对于 2D 情况，有一个方法可以做到:

For 2D case,there is a method to do it:

# a.shape (16L, 10L)
# idx.shape (16L,1)
idx = tf.stack([tf.range(tf.shape(idx)[0]),idx[:,0]],axis=-1)
b = tf.gather_nd(a,idx)

然而，对于 ND 情况，这种方法可能非常复杂

However,For ND case,this method maybe very complex