Linear深入分析

题目

'''
Description: Linear
Autor: 365JHWZGo
Date: 2021-12-15 15:17:55
LastEditors: 365JHWZGo
LastEditTime: 2021-12-15 16:06:23
'''

代码

import torch 
import torch.nn as nn

class Demo(nn.Module):
    def __init__(self,input_size,output_size):
        super(Demo,self).__init__()
        self.input_size = input_size
        self.output_size = output_size
        self.linear = nn.Linear(self.input_size,self.output_size)

    def forward(self,inp,out):
        output1 = self.linear(inp)
        output2 = self.linear(out)
        return output1,output2


if __name__ == '__main__':
    input_size = 4
    output_size =5
    

    demo = Demo(input_size,output_size)
    x = torch.randn(2,4)
    y = torch.randn(3,4)
    out1,out2 = demo(x,y)
    print(out1)
    print(out2)

运行结果

总结

Linear的作用是进行线性变化，实质上矩阵的乘法
$$y=W\ast x+b$$
W是权重，b是偏置

特征	W	b
维度	(out_features，in_features)	(out_features，1)

特征	x	y
维度	(in_features，random_input_size)	(out_features，random_input_size)

Linear类中init的定义

class Linear(Module):
	def __init__(self, in_features: int, out_features: int, bias: bool = True,device=None, dtype=None) -> None:
        factory_kwargs = {'device': device, 'dtype': dtype}
        super(Linear, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.weight = Parameter(torch.empty((out_features, in_features), **factory_kwargs))
        if bias:
            self.bias = Parameter(torch.empty(out_features, **factory_kwargs))
        else:
            self.register_parameter('bias', None)
        self.reset_parameters()

从中可以看出Linear是自己创造出W，bias是根据用户是否输入，如果为True，则
self.bias = Parameter(torch.empty(out_features, **factory_kwargs))，否则self.register_parameter('bias', None)