How to use the Image explainer

[SelComputas]

Describe the bug
A clear and concise description of what the bug is.

I am unable to understand how to use this program to run my CNN model through it, in order to know which pixels are the essential ones.

To Reproduce
Steps to reproduce the behavior:

1. Follow your "Getting Started"
2. Open "notebooks/explain-multiclass-classification-local.ipynb"
3. Look at the image;
   
4. Try to understand how one is supposed to use the Image Kernel mentioned in the image.

Expected behavior
I expect to have a clear and concise way of using your tool in order to obtain some explanation as to what my CNN model is looking at when making predictions.

Screenshots
If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: Windows
• Browser: Chrome
• Version

Additional context
I feel like I am just missing the part of your README-files where you explain how to use the Image Explainer mentioned in the image above.

I can also note that I have a .pb (TensorFlow) model, and a .onnx model.

[imatiach-msft]

@SelComputas currently there is an image explainer in the contrib package azureml-contrib-explain-model package but I wouldn't recommend it, as there is no visualization dashboard for it yet. We have an https://github.com/interpretml/interpret-text repository for text interpretability and logically we would like to add a repository for images eventually, but unfortunately we don't have one yet. Sorry, we should update that image in the notebooks as it isn't accurate anymore, we open-sourced parts of a previous package as this package which has been made an extension to the interpret package.

[pk2005]

Hi! Is there any update on this? I have similar issue and wanted to use InterpretML for my CNN. Thanks!

[imatiach-msft]

@pk2005 currently we don't have image support yet, sorry. However, there is a hierarchical image explainer available in SHAP github repository (https://github.com/slundberg/shap) which works well for images. If you are using pytorch, captum also has great image explainers: https://github.com/pytorch/captum. I'm not sure what is available for tensorflow, I just found this example using integrated gradients with a quick search, there might be better toolkits available:
https://www.tensorflow.org/tutorials/interpretability/integrated_gradients