LRP returns negative values for a simple CNN model.

Hi all,

Thanks for the support of Layer wise propagation. I have a simple CNN model with several convolutional and average pooling layers. From my understanding, LRP should always return positive values and the summation from all pixels should be 1. But I’m getting a negative summation and also lots of negative points. What does the negative value mean? How should I fix it?

Thanks!

class CustomModel(nn.Module):
    
    def __init__(self, model) -> None:
        super(CustomModel, self).__init__()
        self.model = model

    def forward(self, psl):
        psl = psl.reshape(psl.shape[0], psl.shape[1], psl.shape[2], psl.shape[3], n_vars)
        psl = psl[:,:-1]
        out = self.model(psl)
        return out        
custom_model = CustomModel(model)

lrp = LRP(custom_model)

lrps = []
for ind in tqdm(range(test_ds.__len__())):
    sample_lrp = lrp.attribute(test_ds[ind][0].reshape(1, 2, 192, 288, 3).to(device)).to('cpu').data.numpy()
    lrps.append(sample_lrp)

[nanohanno]

That seems to be true, the documentation in appears to be outdated: https://github.com/pytorch/captum/blob/e1575c52f39e403f71cfda412b758df34d857914/captum/attr/_core/lrp.py#L151-L154 Sorry for not being very responsive atm. 😮‍💨

[ShihengDuan]

Hi @bilalsal

Thanks for your reply. I have tried the following but it still gave me a negative attribution.

lrp = LRP(model)
psl = test_ds[ind][0].reshape(1, 2, 192, 288, 3).to(device)
psl = psl.reshape(psl.shape[0], psl.shape[1], psl.shape[2], psl.shape[3], n_vars)
psl = psl[:,:-1]
sample_lrp = lrp.attribute(psl).to('cpu').data.numpy()
print(sample_lrp.shape)
print(np.min(sample_lrp))
print(np.sum(sample_lrp))

Both the min and sum return negative values.