Inverse of torch_geometric.transforms.NormalizeFeatures()

[CodingMaker2]

How do I restore the converted data of the NormalizeFeatures() function to the original data?

[rusty1s]

This is not possible without storing more information, as you cannot go from sum->mean and then back from mean->sum without knowing the factor you how to multiply with.

[CodingMaker2]

for store in data.stores:
for key, value in store.items(*self.attrs):
value = value - value.min()
value.div_(value.sum(dim=-1, keepdim=True).clamp_(min=1.))
store[key] = value
return data

I have noticed that there is a normalization process in the source code. Is it possible to use the 'value. sum (dim=- 1, keepdim=True)' in this code as the factor to multiple with?

[rusty1s]

Yes, that's what you would need to do in this case.

[CodingMaker2]

Thanks for your answer, I will have a try.