Pytorch-lightning: Fix .test() on ddp
This might be broken on notebooks only.
875 solves a few problems with .test()
However, ddp + .test might be broken on notebooks because of the "spawn" option. (likely #747).
williamFalcon
All 3 comments
Anyone interested in looking at this?
@Borda
williamFalcon
on 17 Feb 2020
I am investigating this. Thats how I ran into the spawn issue with shared Q's. I am not using a notebook. This is (I believe) because ddp reduces tensors and parameters back to dev_ids[0] and there is no obvious way to get them back to the spawning CPU. I am slowly making progress, but don't have a lot of consistent time to put in.
sneiman
on 17 Feb 2020
👍3
Isn't ddp broken in notebooks in the first place?
s-rog
on 24 Feb 2020
Was this page helpful?
0 / 5 - 0 ratings
Related issues
justusschock
路
3Comments
anthonytec2
路
3Comments
edenlightning
路
3Comments
Vichoko
路
3Comments
williamFalcon
路
3Comments
Most helpful comment
I am investigating this. Thats how I ran into the spawn issue with shared Q's. I am not using a notebook. This is (I believe) because ddp reduces tensors and parameters back to dev_ids[0] and there is no obvious way to get them back to the spawning CPU. I am slowly making progress, but don't have a lot of consistent time to put in.