Implementation of Grad CAM in tensorflow
Gradient class activation maps are a visualization technique for deep learning networks.
The original paper: https://arxiv.org/pdf/1610.02391v1.pdf
The original torch implementation: https://github.com/ramprs/grad-cam
Clone the repository
git clone https://github.com/Ankush96/grad-cam.tensorflow/Download the VGG16 weights from https://www.cs.toronto.edu/~frossard/vgg16/vgg16_weights.npz
python main.py --input laska.png --output laska_save.png --layer_name pool5| Input | Output |
|---|---|
 |
 |
Model weights (vgg16_weights.npz), Class names (imagenet_classes.py) and example input (laska.png) were copied from this blog by Davi Frossard (https://www.cs.toronto.edu/~frossard/post/vgg16/). TensorFlow model of vgg (vgg16.py) was taken from the same blog but was modified a little. https://github.com/jacobgil/keras-grad-cam also provided key insights into understanding the algorithm.