moboehle/pytorch-lrp

Basic LRP implementation in PyTorch

pytorch-lrp's Issues

How to get "holdout_AD_CN_2Yr15T_plus_UniqueScreening_ID_Hippocampus.h5"

Could you please supply some detail on how to get the "holdout_AD_CN_2Yr15T_plus_UniqueScreening_ID_Hippocampus.h5" file, which is in "Beta Correlation.ipynb" script? Thank you very much.

RuntimeError: shape '[1, 2048, 10, 10]' is invalid for input of size 2048

I'm trying to analyze FaceForensics detector XceptionNet with your LRP implementation, but I get stuck when trying to analyze the first Conv2d layer (the last conv in the network) in the backpass. I'm getting this error message when trying to reshape relevance_in in method conv_nd_inverse (file: inverter_util.py):

relevance_in = relevance_in.view(m.out_shape)
RuntimeError: shape '[1, 2048, 10, 10]' is invalid for input of size 2048

So relevance_in has size torch.Size([1, 2048]) and m.out_shape is [1, 2048, 10, 10]. Why am I getting this error?

Question about skip connections.

I see that you base on sequential relevance propagation

Pytorch-LRP/innvestigator.py

Line 212 in 3263df4

for layer in rev_model:

but isn't it not acceptable for skip connection, where relevance should come from residual and identity branch?

NameError: name 'INSERT' is not defined

When I tried to run this code in colab notebook , it occured the problem as "NameError: name 'INSERT' is not defined" during the process of importing setting.py. I wonder the reason behind it because I couldn't find corresponding useful information on the Internet. Thanks.

Question inverter_util.py

In compute_propagated_relevance function,

Do you only support max_pooling, not avg_pool or adaptive_avg_pooling??

请问一下数据集下载的是那个呢？

数据申请通过了，但是不知道具体下载那个数据集

Version of Pytorch and Python

Hi,

Thanks for the awesome repo. However, I am getting some syntax errors and I guess it might be due to difference of your Pytorch and Python. Can you kindly comment on

1. Version of Pytorch
2. Version of Python

For example I am getting 'syntax error' in the following line:
https://github.com/moboehle/Pytorch-LRP/blob/master/inverter_util.py#L207

Keenly looking forward to your response.

Thanks,
Avisek

Question about unet with upsample or ConvTranspose2d layer support

I want to use LRP to explain the semantic segmentation task using Unet model (Pytorch). I tested the LRP in captum but not support nn.Upsample and nn.ConvTranspose2d. I would like to know if the semantic segmentation model like Unet can be supported, and if not, how should it be implemented? Any help would be appreciated!

when i use lrp in maxpool, the sum of relevance will be changed.

I print the sum of relevance, then I found that it will be changed in maxpool layer.

potential bug in Ln 336 in inverter_util.py

When I tested the b-rule, an error raised saying sum_weights is not defined. I think it could be a bug. I removed this line and could run correctly.