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() and
hidden_units_2(), for use with
brulee::brulee_mlp_two_layer() (#339).
grid_max_entropy() and
grid_latin_hypercube() are deprecated in favor of
grid_space_filling() (#332).
pull_dials_object() has been removed
(#344).
The grid_*() methods for workflow
objects have been removed (#344).
The deprecation of the default argument to the
constructors new_quant_param() and
new_qual_param() has been escalated to an error
(#344).
initial_umap() and target_weight()
for recipes::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() for
recipes::step_impute_mean().
Added num_runs() for
recipes::step_nnmf() (#281).
Added harmonic_frequency() for
recipes::step_harmonic() (#281).
Added validation_set_prop() for
embed::step_discretize_xgb() (#280).
Deprecation of pull_dials_object() has been
escalated to an error. Please use extract_parameter_dials()
instead (#265).
The methods grid_regular.workflow(),
grid_random.workflow(),
grid_max_entropy.workflow(), and
grid_latin_hypercube.workflow() have been deprecated
(#302).
The constructor functions for single parameters,
new_quant_param() and new_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(), and
rate_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 constructors
new_quant_param() and new_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 on
mtry() 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() and
conditional_test_type() now work with the partykit engine
rather than the party engine (#224).
new_quant_param() no longer requires
range and inclusive if values is
supplied (#87).
Chicago data set was removed. It can be found in
the modeldata package.summary_stat() is a new parameter for use in
recipes::step_window().
A general class_weights() parameter was added for
imbalanced models.
prior_outcome_range(),
prior_terminal_node_coef(), and
prior_terminal_node_expo() are new parameters for prior
distribution specification related to
parsnip::bart().
num_knots() and survival_link() are new
parameters for spline survival models (@mattwarkentin, #195).
vocabulary_size() is a new parameter used in
textrecipes::step_tokenize_sentencepiece() and
textrecipes::step_tokenize_bpe().
The new extract_parameter_dials() method supersedes
pull_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() and select_features()
for generalized additive models.
conditional_min_criterion,
conditional_test_statistic,
conditional_test_type for models with the
party engine.
diagonal_covariance(),
regularization_method(),
shrinkage_correlation(),
shrinkage_frequencies(), and
shrinkage_variance() for linear and quadratic discriminant
analysis.
penalty_L1(), penalty_L2(), and
scale_pos_weight() for boosted trees with the
xgboost engine (@joeycouse, #176).
prior_mixture_threshold() and
prior_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 for parameters() 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 to
2^5 instead of 2^-1.
The grid_*() functions no longer generate subclassed
tibbles.
predictor_prop() was added.
The levels argument for tune_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
use svm_margin() instead.
When kept in the original units, a parameter’s range must now be the same data type as the parameter.
Renamed margin() to svm_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)
smoothness() was added.param_set() is being renamed
parameters(). 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 be
2^8 to 2^12.
The range for max_tokens() was changed to be 0 to
1000.
offset() has been renamed kernel_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 the range and the trans. 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() and
neighbors() were changed.
update.param_set() now takes multiple named
arguments.
Two functions for space-filling designs were added:
grid_max_entropy() and
grid_latin_hypercube().
A data set was added for modeling ridership on the Chicago L trains.
spline_degree(), over_ratio(),
under_ratio(), freq_cut(),
unique_cut(), num_breaks(),
min_unique(), num_hash(),
signed_hash(), sample_prop(),
window_size(), min_dist(), and
degree_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 to penalty
in a few models to be consistent with this
change.
batch_size and threshold were
added.
Added a set of parameters for the textrecipes
package issue
16.