TabNetHyperparameters

Properties

batch_sizetunable

integerdefault 1048576≥ 1

Batch size.

penaltytunable

numberdefault 0.001≥ 0

Sparsity regularization penalty.

clip_value

number | number[] | nulldefault null≥ 1 items

Gradient clip value.

losstunable

stringdefault "auto"

Loss function. auto = set from outcome type.

epochstunable

integerdefault 50≥ 1

Number of training epochs.

drop_lasttunable

booleandefault false

Drop the last incomplete batch.

decision_width

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

Decision prediction layer width.

attention_width

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

Attention embedding width.

num_stepstunable

integerdefault 3≥ 1

Number of decision steps.

feature_reusagetunable

numberdefault 1.3≥ 0

Feature reusage coefficient.

mask_typetunable

stringdefault "sparsemax"

Masking function.

one of"sparsemax""entmax"

virtual_batch_sizetunable

integerdefault 65536≥ 1

Virtual batch size (ghost batch normalization).

valid_splittunable

numberdefault 0≥ 0< 1

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

learn_ratetunable

numberdefault 0.02> 0

Learning rate.

lr_decaytunable

numberdefault 0.1≥ 0≤ 1

Learning rate decay.

step_sizetunable

integerdefault 30≥ 1

Learning rate scheduler step size.

checkpoint_epochstunable

integerdefault 10≥ 1

Checkpoint interval in epochs.

cat_emb_dimtunable

integerdefault 1≥ 1

Categorical embedding dimension.

num_independenttunable

integerdefault 2≥ 1

Number of independent GLU layers at each encoder step.

num_sharedtunable

integerdefault 2≥ 1

Number of shared GLU layers at each encoder step.

num_independent_decodertunable

integerdefault 1≥ 1

Number of independent GLU layers for pretraining.

num_shared_decodertunable

integerdefault 1≥ 1

Number of shared GLU layers for pretraining.

momentumtunable

numberdefault 0.02≥ 0

Momentum for batch normalization.

pretraining_ratiotunable

numberdefault 0.5≥ 0≤ 1

Ratio of features to mask during pretraining.

device

stringdefault "auto"

Compute device.

one of"auto""cpu""cuda"

importance_sample_size

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

Sample size for importance calculation.

early_stopping_monitortunable

stringdefault "auto"

Metric monitored for early stopping.

one of"auto""valid_loss""train_loss"

early_stopping_tolerancetunable

numberdefault 0≥ 0

Minimum relative improvement to reset the patience counter.

early_stopping_patiencetunable

integerdefault 0≥ 0

Number of epochs without improvement before stopping.

num_workers

integerdefault 0≥ 0

Number of subprocesses for data loading.

skip_importance

booleandefault false

Skip importance calculation.

ifwtunable

booleandefault false

Inverse Frequency Weighting in classification.

optimizer

stringdefault "adam"

Optimizer name (e.g. "adam").

lr_scheduler

string | null

Learning-rate scheduler: "step" or "reduce_on_plateau". null = none.

one of"step""reduce_on_plateau"

Relationships