Quantization for YOLOv9 Segmentation Models

[GokceSengun]

Hi,

I’m using YOLOv9 for segmentation tasks and noticed that quantization is currently supported for object detection models. Since the backbone is the same across all YOLOv9 variants, I wanted to ask:

• Is it possible to use quantization with YOLOv9 segmentation models?
• If not, are there any plans to support it or recommended workarounds? Thanks for your help!

[levipereira]

It is entirely possible to perform quantization for any YOLO model that uses the same architecture. As you mention the backbone remains the same, even when applied to different tasks like segmentation or detection. Since I have already completed quantization for the backbone, the same approach can be applied to segmentation models. The main challenge lies in implement a dataloader for model evaluation and training (for QAT purposes), which is a part I haven't had time to implement yet. The only aspect that might require minor adjustments is the neck of the YOLO model, which can be adapted as needed. Thus, in theory, the quantization process is ready; we just need to implement the dataloader to enable proper training and evaluation.

[levipereira]

I am working on a new project that involves migrate this project to the modelopt. It is a powerful quantization tool, but I am currently addressing some challenges with NVIDIA. For this reason, I have temporarily put this project on hold.

[felixkarevo]

Hi @levipereira! Have you seen any good examples on Github for quantization (PTQ or QAT) of a YoloV8, 9 or 11 segmentation model? Or maybe examples of INT8 conversion of a YoloV8, 9 or 11 segmentation model?