Training Sparse networks in Caffe

[Gopalakrishna Hegde]

Hi,

         I was studying papers on compressing deep neural networks ( mainly Deep Compression and SqueezeNet). These require pruning certain weights/connections between neurons in different layers if the weights are very weak compared to a threshold.  After removing those connections, the remaining weights need to be retrained. I am wondering if it is possible to construct a sparse network and train in Caffe. Is there any way to input the network connection file similar to weight initialization file?

Appreciate any suggestions on this.

Thanks

Gopal