svrtset.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 svrt.  If not, see <http://www.gnu.org/licenses/>.
  22
  23 import torch
  24 from math import sqrt
  25 from torch import multiprocessing
  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, 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     return [ input, target ]
  41
  42 class VignetteSet:
  43
  44     def __init__(self, problem_number, nb_samples, batch_size, cuda = False, logger = None):
  45
  46         if nb_samples%batch_size > 0:
  47             print('nb_samples must be a multiple of batch_size')
  48             raise
  49
  50         self.cuda = cuda
  51         self.problem_number = problem_number
  52
  53         self.batch_size = batch_size
  54         self.nb_samples = nb_samples
  55         self.nb_batches = self.nb_samples // self.batch_size
  56
  57         seeds = torch.LongTensor(self.nb_batches).random_()
  58         mp_args = []
  59         for b in range(0, self.nb_batches):
  60             mp_args.append( [ problem_number, batch_size, seeds[b] ])
  61
  62         self.data = []
  63         for b in range(0, self.nb_batches):
  64             self.data.append(generate_one_batch(mp_args[b]))
  65             if logger is not None: logger(self.nb_batches * self.batch_size, b * self.batch_size)
  66
  67         # Weird thing going on with the multi-processing, waiting for more info
  68
  69         # pool = multiprocessing.Pool(multiprocessing.cpu_count())
  70         # self.data = pool.map(generate_one_batch, mp_args)
  71
  72         acc = 0.0
  73         acc_sq = 0.0
  74         for b in range(0, self.nb_batches):
  75             input = self.data[b][0]
  76             acc += input.sum() / input.numel()
  77             acc_sq += input.pow(2).sum() /  input.numel()
  78
  79         mean = acc / self.nb_batches
  80         std = sqrt(acc_sq / self.nb_batches - mean * mean)
  81         for b in range(0, self.nb_batches):
  82             self.data[b][0].sub_(mean).div_(std)
  83             if cuda:
  84                 self.data[b][0] = self.data[b][0].cuda()
  85                 self.data[b][1] = self.data[b][1].cuda()
  86
  87     def get_batch(self, b):
  88         return self.data[b]
  89
  90 ######################################################################
  91
  92 class CompressedVignetteSet:
  93     def __init__(self, problem_number, nb_samples, batch_size, cuda = False, logger = None):
  94
  95         if nb_samples%batch_size > 0:
  96             print('nb_samples must be a multiple of batch_size')
  97             raise
  98
  99         self.cuda = cuda
 100         self.problem_number = problem_number
 101
 102         self.batch_size = batch_size
 103         self.nb_samples = nb_samples
 104         self.nb_batches = self.nb_samples // self.batch_size
 105
 106         self.targets = []
 107         self.input_storages = []
 108
 109         acc = 0.0
 110         acc_sq = 0.0
 111         for b in range(0, self.nb_batches):
 112             target = torch.LongTensor(self.batch_size).bernoulli_(0.5)
 113             input = svrt.generate_vignettes(problem_number, target)
 114             acc += input.float().sum() / input.numel()
 115             acc_sq += input.float().pow(2).sum() /  input.numel()
 116             self.targets.append(target)
 117             self.input_storages.append(svrt.compress(input.storage()))
 118             if logger is not None: logger(self.nb_batches * self.batch_size, b * self.batch_size)
 119
 120         self.mean = acc / self.nb_batches
 121         self.std = sqrt(acc_sq / self.nb_batches - self.mean * self.mean)
 122
 123     def get_batch(self, b):
 124         input = torch.ByteTensor(svrt.uncompress(self.input_storages[b])).float()
 125         input = input.view(self.batch_size, 1, 128, 128).sub_(self.mean).div_(self.std)
 126         target = self.targets[b]
 127
 128         if self.cuda:
 129             input = input.cuda()
 130             target = target.cuda()
 131
 132         return input, target
 133
 134 ######################################################################