Machine learning workflow optimization via automatic discovery of resource reuse opportunities
Jialin Liu · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2019
Many state-of-the-art deep learning models rely on dynamic computation logic, making them difficult to optimize. In this thesis, we present a hashing based algorithm that is able to detect and optimize computation logic common to different computation graphs. We show that our algorithm can be integrated seamlessly into popular deep learning frameworks such as TensorFlow, with nearly zero code changes required on the part of users in order to adapt our optimizations to their programs. Experiments show that our algorithm achieves 1.35× speedup on a sentiment classification task trained with the popular Tree-LSTM model.