Keras：将元数据连接成CNN

Question

我有一个CNN图像训练。我还获得了脸部的几何图形（68x2-68个点，x，y坐标）。我想在所有卷积层之后编码几何图形，并在完全连接的图层中使用它们。我正在使用vggFace模型。

''' 
Load the model 
''' 
vgg_model = VGGFace(
    include_top=False, 
    input_shape=(img_width, img_height, 3)) 


''' 
Customize the model 
''' 
# Add geometry input 
geo_input = Input(shape=(1,136,1)) 
geo_input = Flatten(name='flatten')(geo_input) 


last_layer = vgg_model.get_layer('pool5').output 
x = Flatten(name='flatten')(last_layer) 
x = concatenate([x, geo_input], axis=1) 
x = Dense(hidden_dim, activation='relu', name='fc6')(x) 
x = Dense(hidden_dim, activation='relu', name='fc7')(x) 
out = Dense(nb_class, activation='softmax', name='fc8')(x) 

custom_vgg_model = Model(
    [vgg_model.input, geo_input], 
    out) 

但我收到以下错误：

TypeError: Input layers to a `Model` must be `InputLayer` objects. Received inputs: [<tf.Tensor 'input_1:0' shape=(?, 224, 224, 3) dtype=float32>, <tf.Tensor 'flatten/Reshape:0' shape=(?, ?) dtype=float32>]. Input 1 (0-based) originates from layer type `Flatten`. 

Answer 1

这里：

# Add geometry input 
geo_input = Input(shape=(1,136,1)) 
geo_input = Flatten(name='flatten')(geo_input) 

在第二行中，geo_input不是输入了。它是Flatten图层的输出。您将此传递给Model创作。所以你需要保持正确的输入张量：

geo_input_tensor = Input(shape=(1,136,1)) 
geo_input = Flatten(name='flatten')(geo_input_tensor) 

..... 
...... 

custom_vgg_model = Model([vgg_model.input, geo_input_tensor], out)