[DISCUSSION] Subgraph semantics

[dmlc]

[DISCUSSION] Subgraph semantics

[zheng-da]

After thinking more about subgraph training, I don't feel that we need to support the "shared" mode for subgraph training. After the forward computation in a mini-batch, we don't need to write data back to the parent frame directly. Users should explicitly write data back with "inplace=True" to avoid index_copy. In addition, does any underlying framework support non-contiguous view on a tensor?

[jermainewang]

I think it's still good to have for completeness, but you could put it in a lower priority and may not finish it in one PR. I agree that we need an "inplace=True" flag when writing back. I don't think current framework support such view. That's why we have FrameRef for that.

[zheng-da]

OK. let's postpone the implementation of the "shared" mode until we find a real application for this.

[jermainewang]

The basic "non-shared" mode has been done in #90. We are now focusing on performance issue.