Skip to content

External sampler should try harder at generating initial params #2739

@penelopeysm

Description

@penelopeysm

Turing currently generates initial params and passes it to external samplers:

# We need to extract the vectorised initial_params, because the later call to
# AbstractMCMC.step only sees a `LogDensityModel` which expects `initial_params`
# to be a vector.
initial_params_vector = varinfo[:]

However, this might be a bit too simplistic. The HMC implementation for example currently tries a lot of times to generate initial parameters until it can find something where the logp and gradient are finite. Just giving it one attempt might make it hard to actually get started with sampling.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions