Closed
Description
🔧 Proposed code refactoring
Reduce logging on startup in case experiment config yaml has keys that do not match the current dataclass.
When starting the app, get_experiments(q) goes trough all experiments and loads the config yaml.
In case experiment configuration dataclasses code has been changed (as it has been the case recently), the startup logs can be polluted with
2023-08-16 17:56:12,061 - WARNING: Keys {'reward_model', 'ppo_clip_policy', 'advantages_gamma', 'ppo_epochs', 'kl_target', 'use_rlhf', 'initial_kl_coefficient', 'ppo_batch_size', 'adaptive_kl_control', 'scaling_factor_value_loss', 'offload_reward_model', 'advantages_lambda', 'ppo_generate_temperature', 'kl_horizon', 'ppo_clip_value'} are not in the config.
2023-08-16 17:56:12,061 - WARNING: Keys {'number_of_texts'} are not in the config.
Motivation
Remove the startup warnings as
- It is unclear where these warning exactly comes from and what to do
- Warning should only be present when a user starts an experiment from a configuration (or old experiment) that has unseen keys.