PIT: Optimization of Dynamic Sparse Deep Learning Models via Permutation Invariant Transformation