When I use the same onnx converted to trt file, the inference results in tensorrt7 and tensorrt8 are not consistent, where tensorrt 8 has the correct result and tensorrt 7 appears to have more bbox, some of which have some position shift.
The model I am using is retinanet,what is the problem and why tensorrt7 and tensorrt8 behave differently on this model?
When I use the ssd model, the results of tensorrt7 and tensorrt8 are basically the same and correct.
Environment
TensorRT Version: 8.0.1.6
NVIDIA GPU: GTX1650TI
NVIDIA Driver Version: 470
CUDA Version: 10.2
CUDNN Version: 8.2.0