理解 tensor, cat, unsquee, stack

假设一个 tensor 如下
$$[[[1,2],[4,5]],[[6,7],[8,9]]]$$

这个tensor 一共有 $3$ 个维度，shape 是 （2，2，2）
想知道有几个维度，数出有几层中括号就行，有几层中括号就有几维。

而且，随着中括号由外向里走，维度依次增加：从 0 变为 1 变为 2。

 .
 
那么 shape 的每一个分量是怎么得到的呢？还是以上述 tensor 举例子：
第 0 维
想要看第 0 维的分量，就要先褪去一层括号，得到
$$\begin{gathered} [[1,2],[4,5]], \\ [[6,7],[8,9]] \end{gathered}$$
把 $[[1,2],[4,5]]$ 看成一个元素，$[[6,7],[8,9]]$ 看成是另外一个元素，所以第 0 维一共包含 2 个元素，所以第 0 维的分量是2。

 
第 1 维
想要看第 0 维的分量，随便看 $[[1,2],[4,5]]$ 或者 $,[[6,7],[8,9]]$都行，在这里我们看
$$[[1,2],[4,5]]$$
把它的括号褪去，得到
$$[1,2],[4,5]$$
$1,2$是一个元素， $4,5$也是一个元素，所以一共包含两个元素，所以第 1 维也是2.

 
第 2维
以 $[1,2]$举例子，褪去括号是 $1,2$，有两个元素 $1$ 和 $2$，所以第 2 维的大小也是2。
 

举 torch.cat 的例子加以说明：

当torch.cat 的维度参数是 0

可以观察 T3 和 T1、T2的 shape，发现在第 0 维上由 2 变成 4了。

当torch.cat 的维度参数是 1，等价于-1，因为-1代表最高维度

可以观察 T3 和 T1、T2的 shape，发现在第 1 维上由 3 变成6了。

 

举 unsqueeze 的例子加以说明

unsqueeze，为了乘法

 

举 troch.stack 的例子

例子1

import torch

# creating tensors
x = torch.tensor([1., 3., 6., 10.])
y = torch.tensor([2., 7., 9., 13.])

# printing above created tensors
print("Tensor x:", x)
print("x's shape: ",x.shape,"\n")

print("Tensor y:", y)
print("y's shape: ",y.shape,"\n")
# join above tensor using "torch.stack()"
print("Default stack:")
t = torch.stack((x, y))

# print final tensor after join
print(t)
print("shape: ",t.shape,"\n")

print("Stack along dimension 0:")
t = torch.stack((x, y), dim=0)
print(t)
print("shape: ",t.shape,"\n")

print("Stack along dimension 1:")
t = torch.stack((x, y), dim=1)
print(t)
print("shape: ",t.shape,"\n")

 
输出如下：

Tensor x: tensor([ 1.,  3.,  6., 10.])
x's shape:  torch.Size([4]) 

Tensor y: tensor([ 2.,  7.,  9., 13.])
y's shape:  torch.Size([4]) 

Default stack:
tensor([[ 1.,  3.,  6., 10.],
        [ 2.,  7.,  9., 13.]])
shape:  torch.Size([2, 4]) 

Stack along dimension 0:
tensor([[ 1.,  3.,  6., 10.],
        [ 2.,  7.,  9., 13.]])
shape:  torch.Size([2, 4]) 

Stack along dimension 1:
tensor([[ 1.,  2.],
        [ 3.,  7.],
        [ 6.,  9.],
        [10., 13.]])
shape:  torch.Size([4, 2]) 

例子2

import torch

# creating tensors
x = torch.tensor([[1., 3., 6.], [10., 13., 20.]])
y = torch.tensor([[2., 7., 9.], [14., 21., 34.]])

# printing above created tensors
print("Tensor x:", x)
print("x's shape: ",x.shape,"\n")

print("Tensor y:", y)
print("y's shape: ",y.shape,"\n")
# join above tensor using "torch.stack()"
print("Default stack:")
t = torch.stack((x, y))

# print final tensor after join
print(t)
print("shape: ",t.shape,"\n")

print("Stack along dimension 0:")
t = torch.stack((x, y), dim=0)
print(t)
print("shape: ",t.shape,"\n")

print("Stack along dimension 1:")
t = torch.stack((x, y), dim=1)
print(t)
print("shape: ",t.shape,"\n")

print("Stack along dimension 2:")
t = torch.stack((x, y), dim=2)
print(t)
print("shape: ",t.shape,"\n")

 
输出如下：

Tensor x: tensor([[ 1.,  3.,  6.],
        [10., 13., 20.]])
x's shape:  torch.Size([2, 3]) 

Tensor y: tensor([[ 2.,  7.,  9.],
        [14., 21., 34.]])
y's shape:  torch.Size([2, 3]) 

