TensorRT3.0 can't convert FasterRCNN plugin layer when calling caffe2GIEModel

i am going to updating TensorRT2.1 to TensorRT3.0 on TX1, so i coded some test cases, but i met a bug when converting caffe model to GIE model, the FasterRcnn plugin layer convert fail, it’s ok when using TensorRT2.1, i got this error message “Plugin layer output count is not equal to caffe output count” , i defined 3 plugin layers in network, but i don’t know which layer is wrong , anyone have met this issue and give me some advice, thanks!

env: Tx1, cuda9.0 cudnn7.0 TensorRT3.0