Skip to content

Latest commit

 

History

History
2 lines (2 loc) · 496 Bytes

README.md

File metadata and controls

2 lines (2 loc) · 496 Bytes

Reparametrization of FullyConnected/Convolutional Layers to absorb Batch Normalization

This repository serves as quick example on how to apply reparametrization to a Neural Network, allowing the partial or even complete absortion of Batch Normalization layers by previous Dense / Convolutional layers. This technique can be applied at inference/classification step to speed up computations. It has no efect on the training phase, as it still requires standard training to be carried out.