This is the official repository for IEEE ICDM 2021 paper: Dictionary Pair-based Data-Free Fast Deep Neural Network Compression
- add soft link to your dataset folder which should be organized as following:
(or modify the source code to fit your environment)
data | --imagenet | --ILSVRC2012_img_val.lmdb --ILSVRC2012_img_train.lmdb (unused) --cifar10
- run script
It will make directories to save results and reconstructed path files. And pre-trained models from PyTorchcv will be downloaded
sh run_DPL_Compress.sh
Noted that we implement evaluation on ImageNet with lmdb, the code of get_imagenet.py should be modified to fit your environment.
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