Is it possible to compute gradients of the attributions with respect to sample

[cys515]

Hello,

My code is as follows:

from captum.attr import Saliency
...... #load test sample, label, and model
sample = sample.requires_grad_(True)  #a Tensor with the size of [1,50,50]
Grad = Saliency(model)
attributions = Grad.attribution(sample,target=label)
attributions = torch.abs(attribution)
min_attr = torch.min(attribution)
max_attr = torch.max(attribution)
attributions = (attributions-min_attr)/(max_attr-min_attr)
y_mesh, x_mesh = torch.meshgrid(torch.arange(50, dtype=torch.float64), torch.arange(50, dtype=torch.float64)) 
mass_center = torch.stack([torch.sum(attributions*x_mesh)/(50*50),torch.sum(attributions*y_mesh)/(50*50)])
mass_center_loss = -torch.sum((mass_center-mass_center0)**2)  #mass_center0 is a consant
temp = -torch.autograd.grad(mass_center_loss, sample)[0]

I would like to calculate the gradient of mass_center_loss with respect to the input sample,
but it was an error: One of the differentiated Tensors appears to not have been used in graph.
I guess this might be the reason why the sample doesn't record the gradient in captum fuction.
So what can I do to preserve the gradient information of the samples in the calculation of attribution?