This repository contains source code for Part 1 & 2 of our project which is based on improving Activation based techniques
In this research, we implemented latest ABM Techniques on VGG16 classifier and also on state-of-the-art classifier InceptionV3 and Xception and compared its results. We also performed 3 experiments to improve the saliency map by Activation based methods(ABM) such as Grad-CAM, Grad-CAM++, Smooth Grad-CAM++ by using a novel combination with Backpropagation method called as Integrated Gradients
P.S: The main source code for Part1_CamOnDiffClassifiers and Part2_NovelCombination is added as a jupyter notebook for ease of visualization of results. Each cell takes few minutes to run.
Part 1:
- New implementation of Grad-CAM Grad-CAM++ & Smooth Grad-CAM++ in VGG16
- Extending the above techniques for Xception & InceptionV3
Part 2:
- Novel Combination of Integrated Gradients and CAM techniques on VGG16
- Google Colab with 25 GB RAM
- GPU is not mandatory
- Refer requirements.txt
- Create environment with requirements.txt
- Make sure you have hardware requirements
- Download the folder
- Open code/Part1_CamOnDiffClassifiers.ipynb to run Part 1 or code/Part2_NovelCombination.ipynb to run Part 2
- If running on Google Colab, upload this project, mount your Google drive by uncommenting the first cell
- Change the base directory to reflect your local file structure
- Run each cell in sequential order
- Grad-CAM https://arxiv.org/abs/1610.02391
- Grad-CAM ++ https://arxiv.org/abs/1710.11063
- Smooth Grad-CAM ++ https://arxiv.org/abs/1908.01224
- Integrated Gradients https://arxiv.org/abs/1703.01365