在用lenet5做mnist识别时，网络收敛速度很慢经过100多轮训练仍在在50%准确率徘徊；参考其他人一轮变可以达到95+%的代码，进行修改

原始结构及参数：（图像也没有处理为0-1浮点数）

    def get_lenet5(self):
        input = Input(shape=(cf.Img_size, cf.Img_size, cf.Img_channels))
        conv1 = Conv2D(cf.Conv1_chs, kernel_size=cf.Conv1_kernel_size, padding='same', activation='relu')(input)
        pool1 = MaxPooling2D(2, 2, padding='same')(conv1)
        conv2 = Conv2D(cf.Conv2_chs, kernel_size=cf.Conv2_kernel_size, padding='same', activation='relu')(pool1)
        pool2 = MaxPooling2D(2, 2, padding='same')(conv2)
        flat = Flatten()(pool2)
        fc1 = Dense(cf.Fc1_size, activation='relu')(flat)
        fc2 = Dense(cf.Fc2_size, activation='relu')(fc1)
        output = Softmax()(fc2)

        model = Model(inputs=input, outputs=output)
        model.compile(optimizer=Adam(lr=cf.Learning_rate), loss='categorical_crossentropy', metrics=['accuracy'])
        return model

Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 28, 28, 1)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 28, 28, 20)        520       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 14, 14, 20)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 14, 14, 50)        25050     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 7, 7, 50)          0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 2450)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 500)               1225500   
_________________________________________________________________
dense_2 (Dense)              (None, 10)                5010      
_________________________________________________________________
softmax_1 (Softmax)          (None, 10)                0         
=================================================================

修改：

1）将图像处理为浮点数

x_train,x_test = x_train/255,x_test/255

2）去掉最后单独的softmax层，将第二层全连接激活方式改为softmax

fc2 = Dense(cf.Fc2_size, activation='softmax')(fc1)
model = Model(inputs=input, outputs=fc2)

3）经过测试，学习率设置为3e-3