How to get the int8 engine in Tensorrt_yolo?

[amtommy]

Hello!

I would like to create my own model engine for int8 inference.
How to create the onnx file? And how many calibration images do we need for int8 inference?

Thank you!

[storrrrrrrrm]

set mode in launch.xml from FP32 to INT8

<launch>
  <arg name="yolo_type" default="yolov3"/>
  <arg name="label_file" default="coco.names"/>
  <arg name="input_topic" default="/image_raw"/>
  <arg name="output_topic" default="rois"/>
  <arg name="calib_image_directory" default="$(find-pkg-share tensorrt_yolo)/calib_image/"/>
  <arg name="mode" default="FP32"/>

[amtommy]

Thanks for your reply. It's very helpful!
I would like to build my own model. Do you know how to create the onnx file?
How many calibration images do we need for int8 inference?

[storrrrrrrrm]

Thanks for your reply. It's very helpful! I would like to build my own model. Do you know how to create the onnx file?

have a look at this https://github.com/ultralytics/yolov3/blob/master/export.py

How many calibration images do we need for int8 inference?

have no expirence about this,but i think there may not be a general answer, it may be relevant to your specific dataset