Demo GradCAM & Guided GradCAM

On Dogs vs. Cat data

Architecture: ResNet50 & ResNet50 + FC layers

An interactive demo for GradCAM and Guided GradCAM, implemented with Tensorflow 2.x

Detailed analysis and training notebook: https://www.kaggle.com/nguyenhoa/dog-cat-classifier-gradcam-with-tensorflow-2-0

Prerequisite

• Python 3.6
• Required packages

bash requirements.txt

Demo

You can run on your own resources with the file Visualization.ipynb

Otherwise, you can run on Google Colab easily via this link: https://colab.research.google.com/github/nguyenhoa93/GradCAM_and_GuidedGradCAM_tf2/blob/master/Visualization.ipynb (Don't forget to change your runtim to Python3 and choose GPU as your hardware accelerator.)

• Model: There are two trained models, which are
  • VanilaResNet50: Keep the same architecture of ResNet50, replace the output layer on ImageNet and re-train with Dog vs. Cat data.
  • ResNet50PlusFC: Add 2 fully connected layers between Average Pooling layer and output layer and train on Dog vs. Cat data.
• Image: There are some available sample images in assets/samples, if you want to run your own ones, put them in this folder to be displayed on the dropdown list.
• Class: This will be the class for GradCAM and Guided GradCAM visualization.

References:

• GradCAM paper: https://arxiv.org/abs/1610.02391
• GradCAM tutorial by pyimagesearch: https://www.pyimagesearch.com/2020/03/09/grad-cam-visualize-class-activation-maps-with-keras-tensorflow-and-deep-learning/
• Keras GradCAM with Tensorflow 1.x backend:
  • https://github.com/jacobgil/keras-grad-cam
  • https://github.com/eclique/keras-gradcam