Appendix of PyGlove

Appendix A provides a formal definition of symbolic programs in our method, including symbolic counterparts of different program constructs, supported operations, and the description of algorithms used in the materialization process. Appendix B gives a more detailed introduction to PyGlove – our implementation of the method – with an example of dropping neural architecture search (NAS) into an existing Tensorflow program (MNIST [1]). Appendix C provides additional information for experiments used in our case studies, including experiment setup, source code for creating search spaces and complex search flows.