sonsongithub/llvmsrc repository overview

⁠Demitasse

Demitasse is the fast Deep Neural Network Library for mobile devices(non GPGPU). You can check the detail of Demitasse from this white paper.

⁠How to build

• for macOS⁠
• for Linux(ubuntu)⁠
• for macOS(日本語)⁠

⁠Build script

• for macOS⁠
• for Linux⁠

⁠How to convert parameters from caffe

It is very boring to make DNN's model data yourself. Demitasse supports to import pre-trained model data from Caffe Model Zoo.⁠ At this moment, we tested VGG-16 and VGG-CNN-F pre-trained model data for Demitasse. Demitasse import tool places at ./tools/. You have to build Demitasse binaries in order to convert pre-trained datas from the other libraries' models into Demitasse's one. How to use the tools is here⁠.

⁠Reference

Please cite Caffe in your publications if it helps your research:

@inproceedings{kondo2016,
author = {Satoshi Kondo},
booktitle = {ML Systems Workshop NIPS 2016},
title = {{Demitasse: SPMD Programing Implementation of Deep Neural Network Library for Mobile Devices}},
year = {2016}
}

⁠License

• This software is released under the License, see LICENSE.txt⁠.
• The sample query image is created from a public domain content⁠.