Duplicating pixels, DataLoader, pooling functions

[vsainteuf]

Duplicating pixels, DataLoader, pooling functions

[VSainteuf]

Hi again,
I remember I once did a quick test with PyTorch Scatter to handle that more elegantly. But I didn't push much further because I didn't see noticeable gains. I think that the amount of unnecessary computation is quite limited.

[LouisStef]

I am trying to remove the mask to connect the whole model to explainability frameworks (Shap, Captum ...), which usually expect a list of tensors as inputs (our case) but all tensors are interpreted as features (the mask is not one)

I think Torch geometric's DataLoader is an promising candidate, bc it can create batches of graphes which possibly contain different number of nodes ... (A pixel set is a graph without edges)

Will see if I find something interesting !