A new model mode ("quantile regression") was added. Including:
linear_reg() engine for "quantreg".?set_mode.fit_xy() can now take dgCMatrix input for x argument (#1121).
fit_xy() can now take sparse tibbles as data values (#1165).
predict() can now take dgCMatrix and sparse tibble input for new_data argument, and error informatively when model doesn't support it (#1167).
New extract_fit_time() method has been added that returns the time it took to train the model (#853).
Transitioned package errors and warnings to use cli (#1147 and #1148 by @shum461, #1153 by @RobLBaker and @wright13, #1154 by @JamesHWade, #1160, #1161, #1081).
fit_xy() currently raises an error for gen_additive_mod() model specifications as the default engine ("mgcv") specifies smoothing terms in model formulas. However, some engines specify smooths via additional arguments, in which case the restriction on fit_xy() is excessive. parsnip will now only raise an error when fitting a gen_additive_mod() with fit_xy() when using the "mgcv" engine (#775).
Aligned null_model() with other model types; the model type now has an engine argument that defaults to "parsnip" and is checked with the same machinery that checks other model types in the package (#1083).
Make sure that parsnip does not convert ordered factor predictions to be unordered.
Ensure that knit_engine_docs() has the required packages installed (#1156).
Fixed bug where some models fit using fit_xy() couldn't predict (#1166).
For quantile prediction, the quantile argument to predict() has been deprecate in facor of quantile_levels. This does not affect models with mode "quantile regression".
The quantile regression prediction type was disabled for the deprecated surv_reg() model.
Added a missing tidy() method for survival analysis glmnet models (#1086).
A few changes were made to achive more speed-ups (#1075) (#1073) (#1072)
Tightened logic for outcome checking. This resolves issues—some errors and some silent failures—when atomic outcome variables have an attribute (#1060, #1061).
Fixed bug in fitting some model types with the "spark" engine (#1045).
Fixed issues in metadata for the "brulee" engine where several arguments were mistakenly protected. (#1050, #1054)
Fixed documentation for mlp(engine = "brulee"): the default values for learn_rate and epochs were swapped (#1018).
Fixed a bug in the integration with workflows where using a model formula with a formula preprocessor could result in a double intercept (#1033).
We no longer add eval_time arguments to the prediction specification for the engine (#1039).
parsnip now lets the engines for [mlp()] check for acceptable values of the activation function (#1019)
rpart_train() has been deprecated in favor of using decision_tree() with the "rpart" engine or rpart::rpart() directly (#1044).
.filter_eval_time() was moved to the survival standalone file.
Improved errors and documentation related to special terms in formulas. See ?model_formula to learn more. (#770, #1014)
Improved errors in cases where the outcome column is mis-specified. (#1003)
The new_data argument for the predict() method for censoring_model_reverse_km objects has been deprecated (#965).
When computing censoring weights, the resulting vectors are no longer named (#1023).
The predict() method for censoring_model_reverse_km objects now checks that ... are empty (#1029).
Fixed bug where prediction on rank deficient lm() models produced .pred_res instead of .pred. (#985)
Fixed bug where sparse data was being coerced to non-sparse format doing predict().
For BART models with the dbarts engine, predict() can now also return the standard error for confidence and prediction intervals (#976).
augment() now works for censored regression models.
A few censored regression helper functions were exported: .extract_surv_status() and .extract_surv_time() (#973, #980).
Fixed bug where boost_tree() models couldn't be fit with 1 predictor if validation argument was used. (#994)
This release of parsnip contains a number of new features and bug fixes, accompanied by several optimizations that substantially decrease the time to fit() and predict() with the package.
glmnet models fitted with base-R family objects are now supported for linear_reg(), logistic_reg(), and multinomial_reg() (#890).
multi_predict() methods for linear_reg(), logistic_reg(), and multinom_reg() models fitted with the "glmnet" engine now check the type better and error accordingly (#900).
.organize_glmnet_pred() now expects predictions for a single penalty value (#876).
The time argument to predict_survival() and predict_hazard() is deprecated in favor of the new eval_time argument (#936).
