tidytune performs hyperparameter tuning by leveraging the rsample and recipes packages. Adopting a tidy approach improves readability, reproducibility (I had to google that word) and allows the seamless usage of other tidying tools such as dplyr.

Currently, the following methods are implemented:

Here's in a nutshell how you would perform a grid search in tidytune.

In the example below, xgboost_classif_score is a custom scoring function provided by the user.

xgboost_param_grid <- expand.grid(eta = c(0.1, 0.05), max_depth = c(3, 4))

    resamples = resamples, 
    recipe = rec, 
    param_grid = xgboost_param_grid, 
    score_func = xgboost_classif_score, 
    nrounds = 100,
    verbose = FALSE

For grid and random search, there's also a batch version in case you need to save or see your results while the search is progressing (especially useful with big models that take a long time to fine tune).

Currently, model based optimization uses a random forest surrogate model under the hood to map parameter values to predicted model performance.



artichaud1/cook documentation built on May 21, 2019, 9:23 a.m.