Add user data to checkpoint?




I use OpenNMT-py and my translator requires a bit of additional data on
top of what is included in checkpoint (such as BPE codes, for example).
I find it convenient to keep everything I need in one big file, so I add
these extra keys after the training is done.

I think it may be useful to parametrize drop_checkpoint with a dict
with user-defined data. Would you find it valuable?

(Dominik Macháček) #2

just make your fork and help yourself, as you need. I have my own one and I added extra info to checkpoint. It’s easy and possible.


Sure, that’s what I did. I just don’t want my fork to diverge too much, so if my changes were considered valuable for the project, I’d rather had them merged.

(Dominik Macháček) #4

Can you share the repo with me? I have a look and maybe I’ll use it.

Generally I’m not against the merge, I don’t know…


It is a private repo of my company, so, unfortunately, no. Also, the code is a bit hairy at the moment, I didn’t go all the way to make it usable for anyone but myself.

If the problem itself is acknowledged as worth solving, I would make a cleaner patch at my own time, maybe work with maintainers to make sure it doesn’t violate the project’s principles I’m not aware of, and have it merged.