Skip to content

jinhangchoi/approxrep_learning

Repository files navigation

Approximated front-end representation-based Learning

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

About

Heuristic Approximation of Early-Stage CNN Data Representation

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors