I am taking a course on energy-aware embedded systems. I'm proposing a project that will measure and compare the performance of CPU and GPU implementations of NNs. During research, I found Espresso could be a really great point for me to start, because it's really hard to find naive and easily understandable implementations in production systems (like TensorFlow Mobile). Espresso is a minimal system that includes both naive CPU and GPU implementations of components of a CNN. I think I could contribute to the project in such steps:
Port the project to Swift 3;
Implement a new network (maybe Inception) using the layers;
Create a report on comparison of energy impact of different implementations.
Would you allow me to use your project as a starting point and do you think my proposal is doable?
Thanks for your work, it is quite interesting and exciting.
But on your home page, I only found the performance result compare with naive implementation, have you guys compared your performance with tensorflow ?