Updates:
I retrained the model with nn.ReLU() instead of custom swish functions. The result is very similar.
Taking a look at this other post in the pytorch forum one can see that their result for GuidedGradCam is more informative.
As it can be seen in my code, the in which I plot the image is via first converting it to a PIL image. Has anyone used Captum’s GuidedGradCam on a CNN that might be able to shed some light into why I’m having this output?