How to improve the utilization of GPU memory in training?

First of all, thanks for your fantastic work!

I wonder since the batchsize is fixed in this work(equals to the num of exposure images) and can't  be enlarged , how can I improve the utilization of GPU memory in training stage?