AdvancedHMC Extension

The first choice when sampling from the model/image posterior, is AdvancedHMC, which uses Hamiltonian Monte Carlo to sample from the posterior. Specifically, we usually use the NUTS algorithm.

The interface to AdvancedHMC is very powerful and general. To simplify the procedure for Comrade users, we have provided a thin interface. A user needs to specify a sampler and then call the sample function.

To sample a user can use follow the standard AdvancedHMC interface, e.g.,

chain = sample(post, NUTS(0.8), 10_000; n_adapts=5_000)

In addition our sample call has a few additional keyword arguments:

• adtype = Val(:Zygote): The autodiff package to use. Currently the default is Zygote and we recommend using this. Note that you must load Zygote before calling sample.

• saveto = MemoryStore(): Specifies how to store the samples. The default is MemoryStore which stores the samples directly in RAM. For large models this is not a good idea. To save samples periodically to disk use DiskStore, and then load the results with load_samples.

Note that like most AbstractMCMC samplers the initial location can be specified with the initial_params argument.

Example

using Comrade
using AdvancedHMC
using Zygote

# Some stuff to create a posterior object
post # of type Comrade.Posterior

out = sample(post, NUTS(0.9), 2_000; n_adapts=1_000, saveto=DiskStore())
chain = load_samples(out)