Fix learning rate logging when using LRSchedulersContainer

[DNXie]

Now the lr starts from 1e-05 (specified in the config) instead of 0.001

Log:

rl_trainer/learning_rate: 1e-05

[DNXie]

cc @wukaixingxp

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.60%. Comparing base (248126a) to head (b3603ef).
⚠️ Report is 44 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main     #507       +/-   ##
===========================================
+ Coverage   63.57%   84.60%   +21.03%     
===========================================
  Files          77       30       -47     
  Lines        7569     3807     -3762     
===========================================
- Hits         4812     3221     -1591     
+ Misses       2757      586     -2171     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[wukaixingxp]

Thanks for the quick fix, just curious what should we do if self.engine.lr_schedulers.schedulers has more than one lr_scheduler in the list.. any reason we make the self.engine,lr_schedulers.schedulers a list?

[casteryh]

lgtm.
Would it make sense to record all the lrs of each scheduler? (I have no idea why there are multiple schedulers in the first place tbh).

[DNXie]

@casteryh @wukaixingxp
I am not 100% sure why there would be multiple lr schedulers either..
But according to the docstring here, all the lr schedulers are the same. And in this example, it also only accesses the first lr_scheduler.