Proposition on how to specify parameters in the config files
This is linked to #68 (closed) but not urgent.
Reminder, cf #68 (closed), preprocessing just requires the parameter names, the type (regression or classification), and the transform (because transformation is done before normalization of the targets). The loss has to be specified in the training config.
Regarding the definition of the parameters, it would be convenient (to avoid copy-pasting blocks) to be able to define them using some tricks such as:
- (1)"Ne_[1, 21]" or something with a regex "Ne_*" to select multiple parameters at once
- (2) with vectors of param, e.g. [EventTime, PopSize], to which we assign the same info
- (3) [training config only] with a keyword ALL (ou * ?) that applies to all parameters previously defined as targets (in the preprocessing).
Examples :
A) First example
- in preprocessing
learned_params:
[event_size,Ne_*]:
type: regression
log_transform: true
[time2,time_adm]:
type: classification
- in training :
learned_params:
[event_size,Ne_*]:
loss_func: MSE
[time2,time_adm]:
loss_func: my_own_classif_function
B) Second example
- in preprocessing
learned_params:
[event_size,Ne_*]:
type: regression
log_transform: true
- in training
learned_params:
ALL:
loss_func: MSE
or
learned_params:
*:
loss_func: MSE
?