Hi! I was wandering if you could offer some tips on how to extend your code so that Conv3d works for LRP as well.

Could you please suggest which methods/wrappers should be implemented (should I simply make a wrapper similar to conv2d_beta0_wrapper_fct?), or perhaps you already have a prototype Conv3d solution lying around somewhere? The code here is nicely commented, but it is still kinda tough to navigate without much comments

Sorry for any inconvenience, and thanks!

Seul
Thanks for your contributions.
I'm beginner on explaining dnn. it seems that your project for dnn pruning is now unfinished , and it can't reproduce the paper's result.
here is my findings

missing dataset for scene 15, event 8 ...
missing transferred model saving switch
failed argparse for args.relevance, args.resume....
a raised problem , missing self.lrp()

mismatched shape in designing resnet18, here s the code

        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = F.avg_pool2d(out, 4)
        out = out.view(out.size(0), -1)
        out = self.linear(out)

compared to common resnet18,

1. it s lack of maxpool that should apply to the out1
2. the shape of tensor before flatten layer is 25088 , not 512 required, and the avgpool should avg to (1*1)

missing figure printer for result analysis for acc ...
missing pruning for a limited number of samples without fine-tuning. The original code prunes model using the entire training set.
missing dense layer pruning. orginal code works by pruning the conv layer

i have fixed some basical problem . hope this page gives some advice for others
it helps a lot if you provide after work.

Hi Seul-Ki,

Thank you for your great work!

When I ran your code, I faced this bug, when you have time, can you have a look?

"prune_vgg.py", line 407, in prune
R_tot, data_tot, time_tot = self.lrp() # lrp using conventional model
AttributeError: 'PruningFineTuner' object has no attribute 'lrp'

Thanks a lot!

I have been unable to replicate your results from table 3 in your paper, are you willing to supply the weights of the models prior to pruning

also could you supply the version of pytorch/torchvision you had used originally