Adaptivity of Stochastic Gradient Methods for Non-convex Optimization