Abstract
In this paper, we seek to tackle a challenge in training low-precision networks: the notorious difficulty in propagating gradient through a low-precision network due to the non-differentiable quantization function. We propose a solution by training the low-precision network with a fullprecision auxiliary module. Specifically, during training, we construct a mix-precision network by augmenting the original low-precision network with the full precision auxiliary module. Then the augmented mix-precision network and the low-precision network are jointly optimized. This strategy creates additional full-precision routes to update the parameters of the low-precision model, thus making the gradient back-propagates more easily. At the inference time, we discard the auxiliary module without introducing any computational complexity to the low-precision network. We evaluate the proposed method on image classification and object detection over various quantization approaches and show consistent performance increase. In particular, we achieve near lossless performance to the full-precision model by using a 4-bit detector, which is of great practical value.
Original language | English |
---|---|
Title of host publication | Proceedings - 33th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2020 |
Editors | Ce Liu, Greg Mori, Kate Saenko, Silvio Savarese |
Place of Publication | Piscataway NJ USA |
Publisher | IEEE, Institute of Electrical and Electronics Engineers |
Pages | 1488-1497 |
Number of pages | 10 |
ISBN (Electronic) | 9781728171685 |
ISBN (Print) | 9781728171692 |
Publication status | Published - 2020 |
Externally published | Yes |
Event | IEEE Conference on Computer Vision and Pattern Recognition 2020 - Virtual, China Duration: 14 Jun 2020 → 19 Jun 2020 http://cvpr2020.thecvf.com (Website ) https://openaccess.thecvf.com/CVPR2020 (Proceedings) https://ieeexplore.ieee.org/xpl/conhome/9142308/proceeding (Proceedings) |
Conference
Conference | IEEE Conference on Computer Vision and Pattern Recognition 2020 |
---|---|
Abbreviated title | CVPR 2020 |
Country/Territory | China |
City | Virtual |
Period | 14/06/20 → 19/06/20 |
Internet address |
|