In this section, we have implemented the NAS framework for CIFAR-10 data. While this is a great start, there are additional features one can implement, which we will leave to the reader as exercises:
- How can we make the Controller create child networks that solve problems in other domains, such as text and speech recognition?
- How can we make the Controller train multiple child networks in parallel in order to speed up the training process?
- How can we visualize the training process using TensorBoard?
- How can we make the Controller design child networks that include residual connections?
Some of these exercises may require significant changes in the code base but are beneficial for deepening your understanding of NAS. ...