如何在Keras中正确拼接Flatten层与特征向量？

问题重现 #

你尝试在Keras中把VGGFace的Flatten输出和自定义特征向量拼接，但运行时抛出了图断开的错误，你的代码如下：

# custom parameters
n_features = 38
vgg_model = VGGFace(include_top=False, input_shape=(224, 224, 3))
last_layer = vgg_model.get_layer('pool5').output
x = Flatten(name='flatten')(last_layer)
# feature vector
feature_vector = Input(shape = (n_features,))
conc = concatenate(([x, feature_vector]), axis=1)
layer_intermediate = Dense(128, activation='relu', name='fc6')(conc)
layer_intermediate1 = Dense(32, activation='relu', name='fc7')(layer_intermediate)
out = Dense(5, activation='softmax', name='fc8')(layer_intermediate1)
custom_vgg_model = Model(vgg_model.input, out)

得到的错误信息：

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("input_88:0", shape=(?, 38), dtype=float32) at layer "input_88". The following previous layers were accessed without issue: ['input_87', 'conv1_1', 'conv1_2', 'pool1', 'conv2_1', 'conv2_2', 'pool2', 'conv3_1', 'conv3_2', 'conv3_3', 'pool3', 'conv4_1', 'conv4_2', 'conv4_3', 'pool4', 'conv5_1', 'conv5_2', 'conv5_3', 'pool5', 'flatten']

错误原因 #

这个问题的核心很直白：你的模型实际上有两个输入源——VGGFace接收的图像输入，以及你定义的38维特征向量输入，但你在创建Model对象时只传入了vgg_model.input作为唯一输入，完全没把feature_vector这个输入张量包含进去。Keras找不到这个特征向量输入的来源，自然就会报错说计算图断开了。

修正后的代码 #

只需要修改最后一行的Model定义，把两个输入都传进去就行，完整代码如下：

# custom parameters
n_features = 38
vgg_model = VGGFace(include_top=False, input_shape=(224, 224, 3))
last_layer = vgg_model.get_layer('pool5').output
x = Flatten(name='flatten')(last_layer)
# feature vector
feature_vector = Input(shape = (n_features,))
conc = concatenate([x, feature_vector], axis=1)  # 这里可以简化括号，不用套两层
layer_intermediate = Dense(128, activation='relu', name='fc6')(conc)
layer_intermediate1 = Dense(32, activation='relu', name='fc7')(layer_intermediate)
out = Dense(5, activation='softmax', name='fc8')(layer_intermediate1)
# 关键修改：传入两个输入组成的列表
custom_vgg_model = Model(inputs=[vgg_model.input, feature_vector], outputs=out)

额外注意事项 #

训练这个模型的时候，你需要传入两组输入数据，比如：

# 假设X_img是形状为(样本数,224,224,3)的图像数据，X_feat是形状为(样本数,38)的特征向量数据
custom_vgg_model.fit([X_img, X_feat], y_labels, epochs=10, batch_size=32)

预测的时候也要对应传入两个输入，这样模型才能正常运行。

内容的提问来源于stack exchange，提问作者shinigami023