Deploying last RTMW-x model

[davidpagnon]

Is there a way to have access to a full list of the models available under https://download.openmmlab.com/ domain? I believe directory listing is disabled and unfortunately, it seems like project pages are not always updated, so model zoos are not extensive.

For example, this model exists but cannot be found anywhere unless you know the right address:
https://download.openmmlab.com/mmpose/v1/projects/rtmw/onnx_sdk/rtmw-x_simcc-cocktail13_pt-ucoco_270e-384x288-0949e3a9_20230925.zip

[Tau-J]

Hi @davidpagnon , the model list under this URL is not public. Only the person in charge of the corresponding openmmlab repository has access. Since I used to be the maintainer of mmpose, I used it to store some rtmpose series models for easy downloading by community users. The ONNX model you mentioned was converted from the publicly available rtmw-x pth.

I don't understand why you want the entire list of models from mmpose, but you can consider that all models with downloadable links found in the mmpose repository are the ones publicly available to the community users.

[davidpagnon]

Hi @Tau-J,

Thank you for your quick answer! You are right, I definitely don't need the full list of models. Moreover, I pasted the wrong link as an example, sorry about this confusion. Long story short, I think I figured out something.

The address of the RTMW-x onnx file you provide in RTMlib Readme seems outdated.
In our experience, its results are worse than the RTMW-l ones. The reason might be that it has been trained on cocktail14 instead of cocktail13 (also, the release date is earlier). This might be why this link is not available on the RTMPose repo anymore.

We deployed the newer RTMW-x model listed in RTMPose Readme and it works better.
But it is not available on openmmlab as a zip file.

Could you deploy it and update it to openmmlab? It would be cleaner than having it on a baidu or google drive.

In the same vein, the RTMW-x model with 256x192 input size that you provide in the balanced Wholebody solution has been updated as well, so I suppose you could deploy it, too!

Thank you again for your good work, RTMlib is a breeze to install and still flexible, which is much appreciated!

[Tau-J]

Apologies for the late reply. As I am no longer an employee of openmmlab, I have lost access to the official cloud, and my new work has left me with little time to update rtmlib. I will try my best to contact my former colleagues and ask them to upload the new rtmw models.

[davidpagnon]

Oh okay, thanks, I appreciate it!
I hope you or other people will find a way to update RTMlib when needed, because it really is the only option for running state-of-the-art models with a straightforward install, which is incredibly useful. Take it as a pretty clumsy way to thank you again :)

(By the way, it seems like MMpose had slowed down, and some people who wanted to get involved never got an answer. Is it just a feeling, or is there a chance that openmmlab may shut down?)

[davidpagnon]

Related to this, actually:
#14

[Tau-J]

Hi @davidpagnon . OpenMMLab is unlikely to be shut down, but the focus of those who were originally responsible for maintaining these repositories has shifted to large language model work, specifically the InternLM series. The development team that was originally responsible for MMDeploy is now primarily in charge of Xtuner and LMDeploy. The world is changing rapidly.

[davidpagnon]

Okay, thanks for your clarification, I understand a bit better! The world is changing at a scary rate indeed.
I hope there will still be people maintaining the pose and deploy series, as they are still tremendously useful for many of us.

[MiXaiLL76]

open-mmlab/mmdetection#11815

@Tau-J Hi! Can I somehow join the mmlab team in the direction of CV? mmocr \ mmdet \ mmsegm \ mmdeploy?

I lead hundreds of projects in industry using this framework and I want to update them and provide support

[davidpagnon]

In connexion with this: #36 (comment)
The model provided in the balanced wholebody solution was faulty, it detected the toes near the top of the head for a reason that escapes me.

I submitted a pull request that proposes another one: #45

[Tau-J]

Hi @davidpagnon , thank you for your pull request. I will merge it into the main branch as soon as possible to make it accessible to more users!

Regarding the issue with the wrong keypoint you mentioned, this was caused by a bug in the training of the older version (20230925) of the rtmw model. I realized and fixed it in the new version of the model (https://download.openmmlab.com/mmpose/v1/projects/rtmw/rtmw-x_simcc-cocktail14_pt-ucoco_270e-384x288-f840f204_20231122.pth). However, due to personal reasons, I haven't had the time to convert it to the ONNX format.

Would you be willing to help convert it to the ONNX format supported by rtmlib and verify if the issue has been resolved? I greatly appreciate your significant contributions to this project.

[davidpagnon]

Hi @Tau-J,
Thanks! I believe @hunminkim98 already deployed it there: https://huggingface.co/hunminkim98/Models_Pose2Sim/tree/main
Is this right HunMin?

[hunminkim98]

Hello, @davidpagnon, @Tau-J
Thanks for mentioning it.
Yes, you’re right. I’ll also convert and upload the models that haven’t been converted yet, such as https://download.openmmlab.com/mmpose/v1/projects/rtmw/rtmw-x_simcc-cocktail14_pt-ucoco_270e-256x192-13a2546d_20231208.pth

[Tau-J]

Hi @davidpagnon @hunminkim98 , I just merged your new contributions and released the 0.0.13 version to pypi. Thanks for your kind help.

[davidpagnon]

Awesome!