lra.py 文件源码

def low_rank_align(X, Y, Cxy, d=None, mu=0.8):
    """Input: data matrices X,Y,  correspondence matrix Cxy,
              embedding dimension d, and correspondence weight mu
       Output: embedded X and embedded Y
    """
    nx, dx = X.shape
    ny, dy = Y.shape
    assert Cxy.shape==(nx,ny), \
        'Correspondence matrix must be shape num_X_samples X num_Y_samples.'
    C = np.fliplr(block_diag(np.fliplr(Cxy),np.fliplr(Cxy.T)))
    if d is None:
        d = min(dx,dy)
    Rx = low_rank_repr(X,d)
    Ry = low_rank_repr(Y,d)
    R = block_diag(Rx,Ry)
    tmp = np.eye(R.shape[0]) - R
    M = tmp.T.dot(tmp)
    L = laplacian(C)
    eigen_prob = (1-mu)*M + 2*mu*L
    _,F = eigh(eigen_prob,eigvals=(1,d),overwrite_a=True)
    Xembed = F[:nx]
    Yembed = F[nx:]
    return Xembed, Yembed