张量流性能:Numpy矩阵o TF矩阵?

我有以下代码:

with tf.Session() as sess:
    sess.run(init_vars)
    cols = sess.run(tf.shape(descriptors)[1])
    descriptor_matrix = np.zeros((n_batches*batch_size, cols))
    while True:
        batch_descriptor = sess.run(descriptors, feed_dict={dropout_prob: 1})
        descriptor_matrix[i:i+elements_in_batch] = np.array(batch_descriptor)

我混合了张量和numpy向量。这对性能有重要影响吗?为什么会这样?我应该用张量来代替吗?