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)
The code is released for academic research use only. For commercial use, please contact Dr. Chu-Song Chen(song@iis.sinica.edu.tw).
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/
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.
Check out zipper_multiple_train.sh
Excute training in zippering process.
Check out zipper_eval_script.sh
Evaluate the Top-1 accuracy.
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}
}
Please feel free to leave suggestions or comments to Cheng-En Wu(chengen@iis.sinica.edu.tw) , Yi-Ming Chan(yiming@iis.sinica.edu.tw), Chu-Song Chen(song@iis.sinica.edu.tw)