In this study, we focus on reducing training time of disparity networks.We use kernel re-parameterization to build an under-parameterized model to reduce operations required for training. Experiments have shown that we can reduce the time by 13.7% without decreasing accuracy.
