merge 2 models

[Gaoeee]

Hi，thank you for your excellent work.

I'm trying to reproduce your work, and I tried to train the lidar-only model by TransFusion-L , trained the 2D backbone by mask-rcnn as you mentioned，which all reached the ideal metrics separately.

But I failed to used the merged model to train DeepInteraction. I removed the pts_voxel_encoder layers and moved the shared_conv layers to encoder as imgpts_neck.shared_conv_pts layers.

Could you give me some more specific instructions?