Why does the offset_normalizer use switched WH order in deformable attention

[Beanocean]

In the deformable attention, the spatial shapes represented by height (H) and width (W). Why is it necessary to exchange the offset_normalizer to the witdh(W) height(H) order here?

mmcv/mmcv/ops/multi_scale_deform_attn.py

Lines 352 to 353 in c46684c

	offset_normalizer = torch.stack(
	[spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)

Originally posted by @Beanocean in #3197