Added several internal functions (to help work with Surv objects) as a standalone file that can be used in other packages via usethis::use_standalone("tidymodels/parsnip"). These changes provide tooling for downstream packages to handle inverse probability censoring weights (#893, #897, #937).
An internal method for generating inverse probability of censoring weights (IPCW) of Graf et al (1999) is available via .censoring_weights_graf().
Made fit() behave consistently with respect to missingness in the classification setting. Previously, fit() erroneously raised an error about the class of the outcome when there were no complete cases, and now always passes along complete cases to be handled by the modeling function (#888).
Fixed bug where model fits with engine = "earth" would fail when the package's namespace hadn't been attached (#251).
Fixed bug where model fits with factor predictors and engine = "kknn" would fail when the package's namespace hadn't been attached (#264).
Fixed bug with prediction from a boosted tree model fitted with "xgboost" using a custom objective function (#875).
Implemented a number of optimizations in parsnip's backend that substantially decrease evaluation time to fit() and predict() (#901, #902, #910, #921, #929, #923, #931, #932, #933).
logistic_reg() will now warn at fit() when the outcome has more than two levels (#545).
Rather than being implemented in each method, the check for the new_data argument being mistakenly passed as newdata to multi_predict() now happens in the generic. Packages re-exporting the multi_predict() generic and implementing now-duplicate checks may see new failures and can remove their own analogous checks. This check already existed in all predict() methods (via predict.model_fit()) and all parsnip multi_predict() methods (#525).
Functions now indicate what class the outcome was if the outcome is the wrong class (#887).
The minimum version for R is now 3.5 (#926).
Moved forward with the deprecation of req_pkgs() in favor of required_pkgs(). The function will now error (#871).
Transitioned all soft-deprecations that were at least a year old to warn-deprecations. These changes apply to fit_control(), surv_reg(), varying(), varying_args(), and the "liquidSVM" engine.
Various bug fixes and improvements to documentation.
For censored regression models, a "reverse Kaplan-Meier" curve is computed for the censoring distribution. This can be used when evaluating this type of model (#855).
The model specification methods for generics::tune_args() and
generics::tunable() are now registered unconditionally (tidymodels/workflows#192).
Adds documentation and tuning infrastructure for the new flexsurvspline engine for the survival_reg() model specification from the censored package (@mattwarkentin, #831).
The matrix interface for fitting fit_xy() now works for the "censored regression" mode (#829).
The num_leaves argument of boost_tree()s lightgbm engine (via the bonsai package) is now tunable.
A change in our data checking code resulted in about a 3-fold speed-up in parsnip (#835)
A bagged neural network model was added (bag_mlp()). Engine implementations will live in the baguette package.
Fixed installation failures due to undocumented knitr installation dependency (#785).
fit_xy() now fails when the model mode is unknown.
brulee engine-specific tuning parameters were updated. These changes can be used with dials version > 1.0.0.
fit() and fit_xy() doesn't error anymore if control argument isn't a control_parsnip() object. Will work as long as the object passed to control includes the same elements as control_parsnip().
Improved prompts related to missing (or not loaded) extension packages as well as better handling of model mode conflicts.
boost_tree() engine. To supply engine-specific arguments that are documented in xgboost::xgb.train() as arguments to be passed via params, supply the list elements directly as named arguments to set_engine(). Read more in ?details_boost_tree_xgboost (#787).Enable the use of case weights for models that support them.
show_model_info() now indicates which models can utilize case weights.
Model type functions will now message informatively if a needed parsnip extension package is not loaded (#731).
Refactored internals of model specification printing functions. These changes are non-breaking for extension packages, but the new print_model_spec() helper is exported for use in extensions if desired (#739).
Fixed bug where previously set engine arguments would propagate through update() methods despite fresh = TRUE (#704).
Fixed a bug where an error would be thrown if arguments to model functions were namespaced (#745).
predict(type = "prob") will now provide an error if the outcome variable has a level called "class" (#720).
An inconsistency for probability type predictions for two-class GAM models was fixed (#708)
Fixed translated printing for null_model() (#752)
Added a glm_grouped() function to convert long data to the grouped format required by glm() for logistic regression.
xgb_train() now allows for case weights
Added ctree_train() and cforest_train() wrappers for the functions in the partykit package. Engines for these will be added to other parsnip extension packages.
Exported xgb_predict() which wraps xgboost's predict() method for use with parsnip extension packages (#688).
Added a developer function, .model_param_name_key that translates names of tuning parameters.
Fixed a major bug in spark models induced in the previous version (#671).
Updated the parsnip add-in with new models and engines.
Updated parameter ranges for some tunable() methods and added a missing engine argument for brulee models.
Added information about how to install the mixOmics package for PLS models (#680)
Bayesian additive regression trees (BART) were added via the bart() function.
Added the "glm" engine for linear_reg() for numeric outcomes (#624).
Added brulee engines for linear_reg(), logistic_reg(), multinom_reg() and mlp().
A bug for class predictions of two-class GAM models was fixed (#541)
Fixed a bug for logistic_reg() with the LiblineaR engine (#552).
The list column produced when creating survival probability predictions is now always called .pred (with .pred_survival being used inside of the list column).
Fixed outcome type checking affecting a subset of regression models (#625).
Prediction using multinom_reg() with the nnet engine with a single row no longer fails (#612).
When the xy interface is used and the underlying model expects to use a matrix, a better warning is issued when predictors contain non-numeric columns (including dates).
The fit time is only calculated when the verbosity argument of control_parsnip() is 2L or greater. Also, the call to system.time() now uses gcFirst = FALSE. (#611)
fit_control() is soft-deprecated in favor of control_parsnip().
New extract_parameter_set_dials() method to extract parameter sets from model specs.
New extract_parameter_dials() method to extract a single parameter from model specs.
Argument interval was added for prediction: For types "survival" and "quantile", estimates for the confidence or prediction interval can be added if available (#615).
set_dependency() now allows developers to create package requirements that are specific to the model's mode (#604).
varying() is soft-deprecated in favor of tune().
varying_args() is soft-deprecated in favor of tune_args().
An autoplot() method was added for glmnet objects, showing the coefficient paths versus the penalty values (#642).
parsnip is now more robust working with keras and tensorflow for a larger range of versions (#596).
xgboost engines now use the new iterationrange parameter instead of the deprecated ntreelimit (#656).
devtools::load_all() (#653).A model function (gen_additive_mod()) was added for generalized additive models.
Each model now has a default engine that is used when the model is defined. The default for each model is listed in the help documents. This also adds functionality to declare an engine in the model specification function. set_engine() is still required if engine-specific arguments need to be added. (#513)
parsnip now checks for a valid combination of engine and mode (#529)
The default engine for multinom_reg() was changed to nnet.
The helper functions .convert_form_to_xy_fit(), .convert_form_to_xy_new(), .convert_xy_to_form_fit(), and .convert_xy_to_form_new() for converting between formula and matrix interface are now exported for developer use (#508).
Fix bug in augment() when non-predictor, non-outcome variables are included in data (#510).
New article "Fitting and Predicting with parsnip" which contains examples for various combinations of model type and engine. ( #527)
A new linear SVM model svm_linear() is now available with the LiblineaR engine (#424) and the kernlab engine (#438), and the LiblineaR engine is available for logistic_reg() as well (#429). These models can use sparse matrices via fit_xy() (#447) and have a tidy method (#474).
For models with glmnet engines:
penalty (either a single numeric value or a value of tune()) (#481).path_values can be used to set the lambda path as a specific set of numbers (independent of the value of penalty). A pure ridge regression models (i.e., mixture = 1) will generate incorrect values if the path does not include zero. See issue #431 for discussion (#486).The liquidSVM engine for svm_rbf() was deprecated due to that package's removal from CRAN. (#425)
The xgboost engine for boosted trees was translating mtry to xgboost's colsample_bytree. We now map mtry to colsample_bynode since that is more consistent with how random forest works. colsample_bytree can still be optimized by passing it in as an engine argument. colsample_bynode was added to xgboost after the parsnip package code was written. (#495)
For xgboost, mtry and colsample_bytree can be passed as integer counts or proportions, while subsample and validation should always be proportions. xgb_train() now has a new option counts (TRUE or FALSE) that states which scale for mtry and colsample_bytree is being used. (#461)
Re-licensed package from GPL-2 to MIT. See consent from copyright holders here.
set_mode() now checks if mode is compatible with the model class, similar to new_model_spec() (@jtlandis, #467). Both set_mode() and set_engine() now error for NULL or missing arguments (#503).
Re-organized model documentation:
update methods were moved out of the model help files (#479).generics::required_pkgs() was extended for parsnip objects.
Prediction functions now give a consistent error when a user uses an unavailable value of type (#489)
The augment() method was changed to avoid failing if the model does not enable class probabilities. The method now returns tibbles despite the input data class (#487) (#478)
xgboost engines now respect the event_level option for predictions (#460).
An RStudio add-in is available that makes writing multiple parsnip model specifications to the source window. It can be accessed via the IDE addin menus or by calling parsnip_addin().
For xgboost models, users can now pass objective to set_engine("xgboost"). (#403)
Changes to test for cases when CRAN cannot get xgboost to work on their Solaris configuration.
There is now an augument() method for fitted models. See augment.model_fit. (#401)
Column names for x are now required when fit_xy() is used. (#398)
There is now an event_level argument for the xgboost engine. (#420)
New mode "censored regression" and new prediction types "linear_pred", "time", "survival", "hazard". (#396)
Censored regression models cannot use fit_xy() (use fit()). (#442)
show_engines() will provide information on the current set for a model.
For three models (glmnet, xgboost, and ranger), enable sparse matrix use via fit_xy() (#373).
Some added protections were added for function arguments that are dependent on the data dimensions (e.g., mtry, neighbors, min_n, etc). (#184)
Infrastructure was improved for running parsnip models in parallel using PSOCK clusters on Windows.
A glance() method for model_fit objects was added (#325)
Specific tidy() methods for glmnet models fit via parsnip were created so that the coefficients for the specific fitted parsnip model are returned.
glmnet models were fitting two intercepts (#349)
The various update() methods now work with engine-specific parameters.
parsnip now has options to set specific types of predictor encodings for different models. For example, ranger models run using parsnip and workflows do the same thing by not creating indicator variables. These encodings can be overridden using the blueprint options in workflows. As a consequence, it is possible to get a different model fit that previous versions of parsnip. More details about specific encoding changes are below. (#326)tidyr >= 1.0.0 is now required.
SVM models produced by kernlab now use the formula method (see breaking change notice above). This change was due to how ksvm() made indicator variables for factor predictors (with one-hot encodings). Since the ordinary formula method did not do this, the data are passed as-is to ksvm() so that the results are closer to what one would get if ksmv() were called directly.
MARS models produced by earth now use the formula method.
For xgboost, a one-hot encoding is used when indicator variables are created.
Under-the-hood changes were made so that non-standard data arguments in the modeling packages can be accommodated. (#315)
A new main argument was added to boost_tree() called stop_iter for early stopping. The xgb_train() function gained arguments for early stopping and a percentage of data to leave out for a validation set.
If fit() is used and the underlying model uses a formula, the actual formula is pass to the model (instead of a placeholder). This makes the model call better.
A function named repair_call() was added. This can help change the underlying models call object to better reflect what they would have obtained if the model function had been used directly (instead of via parsnip). This is only useful when the user chooses a formula interface and the model uses a formula interface. It will also be of limited use when a recipes is used to construct the feature set in workflows or tune.
The predict() function now checks to see if required modeling packages are installed. The packages are loaded (but not attached). (#249) (#308) (tidymodels/workflows#45)
The function req_pkgs() is a user interface to determining the required packages. (#308)
liquidSVM was added as an engine for svm_rbf() (#300)tidy() was broken on R 4.0.glmnet model.glmnet was removed as a dependency since the new version depends on 3.6.0 or greater. Keeping it would constrain parsnip to that same requirement. All glmnet tests are run locally.
A set of internal functions are now exported. These are helpful when creating a new package that registers new model specifications.
nnet was added as an engine to multinom_reg() #209parsnip and the underlying model function) for spark boosted trees and some keras models. See 897c927.The time elapsed during model fitting is stored in the $elapsed slot of the parsnip model object, and is printed when the model object is printed.
Some default parameter ranges were updated for SVM, KNN, and MARS models.
The model udpate() methods gained a parameters argument for cases when the parameters are contained in a tibble or list.
fit_control() is soft-deprecated in favor of control_parsnip().
A bug was fixed standardizing the output column types of multi_predict and predict for multinom_reg.
A bug was fixed related to using data descriptors and fit_xy().
A bug was fixed related to the column names generated by multi_predict(). The top-level tibble will always have a column named .pred and this list column contains tibbles across sub-models. The column names for these sub-model tibbles will have names consistent with predict() (which was previously incorrect). See 43c15db.
A bug was fixed standardizing the column names of nnet class probability predictions.
Test case update due to CRAN running extra tests (#202)
Unplanned release based on CRAN requirements for Solaris.
The method that parsnip stores the model information has changed. Any custom models from previous versions will need to use the new method for registering models. The methods are detailed in ?get_model_env and the package vignette for adding models.
The mode needs to be declared for models that can be used for more than one mode prior to fitting and/or translation.
For surv_reg(), the engine that uses the survival package is now called survival instead of survreg.
For glmnet models, the full regularization path is always fit regardless of the value given to penalty. Previously, the model was fit with passing penalty to glmnet's lambda argument and the model could only make predictions at those specific values. (#195)
add_rowindex() can create a column called .row to a data frame.
If a computational engine is not explicitly set, a default will be used. Each default is documented on the corresponding model page. A warning is issued at fit time unless verbosity is zero.
nearest_neighbor() gained a multi_predict method. The multi_predict() documentation is a little better organized.
A suite of internal functions were added to help with upcoming model tuning features.
A parsnip object always saved the name(s) of the outcome variable(s) for proper naming of the predicted values.
Small release driven by changes in sample() in the current r-devel.
A "null model" is now available that fits a predictor-free model (using the mean of the outcome for regression or the mode for classification).
fit_xy() can take a single column data frame or matrix for y without error
varying_args() now has a full argument to control whether the full set
of possible varying arguments is returned (as opposed to only the arguments
that are actually varying).
fit_control() not returns an S3 method.
For classification models, an error occurs if the outcome data are not encoded as factors (#115).
The prediction modules (e.g. predict_class, predict_numeric, etc) were de-exported. These were internal functions that were not to be used by the users and the users were using them.
An event time data set (check_times) was included that is the time (in seconds) to run R CMD check using the "r-devel-windows-ix86+x86_64` flavor. Packages that errored are censored.
varying_args() now uses the version from the generics package. This means
that the first argument, x, has been renamed to object to align with
generics.
For the recipes step method of varying_args(), there is now error checking
to catch if a user tries to specify an argument that cannot be varying as
varying (for example, the id) (#132).
find_varying(), the internal function for detecting varying arguments,
now returns correct results when a size 0 argument is provided. It can also now
detect varying arguments nested deeply into a call (#131, #134).
For multinomial regression, the .pred_ prefix is now only added to prediction
column names once (#107).
For multinomial regression using glmnet, multi_predict() now pulls the
correct default penalty (#108).
Confidence and prediction intervals for logistic regression were only computed the intervals for a single level. Both are now computed. (#156)
First CRAN release
set_engine(). There is no engine argumentothers has been replaced by ...regularization was changed to penalty in a few models to be consistent with this change.predict methods, the earth package will need to be attached to be fully operational.snake_case, newdata was changed to new_data.predict_raw method was added.fit interface was previously used to cover both the x/y interface as well as the formula interface. Now, fit() is the formula interface and fit_xy() is for the x/y interface.NEWS.md file to track changes to the package.predict methods were overhauled to be consistent.