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     problem_number, batch_size, cuda, random_seed = s
  36     svrt.seed(random_seed)
  37     target = torch.LongTensor(batch_size).bernoulli_(0.5)
  38     input = svrt.generate_vignettes(problem_number, target)
  39     input = input.float().view(input.size(0), 1, input.size(1), input.size(2))
  40     if cuda:
  41         input = input.cuda()
  42         target = target.cuda()
  43     return [ input, target ]
  44
  45 class VignetteSet:
  46
  47     def __init__(self, problem_number, nb_batches, batch_size, cuda = False):
  48         self.cuda = cuda
  49         self.batch_size = batch_size
  50         self.problem_number = problem_number
  51         self.nb_batches = nb_batches
  52         self.nb_samples = self.nb_batches * self.batch_size
  53
  54         seeds = torch.LongTensor(self.nb_batches).random_()
  55         mp_args = []
  56         for b in range(0, self.nb_batches):
  57             mp_args.append( [ problem_number, batch_size, cuda, seeds[b] ])
  58
  59         # self.data = []
  60         # for b in range(0, self.nb_batches):
  61             # self.data.append(generate_one_batch(mp_args[b]))
  62
  63         self.data = Pool(cpu_count()).map(generate_one_batch, mp_args)
  64
  65         acc = 0.0
  66         acc_sq = 0.0
  67         for b in range(0, self.nb_batches):
  68             input = self.data[b][0]
  69             acc += input.sum() / input.numel()
  70             acc_sq += input.pow(2).sum() /  input.numel()
  71
  72         mean = acc / self.nb_batches
  73         std = sqrt(acc_sq / self.nb_batches - mean * mean)
  74         for b in range(0, self.nb_batches):
  75             self.data[b][0].sub_(mean).div_(std)
  76
  77     def get_batch(self, b):
  78         return self.data[b]
  79
  80 ######################################################################
  81
  82 class CompressedVignetteSet:
  83     def __init__(self, problem_number, nb_batches, batch_size, cuda = False):
  84         self.cuda = cuda
  85         self.batch_size = batch_size
  86         self.problem_number = problem_number
  87         self.nb_batches = nb_batches
  88         self.nb_samples = self.nb_batches * self.batch_size
  89         self.targets = []
  90         self.input_storages = []
  91
  92         acc = 0.0
  93         acc_sq = 0.0
  94         for b in range(0, self.nb_batches):
  95             target = torch.LongTensor(self.batch_size).bernoulli_(0.5)
  96             input = svrt.generate_vignettes(problem_number, target)
  97             acc += input.float().sum() / input.numel()
  98             acc_sq += input.float().pow(2).sum() /  input.numel()
  99             self.targets.append(target)
 100             self.input_storages.append(svrt.compress(input.storage()))
 101
 102         self.mean = acc / self.nb_batches
 103         self.std = sqrt(acc_sq / self.nb_batches - self.mean * self.mean)
 104
 105     def get_batch(self, b):
 106         input = torch.ByteTensor(svrt.uncompress(self.input_storages[b])).float()
 107         input = input.view(self.batch_size, 1, 128, 128).sub_(self.mean).div_(self.std)
 108         target = self.targets[b]
 109
 110         if self.cuda:
 111             input = input.cuda()
 112             target = target.cuda()
 113
 114         return input, target
 115
 116 ######################################################################