Default stack:
tensor([[[ 1.,  3.,  6.],
         [10., 13., 20.]],

        [[ 2.,  7.,  9.],
         [14., 21., 34.]]])
shape:  torch.Size([2, 2, 3]) 

Stack along dimension 0:
tensor([[[ 1.,  3.,  6.],
         [10., 13., 20.]],

        [[ 2.,  7.,  9.],
         [14., 21., 34.]]])
shape:  torch.Size([2, 2, 3]) 

Stack along dimension 1:
tensor([[[ 1.,  3.,  6.],
         [ 2.,  7.,  9.]],

        [[10., 13., 20.],
         [14., 21., 34.]]])
shape:  torch.Size([2, 2, 3]) 

Stack along dimension 2:
tensor([[[ 1.,  2.],
         [ 3.,  7.],
         [ 6.,  9.]],

        [[10., 14.],
         [13., 21.],
         [20., 34.]]])
shape:  torch.Size([2, 3, 2]) 


Process finished with exit code 0

 

例子3

import torch

# creating tensors
x = torch.tensor([1., 3., 6., 10.])
y = torch.tensor([2., 7., 9., 13.])
z = torch.tensor([4., 5., 8., 11.])

# printing above created tensors
print("Tensor x:", x)
print("x's shape: ",x.shape,"\n")
print("Tensor y:", y)
print("y's shape: ",y.shape,"\n")
print("Tensor z:", z)
print("z's shape: ",z.shape,"\n")

print("Default stack:")
t = torch.stack((x, y, z))
# print final tensor after join
print(t)
print("shape: ",t.shape,"\n")

print("Stack along dimension 0:")
t = torch.stack((x, y, z), dim=0)
print(t)
print("shape: ",t.shape,"\n")

print("Stack along dimension 1:")
t = torch.stack((x, y, z), dim=1)
print(t)
print("shape: ",t.shape,"\n")

输出如下：

Tensor x: tensor([ 1.,  3.,  6., 10.])
x's shape:  torch.Size([4]) 

Tensor y: tensor([ 2.,  7.,  9., 13.])
y's shape:  torch.Size([4]) 

Tensor z: tensor([ 4.,  5.,  8., 11.])
z's shape:  torch.Size([4]) 

Default stack:
tensor([[ 1.,  3.,  6., 10.],
        [ 2.,  7.,  9., 13.],
        [ 4.,  5.,  8., 11.]])
shape:  torch.Size([3, 4]) 

Stack along dimension 0:
tensor([[ 1.,  3.,  6., 10.],
        [ 2.,  7.,  9., 13.],
        [ 4.,  5.,  8., 11.]])
shape:  torch.Size([3, 4]) 

Stack along dimension 1:
tensor([[ 1.,  2.,  4.],
        [ 3.,  7.,  5.],
        [ 6.,  9.,  8.],
        [10., 13., 11.]])
shape:  torch.Size([4, 3]) 

例子4

注意，被执行 stack 的两个 tensor 的shape 必须完全一致，不然会报错！

import torch

# creating tensors
x = torch.tensor([1., 3., 6., 10.])
y = torch.tensor([2., 7., 9.])

# printing above created tensors
print("Tensor x:", x)
print("Tensor y:", y)

# join above tensor using "torch.stack()"
print("Default stack")
t = torch.stack((x, y))
# print final tensor after join
print(t)

print("Stack along dimension 0:")
t = torch.stack((x, y), dim=0)
print(t)

print("Stack along dimension 1:")
t = torch.stack((x, y), dim=1)
print(t)

输出如下：

Tensor x: tensor([ 1.,  3.,  6., 10.])
Tensor y: tensor([2., 7., 9.])
Default stack
Traceback (most recent call last):
  File "", line 13, in <module>
    t = torch.stack((x, y))
RuntimeError: stack expects each tensor to be equal size, but got [4] at entry 0 and [3] at entry 1

cat和stack不同

• 2个 tensor 执行cat，新生成 tensor 相比原始 tesnsor 维度不会增加，但是 shape 的分量值会改变，比如两个 shape 都是 (2,3) 的tensor 沿着 0 维 cat，那么得到的新的 tensor 的shape将会是 (4,3)。
• 两个 tensor 执行 stack，得到的新的 tensor相比元素 tensor维度会增加。比如比如 2 个 shape 都是 (2,3) 的tensor 沿着 0 维 stack，那新生成的 tensor 的shape将会是 (2,2,3)。新生成的 tensor 的 shape 的第 0 维的值是 2 是因为 参与 stack 的 tensor 数目是 2。如果有 3 个 tensor 参与 stack，沿着 0 维，那么新 tensor 的shape将会是 (3,2,3)。
• 注意 tensor 执行 stack 新生成的 tensor 的 shape 某一个维度的值会等于参与 stack 的tensor 的数目。有 2 个tensor 参与 stack，那么就会是 2，有 3个，就会是 3.