tensorflow2使用Xception预训练网络完成多输出任务-同时完成分类和回归
本文的数据基于最近在研究的一个项目，数据保存在npy文件，数据格式 [[800，640],1,1,1]，800*640的图片，按键，x坐标，y坐标。
现在要建立神经网络完成输入图片对 按键、x坐标、y坐标的预测。即同时完成对按键的分类和对坐标的回归。
Xception是tensorflow2自带的在imagenet上预训练的模型，在imagenet上取得了验证集top1 0.79和top5 0.945的准确率，该模型的输入数据维度顺序必须为（高度，宽度，通道数），默认尺寸是299*299。
Xception常用参数：
include_top : False 不包含全连接层 True包含全连接层
weights ：None 随机初始化 imagenet 加载在imagenet上的预训练权值
imput_shape: 可选参数，代表输出尺寸元组，仅当include_top =False时有效，否则输入必须是299*299。输入数据宽高不能小于71.
ok，开始加载数据并进行预处理
#WIDTH 图像宽 HEIGHT图像高
train_data = np.load(file_name, allow_pickle=True)
print('training_data-{}.npy'.format(i), len(train))
train = train_data[:-500]
test = train_data[-500:]
#交换宽高，并添加通道数维度，然后进行归一化
train_X = np.array([i[0] for i in train]).swapaxes(0,1) .reshape(-1,HEIGHT,WIDTH,3)/255
train_Y_key = np.array([i[1] for i in train])
train_Y_x = np.array([i[2] for i in train])/WIDTH
train_Y_y = np.array([i[3] for i in train])/HEIGHT
test_X = np.array([i[0] for i in test]).swapaxes(0,1).reshape(-1,HEIGHT,WIDTH,3)/255
test_Y_key = np.array([i[1] for i in test])
test_Y_x = np.array([i[2] for i in test])/WIDTH
test_Y_y = np.array([i[3] for i in test])/HEIGHT构建模型，模型输入为 (HEIGHT,WIDTH, 3) 输出为三个[out_x, out_y,out_key]
xception = tf.keras.applications.Xception(weights='imagenet',include_top=False,input_shape=(HEIGHT,WIDTH, 3))
xception.trainable = False
inputs = tf.keras.layers.Input(shape=(HEIGHT,WIDTH, 3))
x = xception(inputs)
x = tf.keras.layers.GlobalAveragePooling2D()(x)
x1 = tf.keras.layers.Dense(512, activation='relu')(x)
x1 = tf.keras.layers.Dense(256, activation='relu')(x1)
out_x = tf.keras.layers.Dense(1,name='out_x')(x1)
out_y = tf.keras.layers.Dense(1,name='out_y')(x1)
x2 = tf.keras.layers.Dense(512, activation='relu')(x)
out_key = tf.keras.layers.Dense(9,activation='softmax', name='out_key')(x2)
predictions = [out_x, out_y,out_key]
model = tf.keras.models.Model(inputs=inputs, outputs=predictions)
print(model.summary())模型详情：
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_2 (InputLayer) [(None, 400, 320, 3) 0
__________________________________________________________________________________________________
xception (Model) (None, 13, 10, 2048) 20861480 input_2[0][0]
__________________________________________________________________________________________________
global_average_pooling2d (Globa (None, 2048) 0 xception[1][0]
__________________________________________________________________________________________________
dense (Dense) (None, 512) 1049088 global_average_pooling2d[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 256) 131328 dense[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) (None, 512) 1049088 global_average_pooling2d[0][0]
__________________________________________________________________________________________________
out_x (Dense) (None, 1) 257 dense_1[0][0]
__________________________________________________________________________________________________
out_y (Dense) (None, 1) 257 dense_1[0][0]
__________________________________________________________________________________________________
out_key (Dense) (None, 9) 4617 dense_2[0][0]
==================================================================================================
Total params: 23,096,115
Trainable params: 2,234,635
Non-trainable params: 20,861,480
__________________________________________________________________________________________________
None
下面是编译模型的代码
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
loss={'out_x':'mse',
'out_y':'mse',
'out_key':'sparse_categorical_crossentropy'},
metrics=['mae','acc']
)训练代码如下
model.fit(train_X, [train_Y_x,train_Y_y,train_Y_key], epochs=1,batch_size = 32,validation_data = (test_X,[test_Y_x,test_Y_y,test_Y_key])
)so easy!