Note

Click here to download the full example code

Multimodal distributions

ChainConsumer can handle cases where the distributions of your chains are multimodal.

import numpy as np
import pandas as pd

from chainconsumer import Chain, ChainConsumer

First, let's build some dummy data

rng = np.random.default_rng(42)
size = 60_000

eta = rng.normal(loc=0.0, scale=0.8, size=size)

phi = np.asarray(
    [rng.gamma(shape=2.5, scale=0.4, size=size // 2) - 3.0, 3.0 - rng.gamma(shape=5.0, scale=0.35, size=(size // 2))]
).flatten()

rng.shuffle(phi)

df = pd.DataFrame({"eta": eta, "phi": phi})

To build a multimodal chain, you simply have to pass multimodal=True when building the chain.

chain_multimodal = Chain(
    samples=df.copy(),
    name="posterior-multimodal",
    multimodal=True,  # <- Here
)

Now, if you add this Chain to a plotter, it will try to look for sub-intervals and display them.

cc = ChainConsumer()
cc.add_chain(chain_multimodal)
fig = cc.plotter.plot()

Let's compare with what would happen if you don't use a multimodal chain. We use the same data as before but don't tell ChainConsumer that we expect the chains to be multimodal.

chain_unimodal = Chain(samples=df.copy(), name="posterior-unimodal", multimodal=False)

cc.add_chain(chain_unimodal)
fig = cc.plotter.plot()

Let's try with even more modes.

eta = np.asarray(
    [
        rng.normal(loc=-3, scale=0.8, size=size // 3),
        rng.normal(loc=0.0, scale=0.8, size=size // 3),
        rng.normal(loc=+3, scale=0.8, size=size // 3),
    ]
).flatten()


rng.shuffle(eta)

df = pd.DataFrame({"eta": eta, "phi": phi})

chain_multimodal = Chain(samples=df.copy(), name="posterior-multimodal