nn.Flatten()函数

注意默认的start_dim从1开始
https://pytorch.org/docs/stable/generated/torch.nn.Flatten.html?highlight=flatten#torch.nn.Flatten

示例1

卷积公式

$$outsize=\frac{inputsize-ksize+2\ast padding}{stride}+1$$

import torch
import torch.nn as nn
input = torch.randn(32, 1, 5, 5)
m = nn.Sequential(
    nn.Conv2d(1, 32, 5, 1, 1),  # 通过卷积，得到torch.size([32, 32, 3, 3]
    nn.Flatten())

output = m(input)
print(output.size())

>> torch.Size([32, 288])

示例2

import torch
import torch.nn as nn
input = torch.randn(32, 1, 5, 5)
m = nn.Sequential(
    nn.Conv2d(1, 32, 5, 1, 1),  # 通过卷积，得到torch.size([32, 32, 3, 3]
    nn.Flatten(start_dim=0))

output = m(input)
print(output.size())

>>torch.Size([9216])