Output issue after training

[s-kav]

Hello! I saw your great TF realization. I would be thankful for your help. Now, I'm realizing a similar task for the BlazePose model in TF2. I use another architecture of the model, but it's the same as yours. I'm using this model with pre-trained weights by this graph. And I train this model on our specific dataset, after augmentation we have more 30 000 images.
But, one problem appears during this process: as an output, we have a set of points (coordinates), which is the same for any input, so all coordinates are constant. Perhaps, you have some ideas or direction of thinking for this?
Thank you for your work and help.