Pruning less important weights in caffemodel

[Sivagnanam Namasivayamurthy]

Hi,

I'm working on an image classification problem using CNN in Caffe. The trained model has nearly 130Million weights, out of which ~120Million weights are infinitesimally small (i.e., ~0). I would like to prune these weights and re-train with remaining weights. I know its not straightforward to do this in Caffe.

Has anyone tried similar approach? any suggestions would be helpful.

[pius kwao]

Hi, I my research in this area as well. Have you been able to come across any solutions?

[Siva]

Yes, read the weights from .caffemodel into numpy arrays. Do the pruning in the arrays and load it back into the .caffemodel object and re-train it. There are no direct API's available for doing this though.