/DPL_Compression

the official repository for IEEE ICDM 2021 paper

Primary LanguagePython

This is the official repository for IEEE ICDM 2021 paper: Dictionary Pair-based Data-Free Fast Deep Neural Network Compression

Main Process

How to run

  1. add soft link to your dataset folder which should be organized as following: 
    data
 |
 --imagenet
     |
     --ILSVRC2012_img_val.lmdb
     --ILSVRC2012_img_train.lmdb (unused)
 --cifar10
    (or modify the source code to fit your environment)
  2. run script 
    sh run_DPL_Compress.sh
    It will make directories to save results and reconstructed path files. And pre-trained models from PyTorchcv will be downloaded

Noted that we implement evaluation on ImageNet with lmdb, the code of get_imagenet.py should be modified to fit your environment.

Further Improvement

This conference paper is invited for Knowledge and Information Systems (KAIS) publication as Best-ranked paper. New features in KAIS verison are listed as following:

  • Shared dictionary design
  • Auto hyper-parameter tuning
  • Higher compression ratio