Testing things

The new version seems to work, but I am getting worried that there is something wrong with the old when using just a few n_samples -- it seems to be biased. Need to check compared to lundbergs python shap package