批量归一化与未归一化的比较

未归一化的比较：
training on cpu
epoch 1, loss 0.0091, train acc 0.104, test acc 0.100, time 61.8 sec
epoch 2, loss 0.0053, train acc 0.464, test acc 0.611, time 59.1 sec
epoch 3, loss 0.0034, train acc 0.652, test acc 0.557, time 61.1 sec
epoch 4, loss 0.0029, train acc 0.710, test acc 0.645, time 58.6 sec
epoch 5, loss 0.0025, train acc 0.750, test acc 0.732, time 58.9 sec
epoch 6, loss 0.0023, train acc 0.776, test acc 0.778, time 58.1 sec
epoch 7, loss 0.0021, train acc 0.794, test acc 0.801, time 58.4 sec
epoch 8, loss 0.0019, train acc 0.812, test acc 0.778, time 58.1 sec
epoch 9, loss 0.0018, train acc 0.823, test acc 0.815, time 61.2 sec
epoch 10, loss 0.0018, train acc 0.832, test acc 0.823, time 68.8 sec