About ModulatedDeformConv2dPack when batchsize>1

[Xingxu1996]

Thanks for your good job! When converting my backbone resnet101 into trt using mmdet2trt, batchsize=1 is ok. But when batchsize >1, I find only the first images of each batch can output correct result. After longtime to locate problem, I find that the layer of DCNV2(ModulatedDeformConv2dPack) results in error. That is the results is all correct after I delete DCNv2 in config. So, could you help me. Thank you!