假设我有一个2D数组X.
X.shape == (m, n)
我想在X上添加两个维度,同时沿着这些新轴复制值。即我想要
new_X.shape == (m, n, k, l)
并为所有我,j
new_X[i, j, :, :] = X[i, j]
实现这一目标的最佳方法是什么?
你可以简单地用np.broadcast_to
获得张量视图 -
np.broadcast_to(a[...,None,None],a.shape+(k,l)) # a is input array
好处是它没有额外的内存开销,因此实际上是免费的朗姆酒。
如果您需要具有自己的内存空间的数组输出,请附加.copy()
。
样品运行 -
In [9]: a = np.random.rand(2,3)
In [10]: k,l = 4,5
In [11]: np.broadcast_to(a[...,None,None],a.shape+(k,l)).shape
Out[11]: (2, 3, 4, 5)
# Verify memory space sharing
In [12]: np.shares_memory(a,np.broadcast_to(a[...,None,None],a.shape+(k,l)))
Out[12]: True
# Verify virtually free runtime
In [17]: a = np.random.rand(100,100)
...: k,l = 100,100
...: %timeit np.broadcast_to(a[...,None,None],a.shape+(k,l))
100000 loops, best of 3: 3.41 µs per loop