One of my hate library is google/ml_collections which is a library of Python Collections designed for ML use cases. For this purpose, I often see many frameworks of reinforcement learning or machine learning based on this library.
The reason why I don't like this library is that many developers of those frameworks don't seem to understand it, because this library has too rich functions and features to use it well. For most cases, we don't need to use most of them.
Today, I got an error on RL-X as following;
DuplicateFlagError: The flag 'runner' is defined twice. First from rl_x.runner.runner, Second from ml_collections.config_flags.config_flags. Description from first occurrence: ConfigDict instance.
This error sounds like coming from ml_collections to be locked for an flag defined twice. This behavior is quite reasonable but the problem is that this flag stays longer unexpectedly even if you unimport ml_collections .
To resolve this, I read long time to try resolving it reading stackoverflow1, stackoverflow2 and stackoverflow3, but finally it was in vain at all.
The only thing to do is to clear flags from absl. So, only you need to do is to like as following
from absl import flags
for name in list(flags.FLAGS):
delattr(flags.FLAGS, name)
It sucks.
I'm fond of using omegaconf instead of ml_collections btw.