search

NVIDIA-AI-IOT / yolo_deepstream

yolo model qat and deploy with deepstream&tensorrt
Apache License 2.0
533 stars 135 forks source link

yolov5 qat Add graph still use useless data conversion node. #52

Closed maro-jeon closed 11 months ago

maro-jeon commented 11 months ago

Thank you for your repo. it's really good for understanding qat tool.

Now i am currently working on qat for yolov5. I added Q/DQ node with your advice. see the below. image

but still they have useless conversion fp 16 to int8. see the below trex graph. image

how can i remove the useless conversion node ? Thanks, Maro JEON

maro-jeon commented 11 months ago

I added Q/DQ layer on add , concat, resize node. and remove (merge) the redundant q/dq layer. finally got the same fps qat engine compared with ptq engine.

Maro JEON