Training Faster with Compressed Gradient