Explainability is hard to understand

[DrJonnyT]

would be good to show some general plots of what parts of the image in general the model looks at, rather than just on a per-image basis. And some mention of how they work, eg doutput/dinput, how does the output change if we change one pixel?
ScoreCAM goes through the activations maps for the layers and weights them accordingly

Also do not use the penultimate_layer = penultimate_layer_idx part? Seems to work better without it? Possibly...
It goes through to find the layer automatically