v3.0.0dev0 blobs failing when different shapes

General information:

• emcee version: v3.0.0dev0
• platform: any
• installation method (pip/conda/source/other?): any

Problem description:

Expected behavior:

I expect that returning blobs of various shapes should work.

Actual behavior:

An error is raised: ValueError: setting an array element with a sequence.

What have you tried so far?:

When passing blobs_dtype with shapes for each returned blob, it works fine.

Minimal example:

import emcee

import numpy as np
def lnl(p):
    return p[0], 1.0, np.array([1,2,3])
sampler = emcee.EnsembleSampler(
    nwalkers = 10,
    ndim = 2,
    log_prob_fn = lnl
)

p0 = np.random.normal(size=(10,2))
for sample in sampler.sample(p0, iterations=10):
    continue

The offending line is L394 of ensemble.py: dt = np.atleast_1d(blob[0]).dtype. This will work as intended (I think), if it is instead: dt = [(np.atleast_1d(b).dtype, np.atleast_1d(b).shape) for b in blob[0]], though there might need to be a check for 1-dimensional arrays.

[sibirrer]

yes @alessiospuriomancini, for example here: https://github.com/sibirrer/lenstronomy/blob/master/lenstronomy/Sampling/sampler.py in definition mcmc_emcee()

[alessiospuriomancini]

Thanks so much @sibirrer. Just to clarify: you use the script multiprocessing.py within some script for emcee, to directly run using emcee without going through CosmoHammer?