Daniel Dodd
Daniel Dodd
Natgrads
This PR seeks to add natural gradients to GPJax, as well as two new Gaussian variational family parameterisations. Please check the type of change your PR introduces: - [ ]...
# Feature Request GPJax's SVGP computes variational expectations $\int \text{d} q(f(x_i)) \log( p(y_i|f(x_i))$ via Gauss-Hermite quadrature. This is superfluous in cases e.g., of a Gaussian likelihood when the integrand is...
# Feature Request - Define abstract class for basis function kernels. - Allow users to construct such kernels through specifying combinations of basis functions $-$ how the maths is written....
`sparse_gps.py` uses the following function in the `variational_expectation` method: ```python def log_prob(F, y): return self.likelihood.link_function(F, params["likelihood"]).log_prob(y) ``` it might better to move this to the `AbstractLikelihood` class.
## Pull request type This PR will add infrastructure for memory efficient operations with Gram matrices and allow custom solves. - [ ] Bugfix - [X ] Feature - [...
Perhaps we should not require parameter transformations (via `transform`) in objective functions - they should rest with model training. If I have an ELBO or the marginal log-likelihood, shouldn't I...
# Feature Request Desirable to check parameter shapes. For example, consider an ARD RBF kernel on two dimensions. Comparing two dictionaries `params_correct` where the length-scale parameter has the **correct shape**...
# Feature Request Write notebook with mathematical background and illustrate performance advantage of using natural gradients.
Introduce product and sum kernel computations (these are currently dense) by lazily building the gram matrix for each kernel, and then apply product and sum operations across a set of...
# Bug Report **GPJax version:** 0.7.2 Tagging @henrymoss. As noted in #402 there is issue with jitting objective. This functionality should be removed. ## Option (a) Remove these following lines...