Missed TensorFlow Dev Summit? Check out the video playlist. Watch recordings


View source on GitHub

Per-feature configuration for TFL canned estimators.

    name, is_missing_name=None, default_value=None, lattice_size=2,
    monotonicity='none', unimodality='none', reflects_trust_in=None, dominates=None,
    pwl_calibration_always_monotonic=False, pwl_calibration_convexity=0,
    pwl_calibration_num_keypoints=10, pwl_calibration_input_keypoints='quantiles',
    pwl_calibration_clip_min=None, pwl_calibration_clip_max=None,
    pwl_calibration_clamp_min=False, pwl_calibration_clamp_max=False, num_buckets=0,
    vocabulary_list=None, regularizer_configs=None

Used in the notebooks

Used in the tutorials

A feature can either be numerical or categorical. Numeric features will be calibrated using a piecewise-linear function with the given number of keypoints. Categorical features should have num_buckets > 0 and the vocabulary_list represent their categories. Several of the config fields can be filled in automatically based on the FeatureColumns used by the model but can also be provided explicitly. See __init__ args comments for details.

Currently only one dimensional feature are supported.


feature_columns = [
        'age', default_value=-1),
        'thal', vocabulary_list=['normal', 'fixed', 'reversible']),

model_config = tfl.configs.CalibratedLatticeConfig(
            # Monotonically increasing.
            # Per feature regularization.
                tfl.configs.RegularizerConfig(name='calib_hessian', l2=1e-4),
            # Partial monotonicity:
            # output(normal) <= output(fixed)
            # output(normal) <= output(reversible)
            monotonicity=[('normal', 'fixed'), ('normal', 'reversible')],
    # Global regularizers
feature_analysis_input_fn = create_input_fn(num_epochs=1, ...)
train_input_fn = create_input_fn(num_epochs=100, ...)
estimator = tfl.estimators.CannedClassifier(


  • name: The name of the feature, which should match the name of a given FeatureColumn or a key in the input feature dict.
  • is_missing_name: The name of a FeatureColumn or key in the input feature dict that indicates missing-ness of the main feature.
  • default_value: [Automatically filled in from FeatureColumns] If set, this value in the input value represents missing. For numeric features, the output will be imputed. If default_value is provided for a categocial features, it would corresponds to the last bucket counted in num_buckets.
  • lattice_size: The number of lattice verticies to be used along the axis for this feature.
  • monotonicity: - For numeric features, specifies if the model output should be monotonic in this feature, using 'increasing' or 1 to indicate increasing monotonicity, 'decreasing' or -1 to indicate decreasing monotonicity, and 'none' or 0 to indicate no monotonicity constraints.
    • For categorical features, a list of (category_a, category_b) pairs from the vocabulary list indicating that with other features fixed, model output for category_b should be greater than or equal to category_a. If no vocabulary list is specified, we assume implcit vocabulary in the range [0, num_buckets - 1].
  • unimodality: For numeric features specifies if the model output should be unimodal in corresponding feature, using 'valley' or 1 to indicate that function first decreases then increases, using 'peak' or -1 to indicate that funciton first increases then decreases, using 'none' or 0 to indicate no unimodality constraints. Not used for categorical features.
  • reflects_trust_in: None or a list of tfl.configs.TrustConfig instances.
  • dominates: None or a list of tfl.configs.DominanceConfig instances.
  • pwl_calibration_always_monotonic: Specifies if the piecewise-linear calibration should always be monotonic regardless of the specified end-to-end model output monotonicity with respect to this feature.
  • pwl_calibration_convexity: Spefices the convexity constraints of the calibrators for numeric features. Convexity is indicated by 'convex' or 1, concavity is indicated by 'concave' or -1, 'none' or 0 indicates no convexity/concavity constraints. Does not affect categorical features. Concavity together with increasing monotonicity as well as convexity together with decreasing monotonicity results in diminishing return constraints.
  • pwl_calibration_num_keypoints: Number of keypoints to use for piecewise-linear calibration.
  • pwl_calibration_input_keypoints: Indicates what should be used for the input keypoints of the piecewise-linear calibration. It can be one of:
    • String 'quantiles': Input keypoints are set to feature quantiles.
    • String 'uniform': Input keypoints are uniformly spaced in feature range.
    • A list of numbers: Explicitly specifies the keypoints.
  • pwl_calibration_clip_min: Input values are lower clipped by this value.
  • pwl_calibration_clip_max: Input values are upper clipped by this value.
  • pwl_calibration_clamp_min: for monotonic calibrators ensures that the minimum value in calibration output is reached.
  • pwl_calibration_clamp_max: for monotonic calibrators ensures that the maximum value in calibration output is reached.
  • num_buckets: [Automatically filled in from FeatureColumns] Number of categories for a categorical feature. Out-of-vocabulary and missing/default value should be counted into num_buckets (last buckets).
  • vocabulary_list: [Automatically filled in from FeatureColumns] The input vocabulary of the feature.
  • regularizer_configs: None or a list of per-feature tfl.configs.RegularizerConfig instances.



View source


Returns existing or default RegularizerConfig with the given name.