TabNetHyperparameters

Properties

batch_sizetunable

integer≥ 1

Batch size.

penaltytunable

number≥ 0

Sparsity regularization penalty.

clip_value

number | number[] | null≥ 1 items

Gradient clip value.

losstunable

string

Loss function. auto = set from outcome type.

epochstunable

integer≥ 1

Number of training epochs.

drop_lasttunable

boolean

Drop the last incomplete batch.

decision_width

integer | integer[] | null≥ 1≥ 1 items

Decision prediction layer width.

attention_width

integer | integer[] | null≥ 1≥ 1 items

Attention embedding width.

num_stepstunable

integer≥ 1

Number of decision steps.

feature_reusagetunable

number≥ 0

Feature reusage coefficient.

mask_typetunable

string

Masking function.

one of"sparsemax""entmax"

virtual_batch_sizetunable

integer≥ 1

Virtual batch size (ghost batch normalization).

valid_splittunable

number≥ 0< 1

Fraction of data used for (tabnet-internal) validation.

learn_ratetunable

number> 0

Learning rate.

optimizer

string

Optimizer name, resolved by the tabnet backend.

lr_scheduler

string | null

Learning-rate scheduler. NULL = none.

one of"step""reduce_on_plateau"

lr_decaytunable

number≥ 0≤ 1

Learning rate decay.

step_sizetunable

integer≥ 1

Learning rate scheduler step size.

checkpoint_epochstunable

integer≥ 1

Checkpoint interval in epochs.

cat_emb_dimtunable

integer≥ 1

Categorical embedding dimension.

num_independenttunable

integer≥ 1

Number of independent GLU layers at each encoder step.

num_sharedtunable

integer≥ 1

Number of shared GLU layers at each encoder step.

num_independent_decodertunable

integer≥ 1

Number of independent GLU layers for pretraining.

num_shared_decodertunable

integer≥ 1

Number of shared GLU layers for pretraining.

momentumtunable

number≥ 0

Momentum for batch normalization.

pretraining_ratiotunable

number≥ 0≤ 1

Ratio of features to mask during pretraining.

device

string

Compute device.

one of"auto""cpu""cuda"

importance_sample_size

integer | integer[] | null≥ 1≥ 1 items

Sample size for importance calculation.

early_stopping_monitortunable

string

Metric monitored for early stopping.

one of"auto""valid_loss""train_loss"

early_stopping_tolerancetunable

number≥ 0

Minimum relative improvement to reset the patience counter.

early_stopping_patiencetunable

integer≥ 0

Number of epochs without improvement before stopping.

num_workers

integer≥ 0

Number of subprocesses for data loading.

skip_importance

boolean

Skip importance calculation.

ifwtunable

boolean

Inverse Frequency Weighting in classification.

Relationships