In the current implementation of dot, the GP kernel hyperparameters are fixed (at hopefully sensible values). Ideally we'd let these parameters float, but in some of my tests it appears that the NUTS slows down by a factor of >10 when the covariance function's length scale or amplitude are defined as distributions. How do we get around this? Maybe @jejjohnson is interested?
In the current implementation of dot, the GP kernel hyperparameters are fixed (at hopefully sensible values). Ideally we'd let these parameters float, but in some of my tests it appears that the NUTS slows down by a factor of >10 when the covariance function's length scale or amplitude are defined as distributions. How do we get around this? Maybe @jejjohnson is interested?