pytorch中的reshape()、view()、transpose()和flatten()

文章内容皆为个人理解，如有不足欢迎指正。

1、torch.reshape()
reshape()可以由torch.reshape(),也可由torch.Tensor.reshape()调用
其作用是在不改变tensor元素数目的情况下改变tensor的shape

import torch
import numpy as np
a = np.arange(24)
b = a.reshape(4,3,2)
print(np.shape(a))
print(b,np.shape(b))

'''结果
(24,)
[[[ 0  1]
  [ 2  3]
  [ 4  5]]

 [[ 6  7]
  [ 8  9]
  [10 11]]

 [[12 13]
  [14 15]
  [16 17]]

 [[18 19]
  [20 21]
  [22 23]]] (4, 3, 2)
'''

2、view()
view()只可以由torch.Tensor.view()来调用
view()和reshape()在效果上是一样的，区别是view（）只能操作contiguous的tensor，且view后的tensor和原tensor共享存储，reshape（）对于是否contiuous的tensor都可以操作。

3、transpose()

torch.transpose(input, dim0, dim1) -> Tensor

将输入数据input的第dim0维和dim1维进行交换

#官方例子
>>> x = torch.randn(2, 3)
>>> x
tensor([[ 0.9068,  1.8803, -0.5021],
        [-0.6576,  0.6334, -0.8961]])
>>> torch.transpose(x, 0, 1)
tensor([[ 0.9068, -0.6576],
        [ 1.8803,  0.6334],
        [-0.5021, -0.8961]])

4、flatten()
torch.flatten()的输入是tensor

torch.flatten(input, start_dim=0, end_dim=-1) → Tensor

其作用是将输入tensor的第start_dim维到end_dim维之间的数据“拉平”成一维tensor，

#官方例子
>>> t = torch.tensor([[[1, 2],
                               [3, 4]],
                              [[5, 6],
                               [7, 8]]])
>>> torch.flatten(t)
tensor([1, 2, 3, 4, 5, 6, 7, 8])
>>> torch.flatten(t, start_dim=1)
tensor([[1, 2, 3, 4],
        [5, 6, 7, 8]])

torch.nn.Flatten()可以理解为一种网络结构，类似Conv2d、Linear。一般放在卷积层和全连接层之间，将卷积层输出“拉平”成一维，

>>> m = torch.nn.Sequential(
    torch.nn.Conv2d(1, 32, 5, 1, 1),
    torch.nn.Flatten(),
    torch.nn.Linear(160,10))
>>> m
Sequential(
  (0): Conv2d(1, 32, kernel_size=(5, 5), stride=(1, 1), padding=(1, 1))
  (1): Flatten()
  (2): Linear(in_features=160, out_features=10, bias=True)
)

参考：

https://congluwen.top/2018/12/pytorch_reshape_view/
https://blog.csdn.net/GhostintheCode/article/details/102530451