[基础网络] ZF Net Keras 实现

短命女 2022-05-13 09:16 123阅读 0赞

def conv2d_maxpool(x, filters, conv_size=(3,3), padding='valid', strides=(1, 1),
                  pool_strides =(2,2),pool_size = (3,3)
                  name=None):
        x = Conv2D(
            filters, conv_size,
            strides=strides,
            padding=padding,
            use_bias=False,
            name=conv_name)(x)
        #x = BatchNormalization(axis=bn_axis, scale=False, name=bn_name)(x)
        x = Activation('relu', name=name)(x)
        x= MaxPooling2D(pool_size = pool_size,strides = pool_strides)(x)
        return x
    
    class Total_Models():
        def __init__(self, nb_classes):
    
            self.nb_classes = nb_classes
            metrics = ['accuracy']
            if self.nb_classes >= 10:
            self.model,self.original_model = self.zf_model() 
            optimizer = Adam(lr=1e-5)
            self.model.compile(loss='categorical_crossentropy', optimizer=optimizer,
                               metrics=metrics) #
            print(self.model.summary())
        def zf_model(self):
            inputs = Input(shape=(299,299,3))
            layer_1 = conv2d_maxpool(inputs, 96, conv_size=(7,7), padding='valid', strides=(2, 2),
                  pool_strides =(2,2),pool_size = (3,3))
            layer_2 = conv2d_maxpool(layer_1, 256, conv_size=(5,5), padding='valid', strides=(2, 2),
                  pool_strides =(2,2),pool_size = (3,3))
            layer_3 = Conv2D(384, (3, 3),strides=1,padding='same',activation= 'relu', use_bias=True, name='layer3')(layer_2)
            layer_4 = Conv2D(384, (3, 3),strides=1,padding='same',activation='relu', use_bias=True, name='layer4')(layer_3)
            layer_5 = conv2d_maxpool(layer_4, 256, conv_size=(5,5), padding='same', strides=(2, 2),
                  pool_strides =(2,2),pool_size = (3,3))
            flatten = Flatten()(layer_5)
            layer_6 = Dense(4096,activation = 'relu')(flatten)
            dropout_6 = Dropout(0.5)(layer_6)
            layer_7 = Dense(4096,activation ='relu')(dropout_6)
            dropout_7 = Dropout(0.5)(layer_7)
            layer_8 = Dense(self.nb_classes)(dropout_7)
            model = Model(inputs= inputs,outputs = layer_8)
            parallel_model=multi_gpu_model(model,gpus=4)
            return parallel_model,model