2020-ECCV oral-MutualNet Adaptive ConvNet via Mutual Learning from Network Width and Resolution
来源：ChenBong 博客园
- Institute：University of North Carolina at Charlotte
- Author：Taojiannan Yang、Chen Chen
- GitHub：https://github.com/taoyang1122/MutualNet 【】
- Citation： 5
Introduction
加入分辨率自适应的动态slim网络
Motivation
Contribution
Method
Framework
Model Training
有效性的一个解释
Final model
- Grid Search，query table
Experiments
ImageNet (compare with: US-Net)
compare with: individual model+multi resolution
Ablation Study
compare with: US-Net + multi resolution
- Grid Search，query table
compare with: Multi-scale Data Augmentation
MobileNet + Multi-scale data augmentation.
In multi-scale data augmentation, the network may take images of different resolutions in different iterations. But within each iteration, the network weights are optimized in the same resolution direction.
While our method randomly samples four sub-networks which share weights with each other.
the weights are optimized in a mixed resolution direction in each iteration.
US-Net + Multi-scale data augmentation
we randomly choose a scale from {224, 192, 160, 128} and feed the same scaled image to all sub-networks in each iteration.
the weights are still optimized towards a single resolution direction in each iteration, but the direction varies among different iterations.
Effects of Width Lower Bound
Boosting Single Network Performance
As discussed above, the performance of the full-network is greatly improved as we increase the width lower bound.
Conclusion
Summary
- 本质上就是在US-Net加了多个分辨率，并对不同规模的网络做了KD进行训练
- 实验比较充分