rbsathish / Model-pruning

Model-pruning

#Model-pruning

I just used model-pruning in my cnn model. The cnn model is not yet fully completed.

MODEL PRUNING:

          Nowadays, model pruning needs very much for running the machine learning models on mobile and embedded devices.
          
          In production Reduce memory,battery,and hardware consumption without sacrificing accuracy we need to deploy lightweight models on devices and also we need to guarante privacy with private on device computation.
          
          And in Research Learning Dynamics of over-parametrized and under parametrized networks,constrained optimization,initialization & more.
          
          Research and on the production was using their own implementation of pruning.

Network pruning :

            Less number of weights
            Eg. 60m weights to 6m weights

Weights sharing:

            Reduce storage for each remaining weights.
            Eg. 32bit to 4bit

Huffman coding:

            Entropy of total remaining weights

References:

            https://www.youtube.com/watch?v=CrDRr2fxbsg