Keras: compile() should not require arguments when not training

edit: A less obscure situation: after loading a model from disk, it's necessary to call compile() with optimizer and loss, even if you're only using it for prediction. A more obscure situation follows.

I'm using the new functional API to predict intermediate output from my network. First I set up the network and compile the model:

inputs = Input(shape=(1, img_rows, img_cols))
x = Convolution2D(32, 3, 3, activation='relu')(x)
...
x = Flatten()(x)
e = Dense(512, name='e', activation='relu')(x)
x = Dense(nb_classes, activation='softmax')(e)

model = Model(input=inputs, output=x)
model.compile(optimizer='rmsprop', loss='categorical_crossentropy')

Then I train it. After it's done training, I want to evaluate the output at e.

Previously with the Sequential model I did something like this:

def matching_layers(model, name):
    return [l for l in model.layers if l.name == name]
get_encoding = K.function([model.layers[0].input, K.learning_phase()],
                          [matching_layers(model, 'e')[0].output])
encoded = get_encoding([data, 1])[0]

That worked correctly, but I usually have to split the data into chunks and then recombine it because I run out of GPU memory otherwise.

Now I can do something much simpler:

emodel = Model(input=inputs, output=e)
emodel.compile(optimizer='rmsprop', loss='categorical_crossentropy')
encoded = emodel.predict(data)

Awesome. It's less code, and makes more sense. And for some reason I don't have to split up the data (not sure what's different behind the scenes...)

But why do I need to specify optimizer and loss if I'm not training the model?