Which layers should be set to trainable / untrainable? How should last dense layers be structured in accordance with customized classifier needs?
