Though I think CrabNet might need to be refitted for new samples (i.e. if you specify N=10, then you only get 10 samples from the posterior, to get more would probably require refitting, and not sure if these would be directly comparable to the 10 from the first run). Also not exactly sure how this could be converted to individual predictions. Maybe just some basic plumbing in and after:
https://aclanthology.org/2020.emnlp-main.17.pdf
Though I think CrabNet might need to be refitted for new samples (i.e. if you specify
N=10, then you only get10samples from the posterior, to get more would probably require refitting, and not sure if these would be directly comparable to the10from the first run). Also not exactly sure how this could be converted to individual predictions. Maybe just some basic plumbing in and after:https://github.com/anthony-wang/CrabNet/blob/9e0d79c5bff56ceae0600015942c54214d78152f/crabnet/kingcrab.py#L151-L157