Add more logging + ability to push to more downstream models #40
Conversation
facebook-github-bot
added
the
CLA Signed
|
I was wondering why switch to rich based logging vs normal python logging? Also if we're changing to rich based logging, would you want to propagate it through all of the 'examples' as well? @xiaomengy not sure if you have any thoughts on how we should do logging in the code base. |
|
The main reasons I swapped is that in the past (and now too), (1) I've had trouble getting python to actually output what I need (configure log level correctly) + (2) rich does a nice job of colonizing input. I'm not married to it, but I do like it quite a bit :). I'd also be curious about how to do the downstream models, adding it how I did isn't exactly super clean, but I didn't really see another way. You need access to the extra downstream models in inside the train call, but don't really want them to be passed in I think, hence the optional arg. |
|
re: downstream models yeah I think it might be better if we made it a class variable. If we need to wait for the other inference servers to connect, then we could make add/set functions to downstream models (also maybe remove if downstream model nodes go offline). |
EntilZha
force-pushed
the
pr-downstream-logging
branch
from
df39517
to
69a5077
Compare
|
LGTM! |
rlmeta/agents/dqn/apex_dqn_agent.py
Outdated
| @@ -243,7 +266,7 @@ def __call__(self, index: int): | |||
| self._optimizer, self._batch_size, self._grad_clip, | |||
| self._multi_step, self._gamma, | |||
| self._learning_starts, self._sync_every_n_steps, | |||
| self._push_every_n_steps, self._collate_fn) | |||
| self._push_every_n_steps, self._collate_fn, additional_models_to_update=self._additional_models_to_update) | |||
There was a problem hiding this comment.
Please try to format this file by yapf. We have the line limit of 80 characters.
downstream models
EntilZha
force-pushed
the
pr-downstream-logging
branch
from
69a5077
to
87601f4
Compare
This PR adds:
additional_downstream_modelstoagent.trainthat can be used to push to more than one downstream model (e.g., if there are multiple parallel loops, push to the model that each loop is using).