Merging-MobileNets-for-Multitask

Official implementation of On Merging MobileNets for Efficient Multitask Inference.

Created by Cheng-En Wu , Yi-Ming Chan(yiming@iis.sinica.edu.tw), Chu-Song Chen(song@iis.sinica.edu.tw)

Usage

Pretrained and merged checkpoints are available here: https://reurl.cc/96o3x

Datasets in TFRecords are available here: https://reurl.cc/vp3vN
All rights belong to the respective publishers. The datasets are provided only to aid reproducibility.

Place and unzip pretrained checkpoints checkpoints.tar.gz in checkpoints/ , datasets.tar.gz in datasets/ hungarian_algorithm.tar.gz in hungarian_algorithm/

Mergeing

Check out convert_ckpt_to_npy.sh and merge_layers.sh

Convert pretrained checkpoints to numpy files and merge pretrained checkpoints into a unified checkpoints.

Training

Check out zipper_multiple_train.sh

Excute training in zippering process.

Inference

Check out zipper_eval_script.sh

Evaluate the Top-1 accuracy.

Citation

Please cite following paper if these codes help your research:

@inproceedings{wu2019mobilenet_merging,
  Title   = {On Merging MobileNets for Efficient Multitask Inference},
  Author  = {Cheng-En Wu, Yi-Ming Chan and Chu-Song Chen}, 
  booktitle = {2019 IEEE International Symposium on High-Performance Computer Architecture Workshop},
  year    = {2019}
}

Contact

Please feel free to leave suggestions or comments to Cheng-En Wu , Yi-Ming Chan(yiming@iis.sinica.edu.tw), Chu-Song Chen(song@iis.sinica.edu.tw)