Using tensorrt to transform Caffe's trained model

Using tensorrt to transform Caffe’s trained model, it is found that the predicted results are completely incorrect, but not all of them. Has anyone met with it?
My environment is RTX2080ti + CUDA10.2 + TensorRT- + windows10.

Request you to share your model and script, so that we can help you better.

I converted OpenCV image data (BGR) to RGB format, and the problem was solved