Function test_shap_assert

catboost/python-package/ut/medium/test.py:9779–9806

()

def test_shap_assert():
    model_path = test_output_path('model.json')
    pool = Pool([[0, ], [1, ], ], [0, 1])
    model = train(pool, {'iterations': 1, 'task_type': 'CPU', 'devices': '0'})
    model.save_model(model_path, format='json')

    json_model = json.load(open(model_path))
    json_model['scale_and_bias'] = [1, [1]]
    json.dump(json_model, open(model_path, 'w'))
    model = CatBoost().load_model(model_path, format='json')
    shap_values = model.get_feature_importance(type='ShapValues', data=pool)
    predictions = model.predict(pool)
    assert (len(predictions) == len(shap_values))
    for i, pred_idx in enumerate(range(len(predictions))):
        assert (abs(sum(shap_values[pred_idx]) - predictions[pred_idx]) < 1e-9), (sum(shap_values[pred_idx]) - predictions[pred_idx])

    json_model['oblivious_trees'] = [{
        'leaf_values': [1, 2],
        'leaf_weights': [1, 0],
        'splits': [{'border': 0.5, 'float_feature_index': 1, 'split_index': 0, 'split_type': 'FloatFeature'}]
    }]
    json_model['features_info'] = {
        'float_features': [{'borders': [0.5], 'feature_index': 0, 'flat_feature_index': 0, 'has_nans': False, 'nan_value_treatment': 'AsIs'}]
    }

    json.dump(json_model, open(model_path, 'w'))
    model = CatBoost().load_model(model_path, format='json')
    model.get_feature_importance(type='ShapValues', data=pool)


@pytest.mark.parametrize('shrink_mode', ['Constant', 'Decreasing'])