All-Deformable-Butterfly Network

This repository serves as the official code release for the IEEE TNNLS paper titled: Lite It Fly: An All-Deformable-Butterfly Network.

Rui Lin*, Jason Chun Lok Li*, Jiajun Zhou, Binxiao Huang, Jie Ran, Ngai Wong

(*Equal contribution)

Running

This repository has been tested with Ubuntu 20.04.1 LTS, Python 3.8, Pytorch 1.10.1 and CUDA 11.3.

1. Automated Chain Generation & MNIST:

All codes regarding automated chain generation and MNIST experiment shown in the paper is under the AutoChain/. To run the experiment, first download the MNIST and save it to the data/mnist_data, and then run:

python ./AutoChain/main.py

2. PointNet & ModelNet40:

First download the ModelNet40 dataset here and save it to data/modelnet40_normal_resampled/.

We prepare some simple bash scripts, to train a teacher model simply run:

./scripts/run_pointnet_vanilla.sh

To train a student model using CRD framework, run:

./scripts/run_pointnet_distill.sh

To evaluate a saved checkpoint:

python test_pointnet.py \
--model_path [path to the saved checkpoint] \
--r_shape_txt [path to .txt files specifing the structure of debut chains] \

Some useful flags to know:

--path_t: The path to the teacher's checkpoint
--distill: Select the distillation method to use
--model_s: Select the type of student model (i.e SVD, Butterfly, Fastfood, DeBut)
-a: Balancing weight for KD Loss
-b: Balancing weight for CRD Loss
--r_shape_txt: The path to .txt files describing the shapes of the factors in the given monotonic or bulging DeBut chains 

3. CIFAR-100:

First download the CIFAR-100 dataset and save it to data/cifar-100-python.

Similarly, to train a teacher model simply run:

./scripts/run_[vgg/resnet]_vanilla_cifar100.sh

To train a student model using CRD framework, run:

./scripts/run_[vgg/resnet]_distill_cifar100.sh

To evaluate a saved checkpoint:

python test.py \
--model_path [path to the saved checkpoint] \
--r_shape_txt [path to .txt files specifing the structure of debut chains] \

Citation

If you find All-DeBut useful for your research and applications, please consider citing it using this BibTeX:

@article{lin2023lite,
  title={Lite It Fly: An All-Deformable-Butterfly Network},
  author={Lin, Rui and Li, Jason Chun Lok and Zhou, Jiajun and Huang, Binxiao and Ran, Jie and Wong, Ngai},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2023},
  publisher={IEEE}
}

Acknowledgements

Our codes are adapted from official released codes for CRD by Yonglong Tian et al. and Pytorch implementation of PointNet by Xu Yan.