test_clustering_loss.py 文件源码

def check_forward(self, x_data, c_data, gamma, T, y_star, y_pam):
        num_examples = len(x_data)
        x = chainer.Variable(x_data)
        c = chainer.Variable(c_data)

        loss = clustering_loss(x, c, gamma, T)

        sq_distances_ij = []
        for i, j in zip(range(num_examples), y_pam):
            sqd_ij = np.sum((x_data[i] - x_data[j]) ** 2)
            sq_distances_ij.append(sqd_ij)
        f = -sum(sq_distances_ij)

        sq_distances_ij = []
        for i, j in zip(range(num_examples), y_star):
            sqd_ij = np.sum((x_data[i] - x_data[j]) ** 2)
            sq_distances_ij.append(sqd_ij)
        f_tilde = -sum(sq_distances_ij)

        delta = 1.0 - normalized_mutual_info_score(cuda.to_cpu(c_data), y_pam)
        loss_expected = f + gamma * delta - f_tilde

        testing.assert_allclose(loss.data, loss_expected)