# -*- coding: utf-8 -*-
# ELEKTRONN - Neural Network Toolkit
#
# Copyright (c) 2014 - now
# Max-Planck-Institute for Medical Research, Heidelberg, Germany
# Authors: Marius Killinger, Gregor Urban
"""
Supplementary functions to initialise CNN-params with gabor filters
"""
import numpy as np
from matplotlib import pyplot as plt
#def makeGabor(filter_angle, n_modes, size, offset):
# """
# filter_angle: in degree: 0 to 180
# n_modes = 1,2,3 etc.
# size: filter size
# offset: 0 to 180
# """
# sigma = 0.35 # 0.5
# freq = n_modes / np.pi *1.6 #n_modes / np.pi * 1.3 # 1.5
# assert filter_angle<360 and filter_angle>=0
# xMax = yMax = (size-1)/2+1 if size%2==1 else size/2+1
# xMin = yMin = -xMax+1
# assert yMax>0
# assert xMax>0
# sigmaX = sigma
# sigmaY = sigma
# theta = float(filter_angle)/180.0*np.pi
# m_offset = float(offset)/180.0*np.pi
# gabor = np.empty((int(xMax)-int(xMin), int(yMax)-int(yMin)), dtype=np.float32)
# #print shape(gabor)
# for x in range(int(xMin), int(xMax)):
# for y in range(int(yMin), int(yMax)):
# xPrime = float(x)/xMax*np.cos(theta)+float(y)/yMax*np.sin(theta)
# yPrime = -float(x)/xMax*np.sin(theta)+float(y)/yMax*np.cos(theta)
# gabor[x-int(xMin),y-int(yMin)] = np.exp(-0.5*((xPrime*xPrime)/(sigmaX*sigmaX)+(yPrime*yPrime)/(sigmaY*sigmaY)))*np.cos(2.0*np.pi*(freq*xPrime + m_offset))
#
# gabor = gabor - gabor.mean()
# gabor = gabor / np.square(gabor).sum()
# return gabor
[docs]def makeGabor(filter_angle, n_modes, size, offset):
"""
Parameters
----------
filter_angle: in degree: 0 to 180
n_modes = 1,2,3 etc.
size: filter size
offset: 0 to 180
"""
assert 360 > filter_angle >= 0
sigma = 0.35 # 0.5
freq = n_modes / np.pi * 1.6 #n_modes / np.pi * 1.3 # 1.5
theta = float(filter_angle) / 180.0 * np.pi
m_offset = float(offset) / 180.0 * np.pi
Max = (size - 1) / 2 if size % 2 == 1 else size / 2
X, Y = np.meshgrid(np.linspace(-Max, Max, size), np.linspace(-Max, Max, size))
X_ = X / (Max + 1) * np.cos(theta) + Y / (Max + 1) * np.sin(theta)
Y_ = -X / (Max + 1) * np.sin(theta) + Y / (Max + 1) * np.cos(theta)
gabor = np.exp(-0.5 * ((X_**2) + (Y_**2)) / (sigma**2)) * np.cos(2.0 * np.pi * (freq * Y_ + m_offset))
gabor = gabor - gabor.mean()
gabor = gabor / np.square(gabor).sum()
return gabor.astype(np.float32)
[docs]def makeGaborFilters(size, number):
"""
Use this to generate ``number`` first order
and ``number`` second order filters
"""
nd2 = number #int(float(number)*0.1)
nd3 = number #number-nd2
ret = []
for theta in np.arange(0, 360, (360.0 / nd2)):
ret.append(makeGabor(theta, 1, size, 45))
#ret.append( makeGabor(theta, 1 ,size, np.random.random()*360) )
for theta in np.arange(0, 180, (180.0 / nd3)):
ret.append(makeGabor(theta, 2, size, 0))
#ret.append( makeGabor(theta, 2 ,size, 45 + (np.random.random()-0.5)*8) )
return np.array(ret)
[docs]def blob(size):
"""
Return Gaussian blob filter
"""
x, y = np.meshgrid(np.linspace(-2, 2, size), np.linspace(-2, 2, size))
ret = np.exp(-0.5 * (np.square(x) + np.square(y)))
ret = ret / np.square(ret).sum()
return ret
if __name__ == '__main__':
from introspection import embedMatricesInGray
ga = makeGaborFilters(3, 12)
#ga = np.maximum(0, ga)
mat = embedMatricesInGray(ga, 1)
plt.imshow(mat, interpolation='none')
plt.gray()
ga = makeGaborFilters(4, 12)
mat2 = embedMatricesInGray(ga, 1)
plt.figure()
plt.imshow(mat2, interpolation='none')
plt.gray()
