Integrate YoloV8 in Sample Object Detector

Autonomous Vehicles DRIVE AGX Orin DRIVE AGX Orin General
23

Hi @SivaRamaKrishnaNV,

Do you want me to convert onnx to “tensorRT_model.bin” and try to validate this with our python code or simply convert onnx to tensorrt which has extension as “.engine” and verify?

24

Yes. Convert ONNX-> TRT model and use TRT APIs to perform inference to compare the outputs for same input image.

25

Hi @SivaRamaKrishnaNV,

Had already converted onnx to TRT (.engine extension) and validated. The behaviour is same for ONNX & TRT.
Issue is observed when convert ONNX to TensortRT_model.bin using trtexec and use it with sample_object_detector_tracker.

26

I have feed the same input to both TRT and your repo

Added below in prepare_input() method at ONNX-YOLOv8-Object-Detection/yolov8/YOLOv8.py to get input buffer data

        input_tensor = input_img[np.newaxis, :, :, :].astype(np.float32)
        input_tensor.tofile("input.dat")

feed the same buffer as input to TRT model using trtexec

/usr/src/tensorrt/bin/trtexec --onnx=/home/nvidia/yolov8/ONNX-YOLOv8-Object-Detection/yolov8n.onnx --saveEngine=/home/nvidia/yolo8n.trt
/usr/src/tensorrt/bin/trtexec --loadInputs='images:input.dat'  --loadEngine=/home/nvidia/yolo8n.trt --dumpOutput

and noticed the TRT model outputs are matching with some error in decimal position.

def inference(self, input_tensor):
        start = time.perf_counter()
        outputs = self.session.run(self.output_names, {self.input_names[0]: input_tensor})
        print("output")
        for i in range(84):
            for j in range(8400):
              print(outputs[0][0][i][j])

Please double check postprocessing steps.

27

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.