An Efficient Neural Network Model by Weight Roll Algorithm

Abstract

Deploying deep learning models require extraction of the model weights from the training environment and saving them to files that can be shipped to production. Often complex models have large model file size and it is difficult to transport those models, this paper aims to reduce the size of the model file while transferring the trained weights to production environ-ment. Weight rolls is an algorithm that rolls down (reduces) the trained model weights to a smaller size, in some cases even re-duced by a proportion of one thousand (1,000). On the produc-tion environment this is again unrolled to regain the original weights that were learned by the neural network during its train-ing phase. Weight rolls uses a compressed pictorial representa-tion of the weights array along with a pix-to-weight neural net-work to transport the learned weights which can be used on the other end for the unrolling process. The pix-to-weight network maps the pixels of the compressed weight image to the original floating point values which in the unrolling phase is used to transform the pixels into corresponding floating point values of trained weights.