Effective Gradient Sample Size via Variation Estimation for Accelerating Sharpness aware Minimization