Query about weight filter size in AlexNet

I have analysed the pretrained alexnet model. I have attached the screen shot. In that conv2 row, weights are given as 5*5*48*256. what is 48 here?. As per my knowledge, 48 represents the previous layer (number of channels in previous layer). but number of previous layer is 96 after pool1. likewise, same doubt in conv4 and conv5.

Can anyone explain this question?

Answer 1

[Harikrishnan Balachandran Nair]

The reason that the third dimension of weights in the mentioned layer is not the same as the number of channels in input is that the corresponding layer in AlexNet perform grouped Convolution. You can refer to the following Documentation to learn more about AlexNet : https://www.mathworks.com/help/deeplearning/ref/alexnet.html.

[deepika s]

Yes. but how the third dimesion (48) of weight in Conv2 is convolves with number of channels (nothing but dimension 96) in the pool1 output? Both dimensions are different.