In particular we could add the SGNHT algorithm (paper, code).
And perhaps also SGHMC and SGHNT algorithms that use the BAOAB integrator rather than Euler-Maruyama, comprehensive info in Leimkuhler and Matthews textbook (p271 and p350)
In all cases we should be carefully implement with a temperature parameter that accepts the value 0 and to only evaluate the gradient once per iteration.
SamDuffield
commented
1 month ago
In particular we could add the SGNHT algorithm (paper, code).
And perhaps also SGHMC and SGHNT algorithms that use the BAOAB integrator rather than Euler-Maruyama, comprehensive info in Leimkuhler and Matthews textbook (p271 and p350)
In all cases we should be carefully implement with a temperature parameter that accepts the value 0 and to only evaluate the gradient once per iteration.