def test_encode_spatial_bin_l2_max(self):
result = llc_opt.encode_spatial_bin_numba(self.codebook,
np.array([[3, 0], [-1, -1]]),
self.size,
self.alpha,
self.sigma,
pooling='max')
expected = np.array([1.11015, 0.22858, 0.98056])
self.assertArrayAlmostEqual(result, expected)
def test_encode_spatial_bin_l1_sum(self):
result = llc_opt.encode_spatial_bin_numba(self.codebook,
np.array([[3, 0], [-1, -1],
[1, -1]]),
self.size,
self.alpha,
self.sigma,
pooling='sum')
expected = np.array([1.25112, -0.260689, 2.00956])
self.assertArrayAlmostEqual(result, expected)
def _encode_spatial_bin(self, features, pooling='max'):
"""
Computes the LLC codes for a set of features and pools them with the
specified pooling method. In case of an empty set, a zero vector will be
returned.
In further versions it is planned to implement another version of
"encode_spatial_bin", preferably using Tensorflow or Pytorch and this
method can choose which to call based on set options.
:author: Joschka Strüber
"""
return encode_spatial_bin_numba(self._codebook, features, self._size,
self._alpha, self._sigma, pooling)
def test_encode_spatial_bin_empty(self):
result = llc_opt.encode_spatial_bin_numba(self.codebook, np.array([]),
self.size, self.alpha,
self.sigma)
expected = np.zeros(self.encoder._size)
self.assertTrue((result == expected).all())