How to extract saliency map from the ImageNet pretrained with non-RBG input?

[N-Jaro]

I created a model with the non-RGB input, trained it, and saved it in h5 format. I want to get the activation and saliency maps when I predict an input image.

Could anyone please advise how I should proceed with this?

Thanks