Which layer to choose when using gradCAM

[barisic-fzi]

I am not sure which layer should be selected when using gradCAM.

This is my model description (from a YOLOv3 example - repo: https://github.com/YunYang1994/TensorFlow2.0-Examples/tree/master/4-Object_Detection/YOLOV3)

https://www.codepile.net/pile/mqk0b0yQ

If it is better I can also post the model description directly, but i think it looks messy like that.

97 is related, if that would be added, someone casually using this wouldn't have to know, although it would still be cool that it is mentioned somewhere while starting execution and also some comments on this in documentation.

Any help would be great!!!

[RaphaelMeudec]

@barisic-fzi I'll try to integrate #97 as soon as possible. Target layer should be the last convolutional layer (according to the paper). The intuition is that the output filters of the last convolution are the features for the dense layers. That's why it's selected

[RaphaelMeudec]

@barisic-fzi Layer is now automatically detected so that won't be a problem anymore

[barisic-fzi]

will try it out, many thanks!!!

[RaphaelMeudec]

There should be a release this week, you can try it out from master at the moment.

[barisic-fzi]

@RaphaelMeudec tried on a basic example, works good :+1: didn't find that it says which layer is selected, that would be cool if it's not already implemented

[gururajaramesh]

@RaphaelMeudec Can you please share the link to the repository you are referring to? I am trying to implement GradCAM for Yolov3 as well.

[RaphaelMeudec]

@gururajaramesh I'm not sure which reference you're talking about? I'm not pointing to any repository (except this one)?