Class Activation Maps

[smaranjitghose]

Create class activation maps to explain the black-box nature of convolutional neural networks. Remember, that this is primarily aimed at customed trained models rather than pre-trained models so that we can input a h5 or saved_model and an image and analyze.

Tasks ( Implement the following):

• [x] GradCAM
• [ ] SmoothGrad
• [x] GradCAM++
• [x] ScoreCAM

Suggestions:

• Please cite all references
• The code should be properly documented for each and every section
• Modularize into functions as much as possible
• Use meaningful docstrings
• Build it as a module cam

[nkpro2000sr]

I would like to work on this :hand:

[smaranjitghose]

I would like to work on this ✋

@nkpro2000sr Could you please start with GRADCAM first and then if you are comfortable, you can proceed with the others

[r0cketr1kky]

I think the recent CAM implementations are taken directly from tf-explain (without citing them) @smaranjitghose could you confirm this?

[smaranjitghose]

I think the recent CAM implementations are taken directly from tf-explain (without citing them) @smaranjitghose could you confirm this?

Well, I had not considered that comparision earlier. But when you pointed out, I recently checked tf-explain, especially this. I could not find much overlap with the one on the repo. However, if you feel there is considerable inspiration or direct use of the same code, do let me know, so that we can rightfully cite them

[r0cketr1kky]

Oh right cool! Although I do remember seeing an implementation that is very very similar to the one on the repo. I'll let you know if I do find it :-)