'message_hub' keyError when trying to finetune pretrained MMPose models.

[mmayank74567]

When resuming a model to finetune, MMengine expects to have a 'message_hub' key in the checkpoint.
However, it seems like MMPose does not save the 'message_hub' when saving the torch model (reference) This causes the following issue:

  File "/mmengine/mmengine/runner/runner.py", line 2075, in resume
    self.message_hub.load_state_dict(checkpoint['message_hub'])
KeyError: 'message_hub'

[vieenrose]

Hi mmayank74567,

We just got RTMO-s pre-trained weights reproduced in keeping all keys inside PyTorch checkpoint.
You may check Luigi/Retrainable-RTMO-s to get it if you need.