def sample_grid_outside(self, nb_points = 1000, procentage = 0.1):
""" Similar to sample_grid, however this samples from a extension of the
image domain where procentage * image domain is added around the
original image domain
"""
x_ext = procentage * (self.maxbound[0] - self.minbound[0])
y_ext = procentage * (self.maxbound[1] - self.minbound[1])
return create_grid([self.minbound[0] - x_ext, self.minbound[1] - y_ext],
[self.maxbound[0] + x_ext, self.maxbound[1] + y_ext],
[nb_points, nb_points])
def get_transformer_init_weights(n_units, transformer_name='affine'):
""" Get weights that initialize a given transformer to the identity transformation """
dim = get_transformer_dim(transformer_name)
kernel = {'affine': np.zeros((n_units, dim), dtype=np.float32),
'affinediffeo': np.zeros((n_units, dim), dtype=np.float32),
'homografy': np.zeros((n_units, dim), dtype=np.float32),
'CPAB': np.zeros((n_units, dim), dtype=np.float32),
'TPS': np.zeros((n_units, dim), dtype=np.float32)}
bias = {'affine': np.array([1,0,0,0,1,0], dtype=np.float32),
'affinediffeo': np.zeros((dim,), dtype=np.float32),
'homografy': np.array([1,0,0,0,1,0,0,0,1], dtype=np.float32),
'CPAB': np.zeros((dim,), dtype=np.float32),
'TPS': create_grid([-1,-1],[1,1],[8,8]).T.flatten()}
return (kernel[transformer_name], bias[transformer_name])
def sample_grid_image(self, imagesize):
""" Similar to sample_grid, just with varing sample size in x,y direction """
return create_grid(self.minbound, self.maxbound, imagesize)
def sample_grid(self, nb_points = 1000):
""" Samples nb_points in both directions within the image domain and
returns a matrix of size (nb_points^2, 2), where each row is point
"""
return create_grid(self.minbound, self.maxbound, [nb_points, nb_points])
