Maximilian Dreyer

Hey @bugsuse, I have already used Zennit for a UNet with the `nn.Upsample` layer and as @chr5tphr said, this worked even when just using the gradient (no extra rule). In...

@bugsuse, you can basically use the code that is written in `zennit/share/example/feed_forward.py`. After loading your model and data, you only have to adapt the choice of the output_relevance (the output...

@bugsuse also, you set the output_relevance twice: in the beginning `output_relevance = torch.zeros_like(output)` and in the loop later `output_relevance = torch.zeros_like(torch.squeeze(target))` output_relevance should have the same shape as output. `torch.zeros_like(torch.squeeze(target))`...

@bugsuse using the rule `epsilon_plus_flat` I received a heatmap that was similar to the output mask.

@chr5tphr You are right, that might be a problem! In my implementation, a `nn.MaxPool2d(2)` is used instead of `nn.Conv2d(in_channels=in_channels, out_channels=in_channels, kernel_size=3, stride=2, padding=1)`. @bugsuse thus, another option would also be...

I might add that this could be a problem of the gamma-rule, which does not fit (as it is) to the skip-connections of the ResNet model. Usually, the gamma rule...

I have run a small experiment and making the gamma-rule symmetric (for pos. and neg. output, indicated by *) seems to result in much more reasonable results. The total relevance...