Custom keypoint detector

[telliod]

Hi,

I've been able to train my custom keypoint detector using detectron2. The number of keypoints per object is 5.
These objects are not human.
For the configuration, I have these parameters for keypoints :

cfg.merge_from_file(model_zoo.get_config_file("COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x.yaml"))
keypoint_names = ['a', 'b', 'c', 'd', 'e']
keypoint_flip_map = [('a', 'b'), ('a', 'c'), ('a', 'd'),('a','e')]
cfg.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS = len(keypoint_names)
cfg.TEST.KEYPOINT_OKS_SIGMAS = np.ones((len(keypoint_names), 1), dtype=float).tolist()

Nevertheless during inference, I have this message :

Skip loading parameter 'roi_heads.keypoint_head.score_lowres.weight' to the model due to incompatible shapes: (512, 5, 4, 4) in the checkpoint but (512, 17, 4, 4) in the model! You might want to double check if this is expected.
Skip loading parameter 'roi_heads.keypoint_head.score_lowres.bias' to the model due to incompatible shapes: (5,) in the checkpoint but (17,) in the model! You might want to double check if this is expected.
Some model parameters or buffers are not found in the checkpoint:

And the keypoint detection failed with pictures given for inference. It tries to put more than 5 keypoints in bbox.
The 17 keypoints come from the human keypoints detection by default.
Does it come from the "keypoint_rcnn_R_50_FPN_3x.yaml" file (that contains model architecture maybe for human body ) or from the cfg.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS parameter?
How to train the model from scratch, or to set up with only 5 keypoints ?

A similar problem was also described in #702

Thanks for your help