Based on the lenet5 Network,I add a resnet layer between the first conv-layer and the second conv-layer,and still train it on mnist data.
But the acc is not very good.
one with data expand,batch size 100,40000steps acc on test data is 94.68%
second with data expand,batch size 500,8000steps acc on test data is 94.68%
third,four,five with no data expand,batch size 500, but 4000,10000,20000 steps,acc on test data is 97.79%,98.17%,98.34%