ARC-faceloss  

I changed the loss of the original lenet5 handwirtten Network:softmax loss.I use the Arc-loss replace of it.

I havenot finished my training,I only trained 220 steps as fllows,but tha acc on test data is very high 98.93%,which is higher than the 
original softmax loss 98.57% and the cos-loss 98.95%.It is very fast.


After 10 training step(s), loss on test data is 1.98712,acc is 0.772333
After 20 training step(s), loss on test data is 0.786744,acc is 0.903
After 30 training step(s), loss on test data is 0.599833,acc is 0.917667
After 40 training step(s), loss on test data is 0.305006,acc is 0.958333
After 50 training step(s), loss on test data is 0.280962,acc is 0.962333
After 60 training step(s), loss on test data is 0.114537,acc is 0.977667
After 70 training step(s), loss on test data is 0.16493,acc is 0.977333
After 80 training step(s), loss on test data is 0.141905,acc is 0.978333
After 90 training step(s), loss on test data is 0.144544,acc is 0.978
After 100 training step(s), loss on test data is 0.117581,acc is 0.981
After 110 training step(s), loss on test data is 0.107124,acc is 0.983
After 120 training step(s), loss on test data is 0.0971116,acc is 0.982333
After 130 training step(s), loss on test data is 0.0867986,acc is 0.984333
After 140 training step(s), loss on test data is 0.0783345,acc is 0.984667
After 150 training step(s), loss on test data is 0.0639797,acc is 0.987333
After 160 training step(s), loss on test data is 0.0580864,acc is 0.990667
After 170 training step(s), loss on test data is 0.0584896,acc is 0.991
After 180 training step(s), loss on test data is 0.061686,acc is 0.989
After 190 training step(s), loss on test data is 0.0629968,acc is 0.990333
After 200 training step(s), loss on test data is 0.0669648,acc is 0.989
After 210 training step(s), loss on test data is 0.074284,acc is 0.987333
After 220 training step(s), loss on test data is 0.0624132,acc is 0.989333