For age, I used L1 loss since it’s a regression problem. I had to create a custom loss function to combine the losses from the tasks. You can see the code from the notebook: https://gist.github.com/yang-zhang/ec071ae4775c2125595fd80f40efb0d6#file-multi-face-ipyn

For landmarks, I think it’s also a regression problem (I’ve not worked with face landmarks but think it’s similar to object detection problem, which is a regression problem). So one would have to add one more component (perhaps mean squared error) to the custom loss.