How much does the importance of the input pixel change wrt a certain weight

[AlphaBetaGamma96]

Have a look at this thread to get a general idea of what to do with torch.func

Implementing calculation of the Laplacian

Hi @rfmiotto, You can do this quite efficiently within the torch.func namespace, which requires version torch2.1+. Here’s an example below for the laplacian with respect to an input x below, import torch from torch.func import hessian, vmap, functional_call net = Model(*args, **kwargs) #our model x = torch.randn(batch_size, num_inputs) #inputs params = dict(net.named_parameters()) #params (need to be an input for torch.func to work) #functionalize version (params are now an input to the mod…