使用tensorrt推理，结果不对

[jiangxiangchuan]

Hello, I converted model .pth file to .onnx file by tools/export_onnx.py, then I convert .onnx file to .trt file by tensorrt/build/segment. The n_cats is 3. The command is as belows:

python tools/export_onnx.py --config configs/bisenet_my.py --weight-path /project/bisenet/res/model_best.pth --outpath /project/bisenet/res/model_best.onnx --aux-mode eval

./segment compile /project/bisenet/res/model_best.onnx /project/bisenet/res/model_best.trt --fp16 8

When I use .trt file to infer with one image, it encounters an error. The command is as below:

./segment run /project/bisenet/res/model_best.trt ./1.bmp ./1_result.bmp

I debugged the code, and find the retrun numbers of the following code are very large as showed in following image.
the infer code:

the value of res:

Why does it happen?