This repository is the official implementation of Data-free Knowledge Distillation for Segmentation using Data-Enriching GAN .

(Top row) A deployment pipeline where a large model is initially trained on some dataset. Later for mobile deployment when the model needs to be compressed, the dataset needs to be accessed again causing privacy concerns.

(Bottom row) We present a deployment pipeline where no access to data is required. Instead we use a proxy dataset to generate representative samples to perform model compression.

To install requirements:

pip install pytorch-lightning==0.7.0
pip install tensorboard

To train the DeGAN network

python -m code.segmentation.degan_trainer

To train the student network

python -m code.segmentation.datafree_kd_trainer

Create a issue or a pull request if you wish to contribute :)