Effects of l2-regularization and hyperparameters on loss landscape A study connecting training dynamics and loss landscape