BottleFit for Split Computing

The official implementations of BottleFit for ILSVRC 2012 (ImageNet) dataset: "BottleFit: Learning Compressed Representations in Deep Neural Networks for Effective and Efficient Split Computing," IEEE WoWMoM '22
[Paper] [Preprint]

More advanced work "SC2 Benchmark: Supervised Compression for Split Computing" and code are available at https://github.com/yoshitomo-matsubara/sc2-benchmark

Citations

@inproceedings{matsubara2022bottlefit,
  title={{BottleFit: Learning Compressed Representations in Deep Neural Networks for Effective and Efficient Split Computing}},
  author={Matsubara, Yoshitomo and Callegaro, Davide and Singh, Sameer and Levorato, Marco and Restuccia, Francesco},
  booktitle={2022 IEEE 23rd International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM)}, 
  pages={337-346},
  year={2022}
}

Requirements

Python >=3.6
pipenv

How to clone

git clone https://github.com/yoshitomo-matsubara/bottlefit-split_computing.git
cd bottlefit-split_computing/
pipenv install

Download datasets

As the terms of use do not allow to distribute the URLs, you will have to create an account here to get the URLs, and replace ${TRAIN_DATASET_URL} and ${VAL_DATASET_URL} with them.

wget ${TRAIN_DATASET_URL} ./
wget ${VAL_DATASET_URL} ./

ILSVRC 2012 (ImageNet) dataset

# Go to home directory
mkdir ~/dataset/ilsvrc2012/{train,val} -p
mv ILSVRC2012_img_train.tar ~/dataset/ilsvrc2012/train/
mv ILSVRC2012_img_val.tar ~/dataset/ilsvrc2012/val/
cd ~/dataset/ilsvrc2012/train/
tar -xvf ILSVRC2012_img_train.tar
for f in *.tar; do
  d=`basename $f .tar`
  mkdir $d
  (cd $d && tar xf ../$f)
done
rm -r *.tar

wget https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
mv valprep.sh ~/dataset/ilsvrc2012/val/
cd ~/dataset/ilsvrc2012/val/
tar -xvf ILSVRC2012_img_val.tar
sh valprep.sh

Trained models

Baseline methods

Our baseline results (Vanilla, KD, HND, and Autoencoder) and trained model weights are available at https://github.com/yoshitomo-matsubara/head-network-distillation

Proposed methods

Create a directory and download the checkpoints of model weights (densenet.zip, resnet.zip, resnet-other_vers.zip, resnet-other_methods.zip)

mkdir -p resource/ckpt

Unzip the downloaded zip files under ./resource/ckpt/, then there will be ./resource/ckpt/image_classification/.

Test trained models

Taking 3ch-bottleneck-injected models as examples

GHND-KD

pipenv run python image_classification.py --config config/ghnd_kd/custom_densenet169_from_densenet169-3ch -test_only
pipenv run python image_classification.py --config config/ghnd_kd/custom_densenet201_from_densenet201-3ch.yaml -test_only
pipenv run python image_classification.py --config config/ghnd_kd/custom_resnet152_from_resnet152-3ch.yaml -test_only

GHND-KD (FE)

pipenv run python image_classification.py --config config/ghnd_kd-finetune/custom_densenet169_from_densenet169-3ch -test_only
pipenv run python image_classification.py --config config/ghnd_kd-finetune/custom_densenet201_from_densenet201-3ch.yaml -test_only
pipenv run python image_classification.py --config config/ghnd_kd-finetune/custom_resnet152_from_resnet152-3ch.yaml -test_only

GHND-FT

pipenv run python image_classification.py --config config/ghnd_vanilla/custom_densenet169_from_densenet169-3ch.yaml -test_only
pipenv run python image_classification.py --config config/ghnd_vanilla/custom_densenet201_from_densenet201-3ch.yaml -test_only
pipenv run python image_classification.py --config config/ghnd_vanilla/custom_resnet152_from_resnet152-3ch.yaml -test_only

GHND-FT (FE)

pipenv run python image_classification.py --config config/ghnd_vanilla-finetune/custom_densenet169_from_densenet169-3ch.yaml -test_only
pipenv run python image_classification.py --config config/ghnd_vanilla-finetune/custom_densenet201_from_densenet201-3ch.yaml -test_only
pipenv run python image_classification.py --config config/ghnd_vanilla-finetune/custom_resnet152_from_resnet152-3ch.yaml -test_only

Train models

If you would like to train models, you should exclude -test_only from the above commands, and set new file paths for student model in the yaml files.
To enable the distributed training mode, you should use pipenv run python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} --use_env image_classification.py ... --world_size ${NUM_GPUS}