TabM

NeuroTabModels.Models.TabM.EnsembleView Type

EnsembleView(k)

Repeat (D, B) input to (D, K, B). Passes through (D, K, B) unchanged.

source

NeuroTabModels.Models.TabM.LinearBatchEnsemble Type

LinearBatchEnsemble(in_f, out_f; k, scaling_init=:random_signs, bias=true)

Batch-ensemble linear: y = S ⊙ (W(R ⊙ x)) + bias.

Arguments

• in_f, out_f: Input and output dimensions.

• k::Int: Ensemble size.

• scaling_init: :ones, :normal, or :random_signs; or (R, S) tuple.

• bias::Bool: Per-member bias (default true).

source

NeuroTabModels.Models.TabM.ScaleEnsemble Type

ScaleEnsemble(k, d; init=:random_signs, bias=false)

Per-member elementwise scaling on (d, k, batch) input.

source

NeuroTabModels.Models.TabM.TabMConfig Type

TabMConfig(; kwargs...)

Configuration for TabM ensemble backbones.

Arguments

• k::Int: Ensemble size (default 32).

• n_blocks::Int: Number of MLP blocks (default 3).

• d_block::Int: Hidden dimension (default 512).

• dropout::Float64: Dropout rate (default 0.1).

• arch_type::Symbol: :tabm, :tabm_mini, or :tabm_packed (default :tabm).

• MLE_tree_split::Bool: Split output head for Gaussian MLE (default false).

source

NeuroTabModels.Models.TabM.TabMConfig Method

(config::TabMConfig)(; nfeats, outsize)

Build a Lux.Chain from config. Output shape is (outsize, k, batch).

Arguments

• nfeats::Int: Number of input features.

• outsize::Int: Number of output units.

source

NeuroTabModels.Models.TabM.LinearEnsemble Method

LinearEnsemble(in_f, out_f, k; bias=true)

k independent linear layers. Input/output (features, k, batch).

Constructs a GroupedDense with TabM's rsqrt weight and bias init.

source