The project implements SqueezeNet on Android phones. We use RenderScript efficiently to minimize the execution time. I have tested the project on three different Android phones (Nexus 5, Nexus 6P and Samsung Galaxy). The achieved speedups on these platforms are 310X, 133.89X, and 59.54X respectively. The results are submitted to a conference as a paper which is under review.
As the documentation overlaps with the content of the paper, we will publish them when we get the result of submission. However, if you are interested in working on the project, please contact me for more information.
If this project helped your research, please kindly cite our work.
Motamedi, Mohammad, Daniel Fong, and Soheil Ghiasi. "Cappuccino: Efficient Inference Software Synthesis for Mobile System-on-Chips." arXiv preprint arXiv:1707.02647 (2017).