How to calculate the first parameter of nn.Linear()

1

I am new to PyTorch.

When I was reading some examples, I have some questions.

Latest github example of mnist, in line 15 there is a number 320 in nn.Linear(), but I don’t know how to get that number.

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__() 
        self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
        self.conv2 = nn.Conv2d(10, 20, kernel_size=5) 
        self.conv2_drop = nn.Dropout2d()
        self.fc1 = nn.Linear(320, 50) 
        self.fc2 = nn.Linear(50, 10)

I only know that without conv2_drop, in fc1 it should be 2055, but I don’t get how 320 come.

Thank you so much if you can help me.

2 Likes
2

well I forgot it is markdown…so in fc1 it should be 20*5*5, but I don’t get how 320 come.

3

Finally I understand…4*4*16

4

Well it’s 4*4*20 = 320.

You can calculate the shape for the forward pass for each operation.
While a convolution with kernel_size=5 and no padding shrinks the activation by 4 pixels in height and width, a max pooling of kernel_size=2 and stride=2 pools the activation to half its size.
From the forward pass:

# x.shape [1, 28, 28]
x = self.conv1(x) # x.shape: [10, 24, 24]
x = F.max_pool2d(x, 2) # x.shape: [10, 12, 12]
x = F.relu(x)
x = self.conv2(x) # x.shape: [20, 8, 8]
x = self.conv2_drop(x)
x = F.max_pool2d(x, 2) # x.shape: [20, 4, 4]
x = F.relu(x)
x = x.view(-1, 320)
2 Likes