1) training on fragmentet training sample, about the monte carlo integration 2) shear on background, not connected one to one with the light 3) polish A3 4) downsizing the datasets.