Limit numpy version to avoid issue with tensorboard, conv np.int to just int, fix mismatched device issuecuda dev issue