Switch between training and inference. Training runs data forward, sends the error backward and updates the model; inference only runs forward.
Once the weights are fixed, half the machinery is dead weight. Inference is a forward pass and nothing else.
