如何使用中间输出保存/加载模型

Question

我正在用Keras编写自动编码器：

inputs = Input((n_channels,))
l1 = Dense(40, activation="relu")(inputs)
l2 = Dense(19)(l1)
l3 = Dense(40, activation="relu")(l2)
training_layer = Dense(n_channels)(l3)
unify_layer = Model(inputs=inputs, outputs=l2)
training_layer = Model(inputs=inputs, outputs=training_layer)

我使用training_layer进行训练，使用unify_layer进行预测，因此当我在保存后继续学习时，我希望能够访问这两个端点。

根据Marcin的评论进行编辑Model.save只允许我保存一个模型。当我调用时：

unify_layer.save("unify")
training_layer.save("training")

然后

unify_layer = load_model("unify")
training_layer = load_model("training")

两层不再相连，即当我训练training_layer时，unify_layer没有被训练。

Answer 1

哦，我实际上可以使用save_weights和load_weights方法：

class Autoencoder():
    def __init__(self):
        inputs = Input((n_channels,))
        l1 = Dense(40, activation="relu")(inputs)
        l2 = Dense(19)(l1)
        l3 = Dense(40, activation="relu")(l2)
        training_layer = Dense(n_channels)(l3)
        self.unify_layer = Model(inputs=inputs, outputs=l2)
        self.training_layer = Model(inputs=inputs, outputs=training_layer)

    def save(self, filename):
        self.unify_layer.save_weights("unify_" + filename)
        self.training_layer.save_weights("training_" + filename)

    def load(self, filename):
        self.unify_layer.load_weights("unify_" + filename)
        self.training_layer.load_weights("training_" + filename)