ROS Resources: Documentation | Support | Discussion Forum | Index | Service Status | ros @ Robotics Stack Exchange
Ask Your Question

Autoware - How to train my own pointpillars model ?

asked 2019-11-21 03:26:00 -0600

Mackou gravatar image

Hello everyone.

  • Autoware version : 1.13.0-alpha.1-8-gbebcb65
  • ROS Version Melodic
  • Autoware installed from source

I am trying to use pointpillars with autoware and I would like to train my own model, because I would like to train a more complete model with other datasets.

First of all it's very hard to find anything on how to train this model, and to make it compatible with autoware. I used this git repo : SmallMunich/nutonomy_pointpillarsto train the dataset and convert it to the ONNX file format but it doesn't work. Here is the pointpillars log :

Can you help me reproduce the result with the KITTI dataset first and get ONXX files working ? Thanks a lot !

edit retag flag offensive close merge delete


I think right place to search for this is github issues and if you don't find anything then ask for help there.

Choco93 gravatar image Choco93  ( 2019-11-21 05:18:02 -0600 )edit

hi, Thanks for your share, I also wanna use SmallMunich/nutonomy_pointpillars to train and gain the ONNX file to locate in Autoware, so have you succeeded in doing that?

ryontang gravatar image ryontang  ( 2021-01-13 05:55:59 -0600 )edit

Hi, @Mackou, Maybe we can have some communication on this issue, which is located model in Autoware trained by ourselves.

ryontang gravatar image ryontang  ( 2021-01-13 06:01:31 -0600 )edit

Hey @ryontang and sorry for the late reply. I have managed to train my model using the SmallMunich repo. But I have never been able to output the trained model in a format that autoware could accept. I know that a team member rewrote the pointpillar node of autoware from scratch so it could accept this model.

Mackou gravatar image Mackou  ( 2021-01-21 03:49:44 -0600 )edit

Hi, @Mackou, thanks for your reply. You mean there is a way that need to rewrite the pointpillar node, actually it is a little dificulty for me, all right, keep in touch, if we have made any progress in this problem.

ryontang gravatar image ryontang  ( 2021-01-25 21:08:24 -0600 )edit

@ryontang We have totally stopped working on this problem. And unfortunately the new node is proprietary so I cannot share it here.

Mackou gravatar image Mackou  ( 2021-01-26 02:22:40 -0600 )edit

@Mackou, All right, sorry for the late reply, and thanks for your share about the problems you have met !!!

ryontang gravatar image ryontang  ( 2021-02-09 18:32:50 -0600 )edit

Hi, @Mackou, We have the same error ingormation when load the ONNX file exported by git repo : SmallMunich/nutonomy_pointpillars. I am sorry, could you give me some advice on rewritting pointpillar node of autoware. so it could accept this model.

ryontang gravatar image ryontang  ( 2021-03-19 04:48:09 -0600 )edit

3 Answers

Sort by » oldest newest most voted

answered 2019-11-25 04:03:52 -0600

kosuke_murakami gravatar image

I used this repo. But it is not well maintained. Hope this will help.

edit flag offensive delete link more


@kosuke_murakami Thanks for your answer ! How did you convert the model to ONNX ?

Mackou gravatar image Mackou  ( 2019-11-28 02:54:46 -0600 )edit

From the issue of this github project, it seems the ONNX file have some mistakes

ryontang gravatar image ryontang  ( 2021-01-13 05:58:06 -0600 )edit

answered 2019-12-11 03:45:31 -0600

ssfs gravatar image

updated 2019-12-11 05:30:10 -0600

Hi @kosuke_murakami , I am trying to train your repo bu there are some problems. When train using configs/pointpillars/car/xyres_16.config file, code falling the assertion, your comments;

assert 10 == len(ret_dict), f"something wrong with training output size!"

How can I fix this issues, or could you provide proper config file?

edit flag offensive delete link more

answered 2022-03-28 09:01:49 -0600

Hi, I have the same issue of yours that appears in the log files

WARNING: ONNX model has a newer ir_version (0.0.4) than this parser was built against (0.0.3).

it seems that the generated onnx files from the SmallMunich/nutonomy_pointpillars have a different version than the Autoware accept

Did you solve that ? do you know how to convert these versions ?

edit flag offensive delete link more

Question Tools



Asked: 2019-11-21 03:26:00 -0600

Seen: 1,206 times

Last updated: Dec 11 '19