semi_nmf.py 文件源码

def semi_nmf(x, iter = 30):
    '''
    Semi Nonnegative Matrix Factorization.
    It returns a feature matrix F and a representation matrix G by minimizing
    frobenius norm ||X - FG^T||^2. The only contraint is that elements in G to be positive.

    Args:
        x: input matrix X
        int iter: number of iterations of optimization algorithm

    Return:
        f: feature matrix F
        g: representation matrix G
    '''

    x = x.numpy()   # n * m
    f, g, p = svd_initialization(x)

    if  < 2:
        raise ValueError("The number of components (r) has to be >=2.")


    for i in range(iter):

        f = np.dot(x, np.dot(g, la.pinv(np.dot(g.T, g))))

        f = np.nan_to_num(f)

        Ap = (abs(np.dot(x.T, f)) + np.dot(x.T, f))/2   #m * r
        An = (abs(np.dot(x.T, f)) - np.dot(x.T, f))/2
        Bp = (abs(np.dot(g, np.dot(f.T, f))) + np.dot(g, np.dot(f.T, f)))/2
        Bn = (abs(np.dot(g, np.dot(f.T, f))) - np.dot(g, np.dot(f.T, f)))/2

        C = An + Bp
        for m in range(C.shape[0]):
            for n in range(C.shape[1]):
                if C[m, n] is 0:
                    C[m, n] += 0.0001

        for j in range(g.shape[0]):
            for k in range(g.shape[1]):
                g[j, k] = g[j, k] * np.sqrt( (Ap+Bn)[j,k]/(An+Bp)[j,k] )

    g = np.nan_to_num(g)

    return torch.from_numpy(f), torch.from_numpy(g)