kaifishr/pytorchrelevancepropagation

A basic implementation of Layer-wise Relevance Propagation (LRP) in PyTorch.

pytorchrelevancepropagation's Issues

Confusion about Gradient-based Relevance Computation

Could you please explain the formula for Gradient-based Relevance Computation in more detail, I notice you have explained in your blog.
Part code:

z = self.layer.forward(a) + self.eps
        s = (r / z).data
        (z * s).sum().backward()
        c = a.grad
        r = (a * c).data

Part formula:

There are some questions:

1. Why wij can represent by
   , and the i' represents what.
   2.sj is also an equation about a, sj is treated as a constant is not correct.
   3.What is the relationship between zj(a; w) and .

lrp for concatenate layer

Hi @KaiFabi

Look like the lrp layer is worked if the network structure's model is sequential right?. I tried using squeezenet that has FIRE module (contain concatenate layer) got an error channels.

Thanks

lrp for regression

Dear @KaiFabi ,

Thanks for this implementation. I think it is super cool! Can your implementation be used for the regression problem where there is only one output neuron representing a numerical value. How should I change the implementation to do this, if possible?

Looking forward to hearing your comment on this!

Thanks

question in lrp_layers.py

hi @kaifishr
Thanks for your implementation. I'm trying to reimplement lrp on Resnet50, but it has a BatchNorm2D layer in the backbone, I'm a freshman in python and I don't know how to code the RelevancePropagationBatchNorm2D in lrp_layers.py. Can you just give me some ideas？ Thanks a lot.

how to use LRP in regression problems