stack与concat差异

以pytorch为例

例子

import torch

a = torch.randint(0, 10, (2, 3, 4))
b = torch.randint(0, 10, (2, 3, 4))
print(a.shape, a)
print(b.shape, b)
c = torch.stack([a, b], dim=0)
d = torch.stack([a, b], dim=1)
e = torch.stack([a, b], dim=2)
print(c.shape, c)
print(d.shape, d)
print(e.shape, e)

f = torch.cat([a, b], dim=0)
g = torch.cat([a, b], dim=1)
h = torch.cat([a, b], dim=2)
print(f.shape, f)
print(g.shape, g)
print(h.shape, h)

打印结果：

a:  torch.Size([2, 3, 4]) tensor([[[0, 6, 7, 9],
         [8, 9, 3, 9],
         [1, 5, 9, 0]],

        [[2, 6, 7, 2],
         [5, 7, 1, 2],
         [4, 1, 9, 1]]])
b:  torch.Size([2, 3, 4]) tensor([[[5, 7, 9, 0],
         [8, 5, 5, 9],
         [5, 2, 5, 3]],

        [[5, 7, 4, 1],
         [2, 9, 4, 3],
         [1, 4, 3, 2]]])
c:  torch.Size([2, 2, 3, 4]) tensor([[[[0, 6, 7, 9],
          [8, 9, 3, 9],
          [1, 5, 9, 0]],

         [[2, 6, 7, 2],
          [5, 7, 1, 2],
          [4, 1, 9, 1]]],


        [[[5, 7, 9, 0],
          [8, 5, 5, 9],
          [5, 2, 5, 3]],

         [[5, 7, 4, 1],
          [2, 9, 4, 3],
          [1, 4, 3, 2]]]])
d:  torch.Size([2, 2, 3, 4]) tensor([[[[0, 6, 7, 9],
          [8, 9, 3, 9],
          [1, 5, 9, 0]],

         [[5, 7, 9, 0],
          [8, 5, 5, 9],
          [5, 2, 5, 3]]],


        [[[2, 6, 7, 2],
          [5, 7, 1, 2],
          [4, 1, 9, 1]],

         [[5, 7, 4, 1],
          [2, 9, 4, 3],
          [1, 4, 3, 2]]]])
e torch.Size([2, 3, 2, 4]) tensor([[[[0, 6, 7, 9],
          [5, 7, 9, 0]],

         [[8, 9, 3, 9],
          [8, 5, 5, 9]],

         [[1, 5, 9, 0],
          [5, 2, 5, 3]]],


        [[[2, 6, 7, 2],
          [5, 7, 4, 1]],

         [[5, 7, 1, 2],
          [2, 9, 4, 3]],

         [[4, 1, 9, 1],
          [1, 4, 3, 2]]]])
f:  torch.Size([4, 3, 4]) tensor([[[0, 6, 7, 9],
         [8, 9, 3, 9],
         [1, 5, 9, 0]],

        [[2, 6, 7, 2],
         [5, 7, 1, 2],
         [4, 1, 9, 1]],

        [[5, 7, 9, 0],
         [8, 5, 5, 9],
         [5, 2, 5, 3]],

        [[5, 7, 4, 1],
         [2, 9, 4, 3],
         [1, 4, 3, 2]]])
g:  torch.Size([2, 6, 4]) tensor([[[0, 6, 7, 9],
         [8, 9, 3, 9],
         [1, 5, 9, 0],
         [5, 7, 9, 0],
         [8, 5, 5, 9],
         [5, 2, 5, 3]],

        [[2, 6, 7, 2],
         [5, 7, 1, 2],
         [4, 1, 9, 1],
         [5, 7, 4, 1],
         [2, 9, 4, 3],
         [1, 4, 3, 2]]])
h:  torch.Size([2, 3, 8]) tensor([[[0, 6, 7, 9, 5, 7, 9, 0],
         [8, 9, 3, 9, 8, 5, 5, 9],
         [1, 5, 9, 0, 5, 2, 5, 3]],

        [[2, 6, 7, 2, 5, 7, 4, 1],
         [5, 7, 1, 2, 2, 9, 4, 3],
         [4, 1, 9, 1, 1, 4, 3, 2]]])

总结

可以看出，concat是拼接tensor，维度不会扩张，但拼接的维度会变成维度之和，如：2个(2,3,4)的tensor在最后一个维度拼接，结果是(2,3,8)；stack是堆tensor，然后在该维度外层加一个括号，所以维度会在被堆起来的维度前扩张, 扩张为堆的维度数，如：2个(2,3,4)的tensor在最后一个维度stack，结果是(2,3,2,4)。

其他链接

其他stack相关方法：
    (1) stack, vstack, hstack差别