2 # svrt is the ``Synthetic Visual Reasoning Test'', an image
3 # generator for evaluating classification performance of machine
4 # learning systems, humans and primates.
6 # Copyright (c) 2017 Idiap Research Institute, http://www.idiap.ch/
7 # Written by Francois Fleuret <francois.fleuret@idiap.ch>
9 # This file is part of svrt.
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.
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.
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/>.
25 from multiprocessing import Pool, cpu_count
27 from torch import Tensor
28 from torch.autograd import Variable
32 ######################################################################
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))
42 target = target.cuda()
43 return [ input, target ]
47 def __init__(self, problem_number, nb_batches, batch_size, cuda = False):
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
54 seeds = torch.LongTensor(self.nb_batches).random_()
56 for b in range(0, self.nb_batches):
57 mp_args.append( [ problem_number, batch_size, cuda, seeds[b] ])
60 # for b in range(0, self.nb_batches):
61 # self.data.append(generate_one_batch(mp_args[b]))
63 self.data = Pool(cpu_count()).map(generate_one_batch, mp_args)
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()
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)
77 def get_batch(self, b):
80 ######################################################################
82 class CompressedVignetteSet:
83 def __init__(self, problem_number, nb_batches, batch_size, cuda = False):
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
90 self.input_storages = []
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()))
102 self.mean = acc / self.nb_batches
103 self.std = sqrt(acc_sq / self.nb_batches - self.mean * self.mean)
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]
112 target = target.cuda()
116 ######################################################################