Not to update part of the parameters with backpropagation

autograd
1

I am creating my own nn.Module with its own forward function and matrix of parameters, P. But, I would like to impose that part of the matrix of parameters is set to zero and remains zero. Is there a way to do so without defining the matrix using multiple parameters definition?

class Model(nn.Module):
    def __init__(self, n_objects):
        super(Model, self).__init__()

        self.n_objects = n_objects
        P = torch.empty(

            (self.n_objects, self.n_objects),
            requires_grad=True)
        nn.init.uniform_(P)

        self.P = torch.nn.Parameter(P)

Here, for example, I would like that all the elements in the main diagonal of P should not change w.r.t. the original values.
Thanks for the help!