Caffenet norm2 layer output blob

[Ivan Iotzov]

Hi all, I had a quick question about feature extraction from the norm2 layer of the Caffenet reference model.

I am using the python script classify.py to extract features from the norm2 layer by modifying classifier.py to output the full norm2 blob. Originally, the output consisted of a 10x256x13x13 matrix. I modified the deploy.prototxt to limit the batch size to 1 and turned off oversampling and the output matrix was then 1x256x13x13, which can be reduced down to just 256x13x13.

What is this matrix representing? I know that there are 256 kernels that are applied in the conv2 layer, so that explains the 256 dimension. But I can't seem to figure out where the 13x13 dimension comes from. I assume they are the kernel responses to various locations in the image, but I am interested in which locations specifically are being represented by each response. 

If someone that knows better than me how exactly the output from each layer is formed could help me understand what this matrix represents, I would be incredibly appreciative. 

I have attached my classify.py and classifier.py files in case they are relevant.

Thank you all for your help.