Understanding model parameters

vision
1

Hello, I have the following model:

class CovNet(nn.Module):

    
    def __init__(self, imageSize):
        super(CovNet, self).__init__()
        kernel = (3,3)
        self.conv1 = nn.Conv2d(in_channels = 1, out_channels = 3, kernel_size = kernel)
        wOut = self.parametersCalculator(imageSize,3,0,1,1) #Calculate size of data after conv
        wOut = self.parametersCalculator(wOut,2,0,2,1) #Calculate size of data after maxpool
        self.fc2 = nn.Linear(3 * wOut * wOut, 120)
        self.fc3 = nn.Linear(120, 60)
        self.fc4 = nn.Linear(60, 10)

    def forward(self, x):
        x = F.max_pool2d(F.relu(self.conv1(x)), 2)
        x = x.view(-1, self.num_flat_features(x))
        x = F.relu(self.fc2(x))
        x = F.relu(self.fc3(x))
        x = self.fc4(x)
        return x

    def num_flat_features(self, x):
        print(x.size())
        size = x.size()[1:]  # all dimensions except the batch dimension
        num_features = 1
        for s in size:
            num_features *= s
        return num_features
    
    def parametersCalculator(self,W,K,P,S,D):
        return int((W+2*P-D*(K-1)-1)/S + 1)

And I am trying to get the parameters of this model. To this end I wrote this line:

params = list(model.parameters())

This variable is an array with 8 length with the following shapes:
params[0].shape --> (3,1,3,3)
params[1].shape --> (3)
params[2].shape --> (120,507)
params[3].shape --> (120)
params[4].shape --> (60,120)
params[5].shape --> (60)
params[6].shape --> (10, 60)
params[7].shape --> (10)

The image size is 28 by 28.

Can someone explain to me the shapes of the params variable, i dont seem to understand what are the shapes of params 1,3,5 and 6? Are theese the biases?

Thank you in advance!

2

Hi,

Yes these are the biases for the convolution and linear layers. You can set bias=False when creating them and they will disappear.

1 Like