tensorflow构建网络--Keras

方式2：本文将解释如何使用keras构建网络

1.定义输入

这里由于使用Keras因此输入tensor必须使用Input()进行转换
from keras.layers import Input
#假设输入的形状为[224,224,3]
X_input = [224,224,3]
X_input = Input(X_input)

2.搭建网络

def net(X_input,out_put):
	'''
	X_input:输入tensor的维度
	out_put:输出的维度
	'''
	#填0卷积，注意作用
	X = ZeroPadding2D(padding=(3,3))(X_input)
	X = Conv2D(64,kernel_size=(7,7),strides=(2,2),name='c1')(X)
	X = MaxPooling2D(pool_size=(3,3),strides=(2,2),name='mx_pp1')(X)
	X = Activation('relu',name='Ac1')(X)
	
	X = Conv2D(filters=128,kernel_size=(3,3),strides=(1,1),name='c2')(X)
	X = BatchNormalization(name='Bn1')(X)
	X = Activation('relu',name='Ac2')(X)
	
	#均值池化
	X = AveragePooling2D(name='AVG_pool_last')(X)
	#把X拉长
    X = Flatten()(X)
    #全连接
    X = Dense(out_put,activation='softmax',name='fc')(X)
	#此处Model用来画net的图
	My_model_png = Model(inputs=X_input,outputs=X,name='My_model_png')
	return X,My_model_png
	
if __name__ == '__main__':
	X_input = [224,224,3]
	X_input = Input(X_input)
	out_put = 5
	X,My_model_png = net(X_input,out_put)
	My_model_png.summary()
	path_png = '图片存储路径,需构建出.png'
	from keras.utils import plot_model
	plot_model(My_model_png, to_file=path_png,show_shapes=True)

输出：
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 224, 224, 3)       0         
_________________________________________________________________
zero_padding2d_1 (ZeroPaddin (None, 230, 230, 3)       0         
_________________________________________________________________
c1 (Conv2D)                  (None, 112, 112, 64)      9472      
_________________________________________________________________
mx_pp1 (MaxPooling2D)        (None, 55, 55, 64)        0         
_________________________________________________________________
Ac1 (Activation)             (None, 55, 55, 64)        0         
_________________________________________________________________
c2 (Conv2D)                  (None, 53, 53, 128)       73856     
_________________________________________________________________
Bn1 (BatchNormalization)     (None, 53, 53, 128)       512       
_________________________________________________________________
Ac2 (Activation)             (None, 53, 53, 128)       0         
_________________________________________________________________
AVG_pool_last (AveragePoolin (None, 26, 26, 128)       0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 86528)             0         
_________________________________________________________________
fc (Dense)                   (None, 5)                 432645    
=================================================================
Total params: 516,485
Trainable params: 516,229
Non-trainable params: 256