When training a model, don’t we need to subtract the root joint from the model’s output (out)?   Why are GT and OUT inconsistent?