GradCam implementation

[LukeWood]

https://keras.io/examples/vision/grad_cam/

The call signature should be something like:

overlaid_result_images = gradcam(model, images, classes)

[innat]

Relevant issue: #53

Under which directory the visualization tools will be live?

@MrinalTyagi you may like to contribute this. keras-team/keras#15704

[MrinalTyagi]

I feel like creating a new directory named visualization would be a good option as it would be a separate section. If that's not the case adding it to utils can also be a good option. Would love to hear others' thoughts as well on the topic. I will start looking into how to make the feature and share a draft pr soon.

[innat]

or, maybe a directory like keras_cv.utils.grad_cam.

[LukeWood]

I think visualization is deserving of its own directory

[MrinalTyagi]

@LukeWood Would you allow me to work on this during GSoC? I have updated the above proposal deck to be more in TensorFlow template format. Do have a look at it and let me know if I can take this whole visualization section during GSoC. Apart from that if you agree or not I would love to work on the same. Thanks for your time.

[lucasdavid]

I would also very much like to contribute in this part, if you need anything @MrinalTyagi.

If it is of any help, I have implemented Gradients, Smooth-Grad, Full-Grad, Guided-Grad, Grad-CAM, Grad-CAM++, and Score-CAM in this repo, although they need a little polish to work with multiple explaining units at once (namely, to use tf.batch_jacobian instead of tf.gradient).

[bhack]

@lucasdavid I think you could start somewhere with a PR.
as @MrinalTyagi has still #175 under early review.

[innat]

@lucasdavid Here is one reference, https://github.com/keisen/tf-keras-vis. It may help.

[LukeWood]

@LukeWood Would you allow me to work on this during GSoC? I have updated the above proposal deck to be more in TensorFlow template format. Do have a look at it and let me know if I can take this whole visualization section during GSoC. Apart from that if you agree or not I would love to work on the same. Thanks for your time.

Thanks for the proposal, we really appreciate it @MrinalTyagi. There are many visualization tools available that we can write and implement, so even if one is picked up (GradCam) by a contributor we can still find others to keep adding. there are many with unique value. Thanks for your understanding.

[LukeWood]

I would also very much like to contribute in this part, if you need anything @MrinalTyagi.

If it is of any help, I have implemented Gradients, Smooth-Grad, Full-Grad, Guided-Grad, Grad-CAM, Grad-CAM++, and Score-CAM in this repo, although they need a little polish to work with multiple explaining units at once (namely, to use tf.batch_jacobian instead of tf.gradient).

@lucasdavid I opened unique github issues for each of these components. I'd love to begin collecting visualization components that we would like to add. Thanks a lot for your help