Closed sash-a closed 4 months ago
Hey!
Thanks a lot for pointing to this, I would love to offload that to marl-eval.
However, it seems that the implementation you point to is loosing some functionality we are using in BenchMARL, in particular:
write
function can now log only one metric at a time. This means that if I have a dictionary of n
metrics for my evaIuation step I will have to perform n
calls to the function and thus to json.dump(self.data, f, indent=4)
which is a blocking i/o callI just reported this in case it is interesting for you.
I would love to use this in benchmarl, but I would at least need these features in order to make it bc-compatible and not loose efficiency :)
Thanks for pointing these out, we can update the functionality :smile:
Closing as stale, feel free to reopen
Hi there :wave:
Just a suggestion, but we've recently integrated our
MarlEvalLogger
fromMava
into theid-marl-eval
pypi package. We originally got this logger from this repo and made minimal changes to it so that it should be general enough to work with all repos. So if you would like to switch to depending on that logger we're happy to take that (small) maintenance burden off your shoulders.Links to marl eval logger, a quick readme about the tools and how we use it in mava
If you've got any questions about it I'm more than happy to answer them :smile: