Fix normalization of initial guess in linear_cg()

cornellius-gp / linear_operator

A LinearOperator implementation to wrap the numerical nuts and bolts of GPyTorch

MIT License

95 stars 28 forks source link

Fix normalization of initial guess in linear_cg() #71

Closed Turakar closed 1 year ago

Turakar commented 1 year ago

Seems like while rhs was normalized, initial_guess was not. As such, initializing linear_cg() with a good initial guess (e.g. the solution) was of no use. This PR also normalizes initial_guess and adds a corresponding test.

Turakar commented 1 year ago

On a side note: For my example, initializing linear_cg() with the solve from the previous iteration of Gaussian Process training with skip_logdet_forward(True) was of no use regarding convergence speed.