Overview

The Tensorflow implementation of Eigen-CAM

Example

Usage

• The example image is generate from mobilenetv2:

python show.py -d cpu -img assets/n01669191_46.JPEG -layer out_relu

• for custom model
  model path is a saved model folder path including keras_metadata.pb.

python show.py -d cpu -img assets/n01669191_46.JPEG -layer {layer name} -m {your model path} -speedup

• Get predict label
  My program will print the predict label, you can refer to show.py.
  For imagenet label to class name, you can refer to class id to name.

Note

• Remenber to check whether the image preprocess is the same as yours, if not, you should alert the preprocess in the show.py or the result will be wrong.
• If you have cuda, you can just replace the "cpu" to "cuda".
• Generally, the layer is deeper, the interpretaton is better. Accordingly, using the last layer before last global average layer to plot heatmap would be a nice choice.

All inplementation of CAM method

Method	Framework
Grad-CAM	Pytorch/Tensorflow
Eigen-CAM	Pytorch/Tensorflow
Ablation-CAM	Tensorflow
Integrated Gradient	Pytorch/Tensorflow