深入理解pytorch 卷积层 核子和参数

In [19]: conv2 = nn.Conv2d(in_channels=3, out_channels=32, kernel_size=(4,3,2))


In [23]: for (name, param) in conv2.named_parameters():
    ...:     print(name)
  
  
weight
bias


也就是说每一个卷积层个有两个属性， 一个属性是weight， 另一个属性是bias
具体的就是conv2.weight   conv2.bias



 print(conv2.weight.shape)
torch.Size([32, 3, 4, 3, 2])

由于input_channel是3， output_channel=32， 算子的形状(fileter shape)是 4*3*2的
所以每一个算子的参数是  3*（4*3*2）== input_cahnnel * filter shape
这个卷积层有32个算子，所以一共有参数32*3*4*3*2个， 也就是conv2.weight的shape


In [44]: conv2.bias
Out[44]:
Parameter containing:
tensor([ 0.0269, -0.0939, -0.0467,  0.0744,  0.0148, -0.0733,  0.0333, -0.0539,
        -0.0310, -0.0291, -0.1154, -0.1048,  0.0117,  0.0682,  0.0352, -0.0773,
         0.0971,  0.0425, -0.0431, -0.0425, -0.0439,  0.0840, -0.0024,  0.0546,
        -0.0344,  0.0732,  0.0632, -0.0010, -0.1169,  0.1142, -0.0022,  0.0458],
       requires_grad=True)

In [45]: conv2.bias.shape
Out[45]: torch.Size([32])