0.9.0 - 2021-11-30¶
- Wheels for Python 3.6 have been dropped.
- Wheels for Python 3.9 have been added.
anomaly¶
- Moved
base.AnomalyDetector
toanomaly.AnomalyDetector
. - Implemented
anomaly.ConstantThresholder
. - Implemented
anomaly.QuantileThresholder
. - Implemented
anomaly.OneClassSVM
.
base¶
- Renamed
base.WrapperMixin
tobase.Wrapper
. - Introduced
base.WrapperEnsemble
. - Clarified the difference between a
base.typing.Dataset
and abase.typing.Stream
. AStream
is an instance of aDataset
and is stateful. ADataset
is stateless. It's essentially the same difference between anIterable
and anIterator
in the Python standard library.
compat¶
- Added
compat.PyTorch2RiverClassifier
- Implemented median absolute deviation in
stats.MAD
. - Refactored
compat.PyTorch2RiverRegressor
- Fixed an issue where some statistics could not be printed if they had not seen any data yet.
compose¶
- You can now use a
list
as a shorthand to build aTransformerUnion
. - Fixed a visualization issue when using a pipeline with multiple feature unions.
- The prejudiced terms
blacklist
andwhitelist
have both been renamed tokeys
. - Removed
learn_unsupervised
parameter from pipeline methods. - Implemented
compose.TransformerProduct
.
datasets¶
- Added
datasets.Keystroke
.
ensemble¶
- Bug fixes in
ensemble.SRPClassifier
andensemble.SRPRegressor
. - Some estimators have been moved into the
ensemble
module.
feature_extraction¶
- Implemented
feature_extraction.Lagger
. - Implemented
feature_extraction.TargetLagger
.
meta¶
This module has been deleted.
- Move
meta.PredClipper
to thepreprocessing
module. - Removed
meta.BoxCoxRegressor
. - Moved
meta.TargetTransformRegressor
tocompose.TargetTransformRegressor
. - Moved
meta.TargetStandardScaler
topreprocessing.TargetStandardScaler
.
model_selection¶
- This new module replaces the
expert
module. - Implemented
model_selection.GreedyRegressor
. - Added
ModelSelector
base class.
optim¶
optim.Adam
andoptim.RMSProp
now work withutils.VectorDict
s as well asnumpy.ndarray
s.- Added
optim.losses.Huber
.
preprocessing¶
- Enabled
preprocessing.OneHotEncoder
to one-hot encode values that are list or sets.
reco¶
- Added a
debug_one
method toreco.FMRegressor
.
selection¶
- This new module replaces the
expert
module. - Implemented
selection.GreedyExpertRegressor
.
stats¶
- Fixed an issue where some statistics could not be printed if they had not seen any data yet.
- Implemented median absolute deviation in
stats.MAD
. - The
stats.Mean
andstats.Var
implementations have been made more numerically stable.
time_series¶
time_series.Detrender
andtime_series.GroupDetrender
have been removed as they overlap withpreprocessing.TargetStandardScaler
.- Implemented a
time_series.evaluate
method, which performs progressive validation for time series scenarios. - Implemented
time_series.HorizonMetric
class to evaluate the performance of a forecasting model at each time step along a horizon. - Implemented
time_series.HoltWinters
.
utils¶
- Moved
model_selection.expand_param_grid
toutils.expand_param_grid
. - Added
utils.poisson
. - Added the
utils.log_method_calls
context manager. - Added the
utils.warm_up_mode
context manager. - Added the
utils.pure_inference_model
context manager.