training my own dataset and problem

[ghost]

Hey all, i am aiming to do keypoints detection on my own dataset which include only two keypoints and not 17 like human pose estimation.

I tried to use the pre trained weights of COCO for human pose estimation, but of course I get the following error
"ValueError: Dimension 2 in both shapes must be equal, but are 2 and 17. Shapes are [2,2,2,512] and [2,2,17,512]. for 'Assign_724' (op: 'Assign') with input shapes: [2,2,2,512], [2,2,17,512]."

How can i solve this problem?

Thanks

[nmxnql]

I think maybe you need to filter the last layers of maskrcnn and set the keypoint_num to yours

[ghost]

I have tried but it didnt solve the problem. my dataset is not for human pose estimation, maybe using coco pretrained weights is wrong?

[dingfuzhou]

I have tried but it didnt solve the problem. my dataset is not for human pose estimation, maybe using coco pretrained weights is wrong?

Hi adarvit, have you solved this problem?