PyTorch函数解释：cat、stack、transpose、permute、squeeze、unsqueeze

---

原文链接请参考：PyTorch 常用函数解析 | 梦家博客

---

torch.cat() 张量拼接

对张量沿着某一维度进行拼接。连接后数据的总维数不变。，ps:能拼接的前提是对应的维度相同！！！

例如对两个2维tensor（分别为2*3,1*3）进行拼接，拼接完后变为3*3的2维 tensor。

In [1]: import torch

In [2]: torch.manual_seed(1)
Out[2]: <torch._C.Generator at 0x19e56f02e50>

In [3]: x = torch.randn(2,3)

In [4]: y = torch.randn(1,3)

In [5]: x
Out[5]:
tensor([[ 0.6614,  0.2669,  0.0617],
        [ 0.6213, -0.4519, -0.1661]])

In [6]: y
Out[6]: tensor([[-1.5228,  0.3817, -1.0276]])

In [9]: torch.cat((x,y),0)
Out[9]:
tensor([[ 0.6614,  0.2669,  0.0617],
        [ 0.6213, -0.4519, -0.1661],
        [-1.5228,  0.3817, -1.0276]])

以上dim=0 表示按列进行拼接，dim=1表示按行进行拼接。

代码如下：

In [11]: z = torch.randn(2,2)

In [12]: z
Out[12]:
tensor([[-0.5631, -0.8923],
        [-0.0583, -0.1955]])

In [13]: x
Out[13]:
tensor([[ 0.6614,  0.2669,  0.0617],
        [ 0.6213, -0.4519, -0.1661]])

In [14]: torch.cat((x,z),1)
Out[14]:
tensor([[ 0.6614,  0.2669,  0.0617, -0.5631, -0.8923],
        [ 0.6213, -0.4519, -0.1661, -0.0583, -0.1955]])

torch.stack() 张量堆叠

torch.cat()拼接不会增加新的维度，但torch.stack()则会增加新的维度。

例如对两个1*2 维的 tensor 在第0个维度上stack，则会变为2*1*2的 tensor；在第1个维度上stack，则会变为1*2*2 的tensor。

In [22]: x = torch.randn(1,2)

In [23]: y = torch.randn(1,2)

In [24]: x.shape
Out[24]: torch.Size([1, 2])

In [25]: x = torch.randn(1,2)

In [26]: y = torch.randn(1,2)

In [27]: torch.stack((x,y),0) # 维度0堆叠
Out[27]:
tensor([[[-1.8313,  1.5987]],

        [[-1.2770,  0.3255]]])

In [28]: torch.stack((x,y),0).shape
Out[28]: torch.Size([2, 1, 2])

In [29]: torch.stack((x,y),1) # 维度1堆叠
Out[29]:
tensor([[[-1.8313,  1.5987],
         [-1.2770,  0.3255]]])

In [30]: torch.stack((x,y),1).shape
Out[30]: torch.Size([1, 2, 2])

torch.transpose() 矩阵转置

举例说明

torch.manual_seed(1)
x = torch.randn(2,3)
print(x)

原来x的结果：

 0.6614  0.2669  0.0617
 0.6213 -0.4519 -0.1661
[torch.FloatTensor of size 2x3]

将x的维度互换:x.transpose(0,1) ，其实相当于转置操作！
结果

0.6614  0.6213
 0.2669 -0.4519
 0.0617 -0.1661
[torch.FloatTensor of size 3x2]

torch.permute() 多维度互换

permute是更灵活的transpose，可以灵活的对原数据的维度进行调换，而数据本身不变。

In [31]: x = torch.randn(2,3,4)

In [32]: x
Out[32]:
tensor([[[ 0.7626,  0.4415,  1.1651,  2.0154],
         [ 0.2152, -0.5242, -1.8034, -1.3083],
         [ 0.4100,  0.4085,  0.2579,  1.0950]],

        [[-0.5065,  0.0998, -0.6540,  0.7317],
         [-1.4567,  1.6089,  0.0938, -1.2597],
         [ 0.2546, -0.5020, -1.0412,  0.7323]]])

In [33]: x.shape
Out[33]: torch.Size([2, 3, 4])

In [34]: x.permute(1,0,2) # 0维和1维互换，2维不变！
Out[34]:
tensor([[[ 0.7626,  0.4415,  1.1651,  2.0154],
         [-0.5065,  0.0998, -0.6540,  0.7317]],

        [[ 0.2152, -0.5242, -1.8034, -1.3083],
         [-1.4567,  1.6089,  0.0938, -1.2597]],

        [[ 0.4100,  0.4085,  0.2579,  1.0950],
         [ 0.2546, -0.5020, -1.0412,  0.7323]]])

In [35]: x.permute(1,0,2).shape
Out[35]: torch.Size([3, 2, 4])

torch.squeeze() 和 torch.unsqueeze()

常用来增加或减少维度,如没有batch维度时，增加batch维度为1。

• squeeze(dim_n)压缩，减少dim_n维度 ，即去掉元素数量为1的dim_n维度。
• unsqueeze(dim_n)，增加dim_n维度，元素数量为1。

In [38]: x = torch.randn(1,3,4)

In [39]: x.shape
Out[39]: torch.Size([1, 3, 4])

In [40]: x
Out[40]:
tensor([[[-0.4791,  0.2912, -0.8317, -0.5525],
         [ 0.6355, -0.3968, -0.6571, -1.6428],
         [ 0.9803, -0.0421, -0.8206,  0.3133]]])

In [41]: x.squeeze()
Out[41]:
tensor([[-0.4791,  0.2912, -0.8317, -0.5525],
        [ 0.6355, -0.3968, -0.6571, -1.6428],
        [ 0.9803, -0.0421, -0.8206,  0.3133]])

In [42]: x.squeeze().shape
Out[42]: torch.Size([3, 4])

In [43]: x.unsqueeze(0)
Out[43]:
tensor([[[[-0.4791,  0.2912, -0.8317, -0.5525],
          [ 0.6355, -0.3968, -0.6571, -1.6428],
          [ 0.9803, -0.0421, -0.8206,  0.3133]]]])

In [44]: x.unsqueeze(0).shape
Out[44]: torch.Size([1, 1, 3, 4])