Hello, I want to make sure if I understand this right

unsymmetric padding leads to unsymmetric learning, is it right?

unsymmetric padding means choosing filter f = even number

So p = F-1/2 would be a fraction

Right! Iâ€™ll have to go back and find it in the lectures, but I do remember Prof Ng making the comment at some point in Week 1 that the common practice is to use odd filter sizes (e.g. f = 3, 5 or 7) because the math just works out more nicely. When the inputs are images, the height and width dimensions are typically an even number of pixels.

Hi Paulin, got one question-

I have an input size of (64 x 64 x 3) and I am using zero padding of size(3,3), filter size (7,7) , no of filters as 64 and stride (2,2)â€¦wanted to know what would be the size of the output feature map?

Prof Ng gives the formula in the lectures:

n_{out} = \displaystyle \lfloor \frac {n_{in} + 2p - f}{s} \rfloor + 1

Those strange looking brackets are the notation for the â€śfloorâ€ť mathematical function. So letâ€™s plug in the numbers from your example:

n_{out} = \displaystyle \lfloor \frac {64 + 2 * 3 - 7}{2} \rfloor + 1 = \displaystyle \lfloor \frac {63}{2} \rfloor + 1 = \lfloor 31.5 \rfloor + 1 = 32

So with 64 filters, the output should end up as 32 x 32 x 64.

Of course notice that is just the first layer that you show there. BatchNorm and ReLU donâ€™t affect the output size and then you have a pooling layer at the end which will further reduce the h and w dimensions, using the same formula as above. Pooling layers preserve the number of channels.

It was so puzzling for me to get a fraction value (63/2)â€¦floor functions solves thisâ€¦thanks for clarifying this doubt Paulinâ€¦cheers!!