search

facebookarchive / caffe2

Caffe2 is a lightweight, modular, and scalable deep learning framework.
https://caffe2.ai
Apache License 2.0
8.42k stars 1.94k forks source link

Transform caffemodel to codes to increase speed? #647

Open XhqGlorry11 opened 7 years ago

XhqGlorry11 commented 7 years ago

Hi all, I have a question to discuss with anyone who is interested in mobile applications. I have trained a caffemodel and the result is good(~35FPS) when I test it with my desktop, which is equipped with Nvidia Titan XP. But when I transform all these work into Nvidia Tx2, it can work but the detection speed is slow (~2FPS). To solve this problem, I get an idea. Since I have got the trained model, I can convert the forward pass into C++/python scripts and get rid of caffe/caffe2 framework. Anyone has ever done same kind of things before and do you think it's workable and can improve detection speed? Any discussion is appreciated. Thx~

futurely commented 7 years ago

Removing unused operators could speed up the model.