Whats the problem?
It's quite hard to follow the flow with the config because of the nested yaml structure.
Proposed solution
- Collapse the config for every experiment.
- Have a hybrid where you read from classes with types and then use dataset as a YAML.