How to generate normalized offset for deformable roi pooling?

[HB-X]

In deformable convolutional networks [1], a fully conected layer is used to generates the normalized offsets $\Delta{\hat{p}{ij}}$, and the offset normalization is necessary to make the offset learning invariant to roi size.
However, I didn't find normalization operation in this implementation:
https://github.com/open-mmlab/mmcv/blob/4e8972fbf9be5790523950af6edebb4a2ab40ee0/mmcv/ops/deform_roi_pool.py#L140.
But the offsets are transformed to $\Delta{p{ij}}$: https://github.com/open-mmlab/mmcv/blob/4e8972fbf9be5790523950af6edebb4a2ab40ee0/mmcv/ops/csrc/common/cuda/deform_roi_pool_cuda_kernel.cuh#L128.
How can I implement this operation?
[1] Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., & Wei, Y. (2017). Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision (pp. 764-773).