Closed yix081 closed 7 years ago
It is for the historical reason that original AlexNet [1] used group in their structure, and as I am using Caffe's AlexNet pretrained model, I need to build the graph exactly the same as that.
[1] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.
for conv2, conv4, conv5, they all have parameter "group" sets to 2, which I feel is not necessary. What's the reason using it?