You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the current implementation, environment processes are created at the moment when the configuration is being parsed/loaded. I would propose to decouple those two because:
In the situation when we want to start multiple trainer to run training in parallel, it is preferred that configuration is parsed and loaded first, and each trainer process will start the environments by themselves. We need an option to separate configuration parsing and environment process creation.
For readability, it is more clear if parse_conf_file does not create the environment under the hood. In fact, explicitly creating them seems more clear.
Goal
Separate the concept of configuration parsing and environment creation in the current implementation.
Such change should be transparent to most of the alf developers
Update the related documentation
Test and verify the change
The text was updated successfully, but these errors were encountered:
I think there is a very complicated reason why creating env is inside parse_conf_file. I suggest asking @emailweixu first before making any change. It should be related to how alf config works and env random seed initialization.
Closing this because it depends on retiring the .gin files completely (or to put it in another way, waiting for the retirement of .gin is an easier alternative). Will re-open once that is done.
Motivation
In the current implementation, environment processes are created at the moment when the configuration is being parsed/loaded. I would propose to decouple those two because:
parse_conf_file
does not create the environment under the hood. In fact, explicitly creating them seems more clear.Goal
configuration parsing
andenvironment creation
in the current implementation.The text was updated successfully, but these errors were encountered: