             if torch.cuda.is_available():
  50                 input = input.cuda()
  51                 target = target.cuda()
  52             acc += input.sum() / input.numel()
  53             acc_sq += input.pow(2).sum() /  input.numel()
  54             self.targets.append(target)
  55             self.inputs.append(input)
  56
  57         mean = acc / self.nb_batches
  58         std = sqrt(acc_sq / self.nb_batches - mean * mean)
  59         for b in range(0, self.nb_batches):
  60             self.inputs[b].sub_(mean).div_(std)
  61
  62     def get_batch(self, b):
  63         return self.inputs[b], self.targets[b]
  64
  65 ######################################################################
  66
  67 class CompressedVignetteSet:
  68     def __init__(self, problem_number, nb_batches, batch_size):
  69         self.batch_size = batch_size
  70         self.problem_number = problem_number
  71         self.nb_batches = nb_batches
  72         self.nb_samples = self.nb_batches * self.batch_size
  73         self.targets = []
  74         self.input_storages = []
  75
  76         acc = 0.0
  77         acc_sq = 0.0
  78         for b in range(0, self.nb_batches):
  79             target = torch.LongTensor(self.batch_size).bernoulli_(0.5)
  80             input = svrt.generate_vignettes(problem_number, target)
  81             acc += input.float().sum() / input.numel()
  82             acc_sq += input.float().pow(2).sum() /  input.numel()
  83             self.targets.append(target)
  84             self.input_storages.append(svrt.compress(input.storage()))
  85
  86         self.mean = acc / self.nb_batches
  87         self.std = math.sqrt(acc_sq / self.nb_batches - self.mean * self.mean)
  88
  89     def get_batch(self, b):
  90         input = torch.ByteTensor(svrt.uncompress(self.input_storages[b])).float()
  91         input = input.view(self.batch_size, 1, 128, 128).sub_(self.mean).div_(self.std)
  92         target = self.targets[b]
  93
  94         if torch.cuda.is_available():
  95             input = input.cuda()
  96             target = target.cuda()
  97
  98         return input, target
  99
 100 ######################################################################