Christopher

@rachtibat Have a look at the supplied code in colab, the UNet implemementation includes BatchNorms, and the SequentialMergeBatchNorm is applied correctly (the BatchNorms in DoubleConv are assigned in the same...

But I just noticed, that in the `Down` Module ```python class Down(nn.Module): '''Downscaling with conv(stride = 2) then double conv''' def __init__(self, in_channels, out_channels): super().__init__() self.pool_conv = nn.Sequential( nn.Conv2d(in_channels=in_channels, out_channels=in_channels,...

Version 0.3.3 is live on pypi, you can just update to check.

Hey Rodrigo, thank you for raising this issue! Just a short message for now, I will investigate this more thoroughly later. I have seen this behavior before with the Gamma...

Hey Rodrigo, thank you for your insights! After investigating the instability, I have verified my previous expectation that this heavily dependent on the gamma rule, which can lead to vanishing...

As community service, here is the output for resnet18 with the standard values, but bias deactivated. The first column after the layer name is the distance from the maximum to...

Hey everyone, coming back to @maxdreyer's suggestion, it is true that the Gamme rule is only defined on positive inputs. This previously was also the case for the ZPlus rule,...

Hey @HeinrichAD thanks for the bump! I have pushed a draft version in #159 and will try to get around finishing this up the following days. I have not yet...

Hey Reduan, I am aware of this and think this is not really a problem, since the tensor you get with each pass is a new instance, and will be...

Hey @mstaczek the truth is, DenseNet is quite tricky with LRP in the sense that the DenseLayers within the DenseBlocks end with a linear layer without an activation, and start...