Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Accuracy diff of Retinanet int8 and Retinanet FP32 #37873

Closed
lidanqing-intel opened this issue Dec 6, 2021 · 6 comments
Closed

Accuracy diff of Retinanet int8 and Retinanet FP32 #37873

lidanqing-intel opened this issue Dec 6, 2021 · 6 comments
Assignees
@wozna
Labels
Intel status/close 已关闭
Milestone
2022 Q1

Comments

@lidanqing-intel
Copy link
Contributor

Add quantization int8 support to speedup following models

  • Frcnn-r50-fpn
  • Retinanet
  • mobilenetv2
@paddle-bot-old
Copy link

paddle-bot-old bot commented Dec 6, 2021

您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

@lidanqing-intel lidanqing-intel changed the title BML detection models quantization BML detection models int8 quantization Dec 6, 2021
@lidanqing-intel
Copy link
Contributor Author

lidanqing-intel commented Dec 6, 2021 •
edited
Loading

Status update: WW49

  • Retinanet-fpn:
    • INT8 model generated, int8/fp32 1.5X, but there is gap compared with ideal. Currently working on detection_output to speed it up more.
    • Risk:
      • The quant-aware training one epoch model accuracy is bad -> Received T4 and working on generating retinanet quant model
      • Need to ask Diaorenyan
  • Frcnn-r50-fpn:
  • Mobilenetv2
    • Done

@lidanqing-intel lidanqing-intel added this to the Q4 milestone Dec 6, 2021
@lidanqing-intel lidanqing-intel changed the title BML detection models int8 quantization Accuracy issue: BML detection models int8 quantization Jan 7, 2022
@lidanqing-intel
Copy link
Contributor Author

image

@lidanqing-intel
Copy link
Contributor Author

retinanet_fp32_qat
retinanet_int8

@lidanqing-intel lidanqing-intel changed the title Accuracy issue: BML detection models int8 quantization Issue accuracy: BML detection models int8 quantization Jan 7, 2022
@lidanqing-intel
Copy link
Contributor Author

This accuracy issue will be checked again after new INT8 strategy is applied

@lidanqing-intel lidanqing-intel changed the title Issue accuracy: BML detection models int8 quantization Issue accuracy: BML retinanet models int8 quantization Jan 7, 2022
@lidanqing-intel lidanqing-intel changed the title Issue accuracy: BML retinanet models int8 quantization Accuracy diff BML retinanet models int8 quantization Jan 19, 2022
@lidanqing-intel lidanqing-intel removed their assignment Feb 14, 2022
@lidanqing-intel lidanqing-intel modified the milestones: Q4, 2022 Q1 Feb 14, 2022
@lidanqing-intel lidanqing-intel self-assigned this Feb 21, 2022
@lidanqing-intel lidanqing-intel changed the title Accuracy diff BML retinanet models int8 quantization Accuracy diff of Retinanet int8 quantization May 23, 2022
@lidanqing-intel lidanqing-intel changed the title Accuracy diff of Retinanet int8 quantization Accuracy diff of Retinanet int8 May 23, 2022
@lidanqing-intel lidanqing-intel changed the title Accuracy diff of Retinanet int8 Accuracy diff of Retinanet int8 and Retinanet FP32 May 23, 2022
@lidanqing-intel
Copy link
Contributor Author

I got a new quant retinanet model from Baidu Jiajun, I tried using old PaddleDetecion infer.py and new PaddleDetection. However, both version could not work with this quant model. I m still checking if I could fix something to make it work and at the same time asked Jiajun.

@lidanqing-intel lidanqing-intel removed their assignment Aug 1, 2022
@paddle-bot paddle-bot bot added the status/close 已关闭 label Sep 6, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@wozna wozna

Labels
Intel status/close 已关闭
Projects
None yet
Milestone
2022 Q1
Development

No branches or pull requests
4 participants
@jiangjiajun @pangyoki @wozna @lidanqing-intel