To improve performance of a numerical library such as Owl, it is necessary to support multiple hardware platforms. One idea is to “freeride” existing libraries that already support various hardware platforms. We believe that computation graph is a suitable IR to achieve interoperability between different libraries. Along this line, we develop a prototype converter system by using which the users can define a computation in Owl and then run it on TensorFlow. In this talk, we use an example to show the system workflow, and how powerful features of Owl, such as algorithmic differentiation, can be used in TensorFlow. We then briefly introduce system design and implementation before the closing remark on related and future work.
Fri 23 Aug
|10:30 - 10:55|
|10:55 - 11:20|
|11:20 - 11:45|
Jianxin ZhaoUniversity of Cambridge