The data are now correctly loaded, the angles correctly calculated (hard way), but the network seems to have problem with differentiating angles (?). The MSE is good for the first tests but it's impossible to obtain a correct output when testing after training is complete and model is saved. Issue needs to be investigated.