More clear codes are provided at https://github.com/sseung0703/KD_methods_with_TF.
- Define knowledge by Singular value decomposition
- Fast and efficient learning by multi-task learning
- Tensorflow
- Scipy
Unfortunatly SVD is very slow on GPU. so recommend below installation method.
- Install Tensorflow from source which is removed SVD GPU op.(recommended)
- Install ordinary Tensorflow and make SVD using CPU.
- Install Tensorflow version former than 1.2.
The code is based on Tensorflow-slim example codes. so if you used that it is easy to understand. ( If you want recoded dataset and trained teacher weights download them on https://drive.google.com/open?id=1sNPkr-hOy3qKNUE8bu5qJpvCGFI7VOJN )
- Recording Cifar100 dataset to tfrecording file
- Train teacher network by train_teacher.py, and you can find trained_param.mat in training path.
- Train student network using teacher knowledge which trained step 2 or default knowledge which is ImageNet pretrained VGG16.
This research accepted to ECCV2018 poster session and arxiv version paper is available on https://arxiv.org/abs/1807.06819 and ECCV format paper is available on http://openaccess.thecvf.com/content_ECCV_2018/papers/SEUNG_HYUN_LEE_Self-supervised_Knowledge_Distillation_ECCV_2018_paper.pdf