mutual_info_estimation.py 文件源码

def calc_all_sigams(data, sigmas):
    batchs = 128
    num_of_bins = 8
    # bins = np.linspace(-1, 1, num_of_bins).astype(np.float32)
    # bins = stats.mstats.mquantiles(np.squeeze(data.reshape(1, -1)), np.linspace(0,1, num=num_of_bins))
    # data = bins[np.digitize(np.squeeze(data.reshape(1, -1)), bins) - 1].reshape(len(data), -1)

    batch_points = np.rint(np.arange(0, data.shape[0] + 1, batchs)).astype(dtype=np.int32)
    I_XT = []
    num_of_rand = min(800, data.shape[1])
    for sigma in sigmas:
        # print sigma
        I_XT_temp = 0
        for i in range(0, len(batch_points) - 1):
            new_data = data[batch_points[i]:batch_points[i + 1], :]
            rand_indexs = np.random.randint(0, new_data.shape[1], num_of_rand)
            new_data = new_data[:, :]
            N = new_data.shape[0]
            d = new_data.shape[1]
            diff_mat = np.linalg.norm(((new_data[:, np.newaxis, :] - new_data)), axis=2)
            # print diff_mat.shape, new_data.shape
            s0 = 0.2
            # DOTO -add leaveoneout validation
            res = minimize(optimiaze_func, s0, args=(diff_mat, d, N), method='nelder-mead',
                           options={'xtol': 1e-8, 'disp': False, 'maxiter': 6})
            eta = res.x
            diff_mat0 = - 0.5 * (diff_mat / (sigma ** 2 + eta ** 2))
            diff_mat1 = np.sum(np.exp(diff_mat0), axis=0)
            diff_mat2 = -(1.0 / N) * np.sum(np.log2((1.0 / N) * diff_mat1))
            I_XT_temp += diff_mat2 - d * np.log2((sigma ** 2) / (eta ** 2 + sigma ** 2))
            # print diff_mat2 - d*np.log2((sigma**2)/(eta**2+sigma**2))
        I_XT_temp /= len(batch_points)
        I_XT.append(I_XT_temp)
    sys.stdout.flush()
    return I_XT