Change Log

0.2.7

Added

reproducible results on a single machine via the seeds parameter.
proper CUDA support. Automatic detection if you have a CUDA-enabled machine.

0.2.6

Added

enable logistic output/binary classification. Occurs automatically when torch.nn.BCELoss() function passed to the criterion parameter.

0.2.5

Fixed

fixed init_test_size of model selection functions working on the full dataset including NAs in the target variable

0.2.4

Added

make model instantiation and model selection robust to variables with no data in them

0.2.3

Added

feature contribution weighting for data availabilty in interval predict functions

Fixed

got rid of hardcoded references to date column in interval predict code

0.2.2

Added

ability to generate uncertainty intervals via the model.interval_predict() function
ability to generate uncertainty intervals on synthetic vintages via the model.ragged_interval_predict() function

0.2.1

Added

initial_ordering parameter to variable_selection() and select_model() functions. In recursive feature addition (RFA) variable selection, can obtain initial variable order either via their feature contribution in a full model, or from univariate model performances. Former (default) is about 2x faster.

0.2.0

Added

ability to obtain feature contributions to the model via model.feature_contribution() function
automatic variable selection given a set of hyperparameters via variable_selection() function in LSTM.model_selection
automatic hyperparameter tuning given a set of variables via hyperparameter_tuning() function in LSTM.model_selection
automatic variable and hyperparameter tuning via select_model() function in LSTM.model_selection

Changed

hide printing of Training model n when quiet=True