Open AdityaSavara opened 2 years ago
Currently, the discrete_chains_samples do not stack when continue sampling is done (only the flattened post burn in samples gets extended). This means that continueSampling will currently result in convergence diagnostics only reflecting the "newest" sampling.
Continue sampling seems to be working for ESS. And, there is now a warning if the number of walkers has changed between samplings. However, there is not code for handling the cases of different numbers of walkers. For example, going from 1 to 8 or the other way. In either situation, the walkers starting points should be "resampled" from the existing post-burn-in posterior, to have the correct number with correct statistical representation.