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