135a56e0935fbb04811f8ce7b9f514f498212f71,niftynet/layer/crf.py,,ftheta,#Any#Any#Any#Any#Any#Any#,148

Before Change

                U.shape)

    // Weighting Filter Outputs
    Q2 = tf.add_n([Q1 * w for Q1, w in zip(Q1, kernel_weights)])

    // Compatibility Transform
    spatial_dim = infer_spatial_rank(U)
    assert spatial_dim == 2 or 3, \
        "Currently CRFAsRNNLayer supports 2D/3D images."
    full_stride = expand_spatial_params(1, spatial_dim)
    Q3 = tf.nn.convolution(input=Q2,
                           filter=mu,
                           strides=full_stride,
                           padding="SAME")
    // Adding Unary Potentials
    Q4 = U - Q3
    // output logits, not the softmax
    return Q4


def permutohedral_prepare(position_vectors):
    

After Change

    :return: updated mean-field distribution
    
    batch_size, n_ch = U.shape.as_list()[0], U.shape.as_list()[-1]
    n_voxels = np.prod(U.shape.as_list()[:-1])

    H1 = tf.reshape(tf.nn.softmax(H1), [batch_size, -1, n_ch])
    Q1 = 0
    for idx, permutohedral in enumerate(permutohedrals):

Instances