This repository presents source code necessary to reproduce training/inference addressed in the following paper:
@INPROCEEDINGS{Choi2018ICCD_ApproxRepLearning,
title={{Heuristic Approximation of Early-Stage CNN Data Representation for Vision Intelligence Systems}},
author={Jinhang Choi and Jack Sampson and Vijaykrishnan Narayanan},
booktitle={IEEE 36th International Conference on Computer Design (ICCD)},
year={2018},
pages={218-225},
doi={10.1109/ICCD.2018.00041},
ISSN={2576-6996},
month=Oct
}
This repository is composed of the following items:
- notebook/
- Experimental reports
- android-cmake/
- Android NDK compile setup
- caffe-android-demo/
- Android App for testing image classification using Caffe Library
- caffe-android-lib/
- Caffe JNI Library for Android Marshmallow on armv7a uarch
- data/
- ilsvrc2012 synset id for training dataset, validation dataset, and inference tests
- ilsvrc2012 mean substraction for raw images
- models/
- BVLC CaffeNet (AlexNet) - Reference / HOG+modified version / Gabor+modified version
- SqueezeNet - Reference / HOG+modified version / Gabor+modified version
- BVLC GoogLeNet - Reference / HOG+modified version / Gabor+modified version
- It is also possible to train CNNs by using prototxt descriptors
- scripts/
- classify.alexnet.gabor.py3 - inference test for gabor filters by reading LMDB (Gabor feature LMDB)
- classify.alexnet.hog.py3 - inference test for hog by reading LMDB (either original raw image LMDB or HOG feature LMDB)
- cvtlmdb_img2gabor.py - LMDB converter from raw images to Gabor features
- cvtlmdb_img2hog.py - LMDB converter from raw images to HOG features
- cvtlmdb_img2resize.py - LMDB converter by resizing raw images