Method run

lit_nlp/components/shap_explainer.py:111–194

Generate SHAP explanations for model predictions given a set of inputs. Args: inputs: The subset of inputs from the dataset to get prediction for. If empty, falls back to dataset.examples. model: The model making the predictions that get explained. dataset: The dataset

(
      self,
      inputs: list[JsonDict],
      model: lit_model.Model,
      dataset: lit_dataset.Dataset,
      model_outputs: Optional[list[JsonDict]] = None,
      config: Optional[JsonDict] = None
  )

    return {'saliency': types.FeatureSalience(autorun=False, signed=True)}

  def run(
      self,
      inputs: list[JsonDict],
      model: lit_model.Model,
      dataset: lit_dataset.Dataset,
      model_outputs: Optional[list[JsonDict]] = None,
      config: Optional[JsonDict] = None
  ) -> Optional[list[dict[str, dtypes.FeatureSalience]]]:
    """Generate SHAP explanations for model predictions given a set of inputs.

    Args:
      inputs: The subset of inputs from the dataset to get prediction for. If
        empty, falls back to dataset.examples.
      model: The model making the predictions that get explained.
      dataset: The dataset from which the inputs originated.
      model_outputs: Unused, but reqired by the base class.
      config: A dictionary containing the key of the feature to explain, and the
        optional sample size if taking a random sample from the inputs.

    Returns:
      A list of FeatureSalience objects, one for each (randomly sampled) input,
      containing per-input feature salience values in the range of [-1, 1].

    Raises:
      ValueError: if the value of `config[EXPLAIN_KEY]` is not found in the
        model's output spec.
    """
    del model_outputs   # Unused. SHAP calls the model directly

    config_defaults = {k: v.default for k, v in self.config_spec().items()}
    config = dict(config_defaults, **(config or {}))

    default_pred_key = utils.find_spec_keys(
        model.output_spec(), _SUPPORTED_OUTPUT_TYPES)[0]
    pred_key = config.get(EXPLAIN_KEY) or default_pred_key
    pred_spec = model.output_spec().get(pred_key)
    if not pred_spec:
      raise ValueError('SHAP requires a prediction field to explain. Could not '
                       f'find {pred_key} in spec, {str(model.output_spec())}.')

    input_feats = [key for key in model.input_spec() if key in dataset.spec()]

    example_data = inputs or dataset.examples
    examples: pd.DataFrame = pd.DataFrame(example_data)[input_feats]
    sample_size = int(config.get(SAMPLE_KEY, 0))
    if sample_size and len(examples) > sample_size:
      inputs_to_use: pd.DataFrame = examples.sample(sample_size)
    else:
      inputs_to_use: pd.DataFrame = examples

    random_baseline = dataset.sample(1).examples
    background = pd.DataFrame(random_baseline)[input_feats]

    def prediction_fn(examples):
      dict_examples: list[JsonDict] = [
          dict(zip(input_feats, feature_values)) for feature_values in examples
      ]