chainer-desalinet

Chainer implementation of DeConvNet [1] and DeSaliNet [2]. These are the technics that visualize feature maps of each layers of convolutional networks. Currently, this implementation supports VGGNet and AlexNet.

Summary

	Input	conv5	fc8
DeConvNet
DeSaliNet

Requirements

• Chainer
• CuPy (for GPU support)
• OpenCV

Usage

DeConvNet

python run.py --input images/bird.png --arch alex --gpu 0
python run.py --input images/bird.png --arch vgg --gpu 0

DeSaliNet

python run.py --input images/bird.png --arch alex --mask --gpu 0
python run.py --input images/bird.png --arch vgg --mask --gpu 0

References

• [1] Matthew D. Zeiler and Rob Fergus, "Visualizing and Understanding Convolutional Networks", https://cs.nyu.edu/~fergus/papers/zeilerECCV2014.pdf
• [2] Aravindh Mahendran and Andrea Vedaldi, "Salient deconvolutional networks", https://www.robots.ox.ac.uk/~vedaldi/assets/pubs/mahendran16salient.pdf