Skip to content

Speed up infer() via vectorised batch over targets #49

Description

@robaru

bayesian_listener.py > function infer

For repetitions > 1, replace the Python loop over target_num with a single batched call to multiple_logpdfs_vec_input_single_cov. Draw all noise at once (target_num × repetitions × P), reshape to (T*R, P), call the Numba kernel once, reshape back.

Potential issue: intermediate loglik array is (T*R, n_templates) — ~5 GB for 793 targets × 50 reps × 2112 templates. May need a chunked fallback for large inputs.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions