-
Added three new parameters for use in postprocessing in the tailor package (#357).
-
buffer()
sets the distance on either side of a classification threshold within which predictions are considered equivocal intailor::adjust_equivocal_zone()
. -
lower_limit()
andupper_limit()
sets the ranges for numeric predictions intailor::adjust_numeric_range()
.
-
-
All messages, warnings and errors has been translated to use {cli} package (#311).
-
For space-filling designs for
$p$ parameters, there is a higher likelihood of finding a space-filling design for1 < size <= p
. Also, single-point designs now default to a random grid (#363).
-
The space-filling design functionality was expanded to include several new types of designs: Audze-Eglais, max/min L1, max/min L2, and uniform. These are all pre-computed designs accessed from the sfd package (#329).
-
A new function is used to access all of the space-filling designs called
grid_space_filling()
(#329). -
Two new parameters,
activation_2()
andhidden_units_2()
, for use withbrulee::brulee_mlp_two_layer()
(#339).
-
grid_max_entropy()
andgrid_latin_hypercube()
are deprecated in favor ofgrid_space_filling()
(#332). -
pull_dials_object()
has been removed (#344). -
The
grid_*()
methods forworkflow
objects have been removed (#344). -
The deprecation of the
default
argument to the constructorsnew_quant_param()
andnew_qual_param()
has been escalated to an error (#344).
- Added
initial_umap()
andtarget_weight()
forrecipes::step_umap()
(#324).
-
Improving styling of error messages by switching to cli (#315, #317, #321).
-
Update usage of tranformation functions from the scales package to their new names and require the corresponding version v1.3.0 (#323).
-
Added
trim_amount()
forrecipes::step_impute_mean()
. -
Added
num_runs()
forrecipes::step_nnmf()
(#281). -
Added
harmonic_frequency()
forrecipes::step_harmonic()
(#281). -
Added
validation_set_prop()
forembed::step_discretize_xgb()
(#280).
-
Deprecation of
pull_dials_object()
has been escalated to an error. Please useextract_parameter_dials()
instead (#265). -
The methods
grid_regular.workflow()
,grid_random.workflow()
,grid_max_entropy.workflow()
, andgrid_latin_hypercube.workflow()
have been deprecated (#302). -
The constructor functions for single parameters,
new_quant_param()
andnew_qual_param()
, as well as for parameter sets,parameters_constr()
, now have improved handling of the call shown in error messages (#291, #295). -
The constructor for parameter sets,
parameters_constr()
, now checks that all inputs have the same length (#295).
-
Added learning rate scheduler parameters
rate_decay()
,rate_initial()
,rate_largest()
,rate_reduction()
,rate_schedule()
,rate_step_size()
, andrate_steps()
for the new brulee functions (#253). -
Added
num_clusters()
parameter for tidyclust models (#259). -
Added
num_leaves()
parameter for lightbgm models (@joeycouse, #256).
-
The
default
argument to the constructorsnew_quant_param()
andnew_qual_param()
is deprecated.value_seq()
now uses the same logic to generate a sequence of parameter values regardless of how long that sequence is (#153, #229). -
prior_terminal_node_expo()
for Bayesian adaptive regression trees (BART) now defaults to a range greater than 1 to limit explosive tree growth (#251). -
The label for
spline_degree()
was improved.
-
The new parameter
mtry_prop()
is a variation onmtry()
where the value is interpreted as the proportion (rather than the count) of predictors that will be randomly sampled at each split (#233). -
conditional_test_statistic()
andconditional_test_type()
now work with the partykit engine rather than the party engine (#224). -
new_quant_param()
no longer requiresrange
andinclusive
ifvalues
is supplied (#87).
- The
Chicago
data set was removed. It can be found in themodeldata
package.
-
summary_stat()
is a new parameter for use inrecipes::step_window()
. -
A general
class_weights()
parameter was added for imbalanced models. -
prior_outcome_range()
,prior_terminal_node_coef()
, andprior_terminal_node_expo()
are new parameters for prior distribution specification related toparsnip::bart()
. -
num_knots()
andsurvival_link()
are new parameters for spline survival models (@mattwarkentin, #195). -
vocabulary_size()
is a new parameter used intextrecipes::step_tokenize_sentencepiece()
andtextrecipes::step_tokenize_bpe()
.
-
The new
extract_parameter_dials()
method supersedespull_dials_object()
which has been deprecated. -
activation()
now supports values of"tanh"
. -
New link to article explaining how to make custom parameter objects was added to the pkgdown site.
-
adjust_deg_free()
andselect_features()
for generalized additive models. -
conditional_min_criterion
,conditional_test_statistic
,conditional_test_type
for models with theparty
engine. -
diagonal_covariance()
,regularization_method()
,shrinkage_correlation()
,shrinkage_frequencies()
, andshrinkage_variance()
for linear and quadratic discriminant analysis. -
penalty_L1()
,penalty_L2()
, andscale_pos_weight()
for boosted trees with thexgboost
engine (@joeycouse, #176). -
prior_mixture_threshold()
andprior_slab_dispersion()
for sparse PCA. -
stop_iter()
for early stopping.
-
Re-licensed package from GPL-2 to MIT. See consent from copyright holders here.
-
param_set()
, scheduled for removal in version 0.0.5, is now removed.
-
Quantitative parameters now print the number of possible values if they have been set with
value_set()
(@kmdupr33, #138). -
The
print()
method forparameters()
has changed to be more clear. -
A new function,
pull_dials_object()
was also added. -
Duplicate parameter combinations are now automatically removed from grid results.
-
The range for
epochs()
was change to start at 10 iterations instead of 1. -
The lower range for
spline_degree()
now starts at 1 instead of 3. -
The upper range for
cost()
now goes to2^5
instead of2^-1
.
- A number of new parameter objects associated with engine-specific functions were added for engines "ranger", "randomForest", "earth" and "C5.0".
-
The
grid_*()
functions no longer generate subclassed tibbles. -
predictor_prop()
was added. -
The
levels
argument fortune_grid()
can now handle a named vector, to account for differences in ordering.
-
The range of
dist_power()
was changed so that the lower limit is 1. -
The deprecation period for
margin()
is over; please usesvm_margin()
instead.
- Quick bug fix release related to range checks in 0.0.5. The check is more forgiving when the required type is integer and a double is provided.
-
When kept in the original units, a parameter's range must now be the same data type as the parameter.
-
Renamed
margin()
tosvm_margin()
(@gabrielodom and @gralgomez, #85) -
A bug in space filling designs with qualitative parameters was fixed (#94)
-
A better error message was created when grids are used with parameters sets that contain parameters that require finalization (#99)
-
Space-filling desings now share the same attributes as other grid objects (#100)
-
The range for
sample_frac()
was fixed (#96)
- Parameters
smoothness()
was added.
-
param_set()
is being renamedparameters()
. The old name implied that you only use it to set parameters (say from a recipe or model rather than de novo).param_set()
will be available until version 0.0.5. -
The range for
num_hash()
was increase to be2^8
to2^12
. -
The range for
max_tokens()
was changed to be 0 to 1000.
offset()
has been renamedkernel_offset()
to avoid name conflicts.
-
All parameter objects are now parameter functions. For example, the pre-configured object
mtry
is now a function,mtry()
, with arguments for therange
and thetrans
. This provides greater flexibility in parameter creation, and should feel more natural. -
deg_free()
erroneously produced real values; integers are now returned. -
Default ranges were changed for
learn_rate()
andneighbors()
were changed. -
update.param_set()
now takes multiple named arguments.
-
Two functions for space-filling designs were added:
grid_max_entropy()
andgrid_latin_hypercube()
. -
A data set was added for modeling ridership on the Chicago L trains.
- Parameters
spline_degree()
,over_ratio()
,under_ratio()
,freq_cut()
,unique_cut()
,num_breaks()
,min_unique()
,num_hash()
,signed_hash()
,sample_prop()
,window_size()
,min_dist()
, anddegree_int()
were added.
-
Parameter objects now contain code to finalize their values and a number of helper functions for certain data-specific parameters. A
force
option can be used to avoid updating the values. -
Parameter objects are printed differently inside of tibbles.
-
regularization
was changed topenalty
in a few models to be consistent with this change. -
batch_size
andthreshold
were added. -
Added a set of parameters for the
textrecipes
package issue 16.
- First CRAN version