loss_b_2 in positional loss has no grad_fn.

[mhjc01]

gnn-lspe/nets/ZINC_graph_regression/gatedgcn_net.py

Line 156 in 6b38443

loss_b_2 = torch.tensor(norm(PTP_In, 'fro')**2).float().to(self.device)

'loss_b_2' has no attributes of grad_fn and requires_grad as False, then, it give no effect in the gradient of positional embedding p when gradients of the loss propagate backward.

It is because the P and PTP_In is a scipy sparse matrix instead of a pytorch tensor.

Was that intended?