ViTPose using a pretrained model.

[amartincirera]

Hello,

I'm attempting to train a model using the VitPose architecture starting from a pretrained model and using a custom dataset with a skeleton and a different number of key points that used for human pose. I'm encountering the following message:

AssertionError: The length of pos_embed (197) doesn't match the expected shape (14*14+0). Please check the img_size argument.

However, when I do not use a pretrained model, the training proceeds without any messages.

These are the versions I am using on my system:

torch version: 2.0.1 True
torchvision version: 0.15.2
mmpose version: 1.1.0
cuda version: 11.8
compiler information: GCC 9.3

Is there any way to train a pretrained ViTPose model without getting this message?

Thank you very much for your assistance.

[52THANOS]

i got the same issue

[Ben-Louis]

Sorry for the late reply. This issue has been resolved in the latest version. Please install mmpose 1.2.0 and use the config from that version. Additionally, make sure you use mmpretrain>=1.0.0.

[amartincirera]

Thank you so much!