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 from multiprocessing import Pool, cpu_count
  26
  27 from torch import Tensor
  28 from torch.autograd import Variable
  29
  30 import svrt
  31
  32 ######################################################################
  33
  34 def generate_one_batch(s):
  35     svrt.seed(s)
  36     target = torch.LongTensor(self.batch_size).bernoulli_(0.5)
  37     input = svrt.generate_vignettes(problem_number, target)
  38     input = input.float().view(input.size(0), 1, input.size(1), input.size(2))
  39     if self.cuda:
  40         input = input.cuda()
  41         target = target.cuda()
  42     return [ input, target ]
  43
  44 class VignetteSet:
  45
  46     def __init__(self, problem_number, nb_batches, batch_size, cuda = False):
  47         self.cuda = cuda
  48         self.batch_size = batch_size
  49         self.problem_number = problem_number
  50         self.nb_batches = nb_batches
  51         self.nb_samples = self.nb_batches * self.batch_size
  52
  53         seed_list = torch.LongTensor(self.nb_batches).random_().tolist()
  54
  55         # self.data = []
  56         # for b in range(0, self.nb_batches):
  57             # self.data.append(generate_one_batch(seed_list[b]))
  58
  59         self.data = Pool(cpu_count()).map(generate_one_batch, seed_list)
  60
  61         acc = 0.0
  62         acc_sq = 0.0
  63         for b in range(0, self.nb_batches):
  64             input = self.data[b][0]
  65             acc += input.sum() / input.numel()
  66             acc_sq += input.pow(2).sum() /  input.numel()
  67
  68         mean = acc / self.nb_batches
  69         std = sqrt(acc_sq / self.nb_batches - mean * mean)
  70         for b in range(0, self.nb_batches):
  71             self.data[b][0].sub_(mean).div_(std)
  72
  73     def get_batch(self, b):
  74         return self.data[b]
  75
  76 ######################################################################
  77
  78 class CompressedVignetteSet:
  79     def __init__(self, problem_number, nb_batches, batch_size, cuda = False):
  80         self.cuda = cuda
  81         self.batch_size = batch_size
  82         self.problem_number = problem_number
  83         self.nb_batches = nb_batches
  84         self.nb_samples = self.nb_batches * self.batch_size
  85         self.targets = []
  86         self.input_storages = []
  87
  88         acc = 0.0
  89         acc_sq = 0.0
  90         for b in range(0, self.nb_batches):
  91             target = torch.LongTensor(self.batch_size).bernoulli_(0.5)
  92             input = svrt.generate_vignettes(problem_number, target)
  93             acc += input.float().sum() / input.numel()
  94             acc_sq += input.float().pow(2).sum() /  input.numel()
  95             self.targets.append(target)
  96             self.input_storages.append(svrt.compress(input.storage()))
  97
  98         self.mean = acc / self.nb_batches
  99         self.std = sqrt(acc_sq / self.nb_batches - self.mean * self.mean)
 100
 101     def get_batch(self, b):
 102         input = torch.ByteTensor(svrt.uncompress(self.input_storages[b])).float()
 103         input = input.view(self.batch_size, 1, 128, 128).sub_(self.mean).div_(self.std)
 104         target = self.targets[b]
 105
 106         if self.cuda:
 107             input = input.cuda()
 108             target = target.cuda()
 109
 110         return input, target
 111
 112 ######################################################################