I noticed that variable and tensor has been merged together after 0.4. But I still encounter some code accessing data field directly. So I just want to clarify a few things.
It seems like tensor.data is still a tensor object. Does autodiff engine track computation when operating on tensor.data directly (e.g. c = a.data + b.data)? If not, is tensor.data just another view of tensor but with requires_grad=False?