zhaohui-yang / legonet Goto Github PK

Hello. I was wondering whether in your implementation of the Lego version of the ResNet50 which is described in your paper, you just replaced all regular convolutional layers with their Lego counterparts with corresponding kernel sizes (just like in VGG16). Also would it be possible to share your code for implementing LegoNet based on ResNet and MobileNet architectures for reproducibility reasons.

Thanks in advance.

The current given implementation does not provide any speed up in training. The usage of lego filters actually slows down the training process compared to regular convolutional filters. Can you explain why this is happening?