janknaup / skl-efs-predictor Goto Github PK

Following a security breach warning from CircleCI, tokens need rotating

Storing of fitness trace should be optional and disabled by default, just like population trace.
Storing full traces over big fits consumes considerable memory and may lead to problems in hyperparameter tuning. The trace should only be stored when the user actively intends to diagnose it.

Fitness calculation is a time consuming step and embarassingly parallel. It should be parallelized via joblib to take advantage of multiple cores.

Since a predictor is strictly required for calculating the fitness of a feature selection, the uer should be able to use the fitted predictor directly.

Requires refitting the predictor to the fittest specimen after iteration is stopped. This step should be made optional in case EFS is used in a pipeline with subsequent predictors. Cost of predictor fitting should not be an issue as one more fitness evaluation must be insignificant compared to the number of evaluations during iterations.

Standardized helper functions to create plots of evolution progress would be helpful for diagnosing evolution quality and problems.
At least the following plots should be provided:

• Fitness trace over generations
• Trace of features selected in whole population
• Trace of features selected in fittest specimen
• Trace of population diversity

build oc circleci fails because miniconda.sh has a bug (apparently a regression of an earlier issue) where miniconda.sh can only run successfully in bash.

To Reproduce
CI build of any branch fails with error message:

./miniconda.sh: 444: [[: not found