Gramian
R
"""
Image style reconstruction with Gram matrices.
https://en.wikipedia.org/wiki/Gram_matrix
https://en.wikipedia.org/wiki/Neural_style_transfer
https://arxiv.org/pdf/1603.08155#page=7&zoom=auto,-294,3
"""
import numpy as np
def gram_matrix(mat: np.ndarray) -> np.ndarray:
"""
Returns the Gram (Gramian) matrix of an image.
:param mat: matrix of shape (C, H, W); C = color channels, H = height, W = width.
:type mat: np.ndarray
:return: matrix of shape (C, C).
:rtype: np.ndarray
Examples
--------
>>> gram_matrix(np.ones((2,5,5)))
array([[0.5, 0.5],
[0.5, 0.5]])
>>> gram_matrix(np.ones((3,5,5)))
array([[0.33333333, 0.33333333, 0.33333333],
[0.33333333, 0.33333333, 0.33333333],
[0.33333333, 0.33333333, 0.33333333]])
>>> gram_matrix(np.ones((3,5,5))).shape
(3, 3)
"""
color, height, width = mat.shape
vec = mat.reshape(color, height * width)
gram = vec @ vec.T
return gram / (color * height * width)
def gram_loss(input_features: np.ndarray, reference_features: np.ndarray) -> np.float64:
"""
Calculates the squared Frobenius norm of the difference between
the Gram matrices of the input and reference image.
:param input_features: Feature map of shape (C, H, W)
:type input_features: np.ndarray
:param reference_features: Feature map of shape (C, H, W)
:type reference_features: np.ndarray
:return: Gram loss between the two feature maps.
:rtype: float64
Examples
--------
>>> a = np.random.randn(3,5,5)
>>> gram_loss(a, a)
np.float64(0.0)
>>> a = np.zeros((3,5,5))
>>> b = np.ones((3,5,5))
>>> gram_loss(a, b)
np.float64(1.0)
"""
input_gram = gram_matrix(input_features)
reference_gram = gram_matrix(reference_features)
return np.sum(np.square(input_gram - reference_gram)).astype(np.float64)
if __name__ == "__main__":
import doctest
doctest.testmod()