Skip to content

Latest commit

 

History

History
89 lines (76 loc) · 5.04 KB

README.md

File metadata and controls

89 lines (76 loc) · 5.04 KB
Raw

FrankenSplit

Official Repository for the paper on FrankenSplit: Efficient Neural Feature Compression with Shallow Variational Bottleneck Injection for Mobile Edge Computing

If you find this work interesting, please check the follow up

Setup

  1. Create virtual environment and install dependencies (requirements.txt). 2. Optionally: Install cupy for (significantly) faster cam map generation. The required package depends on your cuda version and you can find more information here.
  2. Download the ImageNet dataset (We cannot directly link to it) and execute the script below
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_train.tar ./
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_val.tar ./
mkdir ~/dataset/ilsvrc2012/{train,val} -p
mv ILSVRC2012_img_train.tar ~/resources/datasets/ilsvrc2012/train/
mv ILSVRC2012_img_val.tar ~/resources/datasets/ilsvrc2012/val/
cd ~/resources/datasets/ilsvrc2012/train/
tar -xvf ILSVRC2012_img_train.tar
mv ILSVRC2012_img_train.tar ../
for f in *.tar; do
  d=`basename $f .tar`
  mkdir $d
  (cd $d && tar xf ../$f)
done
rm -r *.tar
cd ../../../../../

wget https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
mv valprep.sh ~/resources/datasets/ilsvrc2012/val/
cd ~/resources/datasets/ilsvrc2012/val/
tar -xvf ILSVRC2012_img_val.tar
mv ILSVRC2012_img_val.tar ../
sh valprep.sh
mv valprep.sh ../
cd ../../../../../
  1. Set PYTHONPATH as repository root
  2. Optional: For saliency guided distortion create CAM maps: python saliency_maps/cam_prep/cam_map_generation.py --model swin_s3_base_224 --input ~/resources/datasets/ilsvrc2012/train --output resources/dataset/tmp --batch_size 16 --device cuda:0 --mix_layers --saliency_type XGradCAM --eigen_smooth --patch_cam --target_dim 28 4. Set batch size as memory permitting 5. The models are loaded from torch image models (timm). You can replace swin_s3_base_224 with other target models that are present in the timm registry. 6. To pass custom weights for a timm model use --weights_path path/to/weights 7. There are some other options, but they are not relevant to the experiments in the paper
  3. Optional: Download some pre-trained weights from: TODO

Train & Test

Training Setup

To train from scratch: python main_classification_torchdistill.py --config path/to/config --skip_ckpt

To test with pre-trained weights: python main_classification_torchdistill.py --config path/to/config --test_only (make sure that the weights are found)

Notes

  • I've removed most code that was out of scope to include in the paper to avoid confusion but there are still some references to unpublished implementations/results. Specifically, this repository was created by extracing relevant parts into its own Repository.
  • This was the first time I worked on a large PyTorch project and it started with me hacking some ideas together. I've understimated, the scope and times things need to be updated and extended, so the repository is a bit of a mess. I'll clean things up iteratively and include more detailed instructions over time. For now, if you need my assistance with anything, feel free to write me a mail at [email protected]
  • Check out torchdistill (It's awesome!) for documentation on how configurations are loaded and how you can adjust them if you want to perform your own experiments

Citation

[IEEE] [Preprint]

@misc{furutanpey2023frankensplit,
      title={FrankenSplit: Efficient Neural Feature Compression with Shallow Variational Bottleneck Injection for Mobile Edge Computing}, 
      author={Alireza Furutanpey and Philipp Raith and Schahram Dustdar},
      year={2023},
      eprint={2302.10681},
      archivePrefix={arXiv},
      primaryClass={eess.IV}](https://ieeexplore.ieee.org/document/10480247)
}

References

  • Matsubara, Yoshitomo. "torchdistill: A modular, configuration-driven framework for knowledge distillation." Reproducible Research in Pattern Recognition: Third International Workshop, RRPR 2021, Virtual Event, January 11, 2021, Revised Selected Papers. Cham: Springer International Publishing, 2021.
  • Matsubara, Yoshitomo, et al. "SC2: Supervised compression for split computing." arXiv preprint arXiv:2203.08875 ( 2022).
  • Wightman, Ross. "Pytorch image models." (2019).
  • Bégaint, Jean, et al. "Compressai: a pytorch library and evaluation platform for end-to-end compression research." arXiv preprint arXiv:2011.03029 (2020).
  • Gildenblat, Jacob. "contributors. Pytorch library for cam methods." (2021).
  • Ballé, Johannes, et al. "Variational image compression with a scale hyperprior." arXiv preprint arXiv:1802.01436 ( 2018).
  • Minnen, David, Johannes Ballé, and George D. Toderici. "Joint autoregressive and hierarchical priors for learned image compression." Advances in neural information processing systems 31 (2018).