The result of yolox is wrong in int8 mode

[Mike_deng]

Please provide complete information as applicable to your setup.

**• Jetson **
• DeepStream 6.0 and 5.1
• jetson Pack 4.6
• TensorRT 8.0
I try to run int8 in yolox, the result is wrong, but fp16 is ok.and in othe model, the results are ok in int8 mode. I don’t know how to check the issue, can you give some suggestion?

the follow is the github:
inference:
GitHub - egbertYeah/yolox_deepstream to do inference.
int8 calibrator:
tensorrt_demos/onnx_to_tensorrt.py at master · jkjung-avt/tensorrt_demos · GitHub to do int8 calibrator

the step：
1.use onnx_to_tensorrt.py to create calib bin file and trt file.
2. change deepstream config int8-calib-file ,model-engine-file and network-mode.
3. run deepstream-app

[mchi]

@Mike_deng

int8-calib-file

did you generate the INT8 calibration table based on the tipical inference images? And, how many images did you use to generate the INT8 calibration table?

[Mike_deng]

yes.I create the calibration table, the number of images is 1000.

[mchi]

@Mike_deng

the result is wrong,

what do you mean by “wrong”? Does it totally not work or just accuracy is not so good?
You may could refer to TensorRT/INT8 Accuracy - eLinux.org to narrow down the INT8 accuracy.