tf.keras 输入层仅用于推理期间

Question

I'd like to create an auxiliary input layer that I use solely during inference time, but can't figure out how to do it. I'd like to do something like the following:

inputs = tf.keras.Input(shape=(784,))

x = tf.keras.layers.Dense(64, activation='relu')(inputs)
x = tf.keras.layers.InputLayer((64,), name='input_foo')(x)
predictions = tf.keras.layers.Dense(64, activation='relu')(x)

model = tf.keras.Model(inputs=inputs, outputs=predictions)
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
model.fit(data, labels)  # starts training
model.save('foo.h5')

model = tf.keras.models.load_model('foo.h5')
inference_model = tf.keras.Model(model.get_layer('input_2'), model.output)

However, I get the following error upon loading:

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("input_foo_2:0", shape=(None, 64), dtype=float32) at layer "input_foo". The following previous layers were accessed without issue: []

Answer 1

Just create a new model with input2 and output and load the weights from the trained model. or try to write a subclassing model by utilizing the training parameter in the call method as below.

def customModel(tf.kears.models.Model):
    def call(inputs, training=True):
        if training == True:
            #do your model for training
        else:
            #do your model for inference