v1.5.0

NeuroTreeModels v1.5.0

Diff since v1.4.0

Changes

• Add a scaler option to be applied at the per-node signal to the learner
• Apply a normalisation so that sum of feature weights == 1

These changes may result in some changes in the optimal learning rate (lr kwargs to the learner) compared to previous release (<=v1.4)

Merged pull requests:

• Jdb/scaling (#17) (@jeremiedb)

v1.4.0

NeuroTreeModels v1.4.0

Diff since v1.3.1

Model constructors (NeuroTreeRegressor, NeuroTreeClassifier) now include the following arguments:

• metric: the evaluation metric to be tracked
• early_stopping_rounds
• device=:cpu: either :cpu or :gpu
• gpuID=0: the gpu device ID

Example:

config = NeuroTreeRegressor(
    loss = :mse,
    nrounds = 10,
    num_trees = 16,
    depth = 5,
    device = :cpu
)

These arguments have been removed from fit

m = NeuroTreeModels.fit(config, dtrain; feature_names, target_name)

This provides the ability to train on GPU when using the MLJ interface:

using MLJBase, NeuroTreeModels
m = NeuroTreeRegressor(depth=5, nrounds=10, device=:gpu)
X, y = @load_boston
mach = machine(m, X, y) |> fit!
p = predict(mach, X)

Note that inference is always performed on :cpu when using MLJ's predict.
For GPU inference, should use p = mach.fitresult(X; device=:gpu)

Merged pull requests:

• Jdb/api (#14) (@jeremiedb)

v1.3.1

NeuroTreeModels v1.3.1

Diff since v1.3.0

Merged pull requests:

• Dev (#13) (@jeremiedb)

Closed issues:

• Allow Tweedie loss function within NeuroTreeModels.jl (#2)

v1.3.0

NeuroTreeModels v1.3.0

Important changes

• The kwarg device (:cpu / :gpu) is moved from NeuroTreeRegressor to fit. Same for gpuID.
• Removal of outsize argument.
• Introduction of NeuroTreeClassififier, respecting the MLJ interface
• It's no longer needed to specify the number of classes (through the deprecated outsize kwarg). They're now automatically detected throuh the target variable number of levels.
• Classification tasks (using NeuroTreeClassififier) require the target variable to be <:Categorical

Diff since v1.2.0

Merged pull requests:

• add tweedie (#11) (@jeremiedb)
• Dev (#12) (@jeremiedb)

Closed issues:

• Robust preprocessing of target for classifiction tasks (#10)

v1.2.0

NeuroTreeModels v1.2.0

Diff since v1.1.1

Merged pull requests:

• Compat entry for Optimisers (#6) (@pat-alt)
• set compat to julia >= 1.9 (#7) (@jeremiedb)

Closed issues:

• Warning regarding rewrite of supports_weights (#8)
• Memory errror when transforming model to gpu (#9)

v1.1.1

NeuroTreeModels v1.1.1

Diff since v1.1.0

Merged pull requests:

• Error when fitting with CUDA not functional (adding conditional to fix) (#4) (@pat-alt)
• Dev (#5) (@jeremiedb)

v1.1.0

NeuroTreeModels v1.1.0

Diff since v1.0.1

Merged pull requests:

• Dev (#3) (@jeremiedb)

v1.0.1

NeuroTreeModels v1.0.1