- Adds
search_experiments
,search_registered_models
, andsearch_experiments_without_nodelete
. Deprecateslist_experiments
,list_registered_models
, andlist_experiments_without_nodelete
.
- Fixes an
rbind
bug inlist_experiment
. (The same issue resolved inseach_runs
in version0.6.7
.) Allrbind
calls replaced with the internal "lite" version ofdplyr::bind_rows()
.
- Fixes a bug with
search_runs
whererbind
-ing runs with different numbers of columns (e.g. possible due to no logged parameters or no end time for a run) resulted in an error.
- Uses
purrr::insistently
to retry artifact saves and loads, since S3's API has a habit of returning a500
every once in a while. Five retries by default, with an exponential backoff.
- Fixes a bug where MLFlow
1.26.0+
now requires thesource
parameter when creating a model version. InlightMLFlow
, this is defaulted to the artifact URI of the current run. - CI improvements
- Rips out lots of unused code that's only useful for MLFlow running on local (which is not the goal of this project to enable)
- Rips out the MLFlow
models
infrastructure. The idea for this package -- at least in the short-term -- is to save models as artifacts with the intention of using them in Plumber APIs, etc. If you wanted to load an (e.g.) XGBoost model into Sagemaker, you could uselog_artifact
in tandem withFUN = xgboost::xgb.save
to make this easy. - Adds some more helpful hints for MLFlow clients
- Fixes a bug that would leave a GH token in plain text in the run
- Added a
NEWS.md
file to track changes to the package. - Adds
get_metric
- Adds
get_param
- Abstracts away more timestamp headaches from the user