Reproduce CBNetV2 with other models

[ghost]

Hi！
Recently I want to utilize CBNetV2 in PaddleDetection. After reading your paper, I connect two ResNet50 together directly while find some questions in "Assistant Supervision" part. You have mentioned the detection head and neck are "weight sharing" so that I simply compute loss after forwarding the feats from two backbones to the same neck and head. Then I backward total_loss(=lead_loss+assist_loss) to train entire network.
I want to know if this way is right and is same to your cbnetv2 in mmdetection? If my method exists mistakes, hoping you can give me some advice.
Hoping receive your reply!