This is fantastic — well done. I have BC at least one application I’ve kept on least squares precisely because it can say no.

Can Keras/Tensorflow do this too? Seems like a massive advantage for PyTorch if they can’t.

Nvm, found it