search

ModelTC / MQBench

Model Quantization Benchmark
Apache License 2.0
768 stars 140 forks source link

how to ptq for Faster RCNN or SSD？ #134

Closed wangshankun closed 2 years ago

wangshankun commented 2 years ago

From QDROP paper，i notice the benchmark result include Faster RCNN； image

Could you provide this examples？

In addition, it's best to provide PTQ of SSD，another import object detection network；

PannenetsF commented 2 years ago

https://github.com/ModelTC/United-Perception/tree/main/configs/quant/det https://github.com/ModelTC/United-Perception/tree/main/configs/quant/det

Here are configs for the FRCNN and RetinaNet, with which you can easily re-produce qdrop/brecq/adaround on detection task via the united-perception framework.

We did not do PTQ on SSD, you can try it out just like the other cls networks.

On Jul 8, 2022, at 17:59, sandy @.***> wrote:

From QDROP paper，i notice the benchmark result include Faster RCNN； https://user-images.githubusercontent.com/3970271/177967270-99e64612-3384-46bb-b4db-d2fbb1356067.png Could you provide this examples？

In addition, it's best to provide PTQ of SSD，another import object detection network；

— Reply to this email directly, view it on GitHub https://github.com/ModelTC/MQBench/issues/134, or unsubscribe https://github.com/notifications/unsubscribe-auth/AK556VAE42MBBP4FI7ADYDLVS73WRANCNFSM53AOELFQ. You are receiving this because you are subscribed to this thread.

wangshankun commented 2 years ago

@PannenetsF 感谢及时答复； 我的理解检测模型的ptq不需要train会比给出的sample config要简单不少，但是检测模型的ptq没办法直接比对cls网络；因为网络的neck部分是需要做处理的；

我看工具链team在B站视频里面介绍要配合EOD的代码，注册submodule···看起来还是很麻烦的

PannenetsF commented 2 years ago

理解部分可以说的详细一点，有点不明白；

注册的话还是比较简单，和正常搭建网络工作量差别不大。

wangshankun commented 2 years ago

我看分析了你给我的链接， https://github.com/ModelTC/United-Perception/blob/main/configs/quant/det/yolox/yolox_fpga_quant_vitis_qat.yaml 这是检测网络的qat，有点复杂，以至于我没有能力从这个config里面摘录出yolox的ptq

PannenetsF commented 2 years ago

可以参考同目录下的ptq config

wangshankun commented 2 years ago

可以参考同目录下的ptq config

@PannenetsF 有det系列模型的 ptq config参考吗？

PannenetsF commented 2 years ago

https://github.com/ModelTC/United-Perception/tree/main/configs/quant/det

wangshankun commented 2 years ago

https://github.com/ModelTC/United-Perception/tree/main/configs/quant/det

多谢！找到了一个合适的参考 https://github.com/ModelTC/United-Perception/blob/main/configs/quant/det/retinanet/retinanet-r50-w2a4-ada.yaml

wangshankun commented 2 years ago

https://github.com/ModelTC/United-Perception/tree/main/configs/quant/det

EOD自己给的例子，config中的quant_ptq是不支持的 image

问题提交 https://github.com/ModelTC/United-Perception/issues/29