Impressively fast

• By Hang Yu
• on 11/09/2017

My research group plans to train a Go playing deep neural net in the way of supervised learning. We used x8large instance with 2048 mini-batch divided evenly to each volta GPU. Each GPU host a 20 block deep residual net. Total params are about 70 millions. We are able to maxout the ability and it only take about 45s to evaluate 65536 training samples. We get about 70% training accuracy after about 8 hours of training.