Pytorch-lightning: Custom callbacks
alainjungo
All 4 comments
We've been talking about this for a while. I'm very for this.
One aspect I would like to add to this would be deterministic ordering of callbacks. It can be very difficult to maintain callbacks if the order of callback execution is not pre-defined, we end up having to handle a lot of edge cases for each callback as a result.
jeffling
on 9 Dec 2019
a very simple yet flexible way to achieve deterministic ordering is to use the order in which they're provided.
eg. Trainer(..., callbacks=[custom_callback_1, custom_callback_2, model_checkpoint_callback]) would first call custom_callback_1, then custom_callback_2, etc. for each hook
otherwise, enforcing order seems like a very daunting task.
see keras for an example.
jeremyjordan
on 12 Feb 2020
As @jeremyjordan suggested, using a Keras-style ordered list of custom callbacks would be incredibly useful.
davidlrobinson
on 18 Feb 2020
Yeah, after splitting callback in #849 we opened a discussion in #896...
@alainjungo do you have anything else you would add to this topic?
I would close this request in favour of #896 so pls let's continue the discussion there :robot:
btw @davidlrobinson do you remember the issue it was discussed so we can link it here too?
Borda
on 20 Feb 2020
Related issues
maxime-louis
路
3Comments
versatran01
路
3Comments
iakremnev
路
3Comments
williamFalcon
路
3Comments
edenlightning
路
3Comments
Most helpful comment
We've been talking about this for a while. I'm very for this.
One aspect I would like to add to this would be deterministic ordering of callbacks. It can be very difficult to maintain callbacks if the order of callback execution is not pre-defined, we end up having to handle a lot of edge cases for each callback as a result.