Model Pruning

IR models are extremely memory intensive and call for memory-oriented compression, we design a pruning flow to cut down the peak memory usage with special handling of the long-range skip connections (which incurs large peak memory overhead and is redundant for the task performance).