新手速成---TensorFlow实现jaffe表情识别--step2补充+step3+step4

觉得上一篇写的太粗糙了，补充点内容
网络结构
[48×48]⇀conv2d［3×3］×128⇀pooling⇀conv2d[5×5]×64⇀pooling⇀conv2d[3×3]×32⇀pooling⇀fc[6×6×32]×200⇀dropout⇀softmax[200×7]
其实我训练了两种，epoch=50 和epoch=100，结果显示，epoch=100时结构更好（网上都是50）
epoch=50时
Epoch: 45, Loss= 21.501, Training Accuracy= 0.800
Epoch: 45, Test Loss= 27.431, Test Accuracy= 0.550
Epoch: 46, Loss= 20.672, Training Accuracy= 0.800
Epoch: 46, Test Loss= 26.583, Test Accuracy= 0.600
Epoch: 47, Loss= 21.156, Training Accuracy= 0.750
Epoch: 47, Test Loss= 26.609, Test Accuracy= 0.550
Epoch: 48, Loss= 21.344, Training Accuracy= 0.850
Epoch: 48, Test Loss= 26.205, Test Accuracy= 0.650
Epoch: 49, Loss= 19.028, Training Accuracy= 0.850
Epoch: 49, Test Loss= 25.373, Test Accuracy= 0.700
Epoch: 50,