Function add_

train.py:711–712

(self, *args, **kwargs)

            # backward() on future micro-batches. But we can modify .data directly so the autograd engine doesn't detect in-place modifications.
            # TODO: this is unbelievably hacky and not mathematically sound, I'm just seeing if it works at all.
            def add_(self, *args, **kwargs):
                self.data.add_(*args, **kwargs)
            for p in model_parameters:
                p.add_ = add_.__get__(p)