Consistent documentation of the same parameter for each builder
Some parameters in the configuration file may be used by multiple builders, but should be documented differently. This is the case for the sample_augmentation
parameter, as sample augmentation is not done in exactly the same way for classification and regression.
Therefore, in addition to the long_desc
and short_desc
attributes, we could add a specific_desc
attribute to the fields that specify the parameters. This attribute could be a dictionary whose keys are builder names and whose value is builder-specific documentation.
The question of fitting a parameter field to a builder can also arise for the default value of certain parameters. For example, in the deep_learning
section, the default loss is cross entropy, which is not appropriate for a regression problem.
What do you think?