tensorflow2.0 - Layerwise Relevance Propagation in tensorflow 2.x

Question

I am looking into neural networks and trying to implement LRP for a simple example. I tried to follow the example on: https://git.tu-berlin.de/gmontavon/lrp-tutorial/-/blob/main/tutorial.ipynb for the VGG network. But I want to use it in Tensorflow 2.x instead of Pytorch. Maybe someone out there already tried it successfully and can share their code?

My problem is already at the beginning of the example code:

Example:

model = torchvision.models.vgg16(pretrained=True); model.eval()
layers = list(model._modules['features']) + utils.toconv(list(model._modules['classifier']))
L = len(layers)
A = [X]+[None]*L
for l in range(L): A[l+1] = layers[l].forward(A[l])

My version in tensorflow:

model = vgg16.VGG16()
layers = model.layers 
del layers[0]
L =len(layers)


A = [input] + [None]*L
for l in range(L):
    A[l+1] = layers[l].forward(A[l])

It produces the following error:

'Conv2d' has no attribute 'forward'

question from:https://stackoverflow.com/questions/65884661/layerwise-relevance-propagation-in-tensorflow-2-x

Answer 1

Waitting for answers