LRP with pretrained ResNet returns weird heatmaps

[razla]

Hey,

I’m trying to use your LRP implementation on a pretrained model resnet34/32 (ImagetNet and CIFAR10/100 respectively).

I needed to convert the ReLU layers to inplace=False in order to make it work and then when I execute this - it works but the output is weird and inconsistent…

these are the images for example:

and these are the explanations (heat maps)

which doesnt make sense, since LRP tends to be MUCH more readable and coherent.

thanks!

[NarineK]

@razla, are you getting most of the LRP scores zero or close to zero or are you visualizing only the positive scores ?
It can be that we need to look into the layer types and the rules applied on them. You might need to readjust used rules define custom ones.

There is a note from @johannesk, that we might need to tinker the rules a little bit for Resnet-18.

#668 (comment)