Added support for emcee in arviz

Wanted to let you know we have added support for emcee to the arviz library - see https://github.com/arviz-devs/arviz/pull/302 for a description of the change.

Roughly, the plan is for arviz to replace plotting and diagnostics for pymc3 (so we do not have to rewrite it for pymc4), by converting data to a common data format (netCDF/xarray datasets), and then implementing all the plots and diagnostics on that object. emcee is the third library to have support (pymc3 and pystan), after hearing so many astronomers saying nice things about it at a conference.

In addition to the plots you can see in the pull request above, you can also calculate things like the effective sample size and the gelman rubin statistic. The summary function below shows a number of these.

Some questions I had -

• we store sample_stats from pystan and pymc3. Is there a way to access whether a sample was accepted at a per-draw basis in emcee?
• Are there particular diagnostics or plots you would like to see?

import arviz as az
from arviz.tests import helpers

sampler = helpers.emcee_linear_model(None, 500, 10)  # this is the sampler from http://dfm.io/emcee/current/user/line/

# This converter turns the data into a netCDF object that can be serialized and shared
data = az.from_emcee(sampler, var_names=['ln(f)', 'b', 'm'])

# all arviz functions should work with this object
az.summary(data).to_dataframe()

	ln(f)	b	m
metric
mean	-0.901523	4.186736	-0.493168
standard deviation	0.083640	0.387793	0.146651
mc error	0.002052	0.008238	0.003727
hpd 3.00%	-1.061138	3.448212	-0.759357
hpd 97.00%	-0.742526	4.923531	-0.201992
effective samples	1429.248983	1408.863122	1287.373414
gelman-rubin statistic	1.030000	1.030000	1.040000

[dfm]

I haven’t tested it thoroughly, but the assumptions are not satisfied. I’d recommend checking out this blog post for more discussion and recommended methods for ensemble methods: https://dfm.io/posts/autocorr/

[nedahejazi]

Hello, I seriously need help with emcee. I am trying to fit model spectra to observed ones. After working hard on code, the results are very strange. The chain values of parameters are the same for each walker. I mean the initial values do not change over the chain. Do you have any clue? If needed, I can send the code.