This is a PyTorch implementation of the IJCNN 2019 paper:
The contributions of this paper are summarized as follows.
- We present TCP as a unified approach for accelerating deep unsupervised domain adaptation models. TCP is a generic, accurate, and efficient compression method that can be easily implemented by most deep learning libraries.
- TCP is able to reduce negative transfer by considering the cross-domain distribution discrepancy using the proposed Transfer Channel Evaluation module.
- Extensive experiments on two public UDA datasets demonstrate the significant superiority of TCP.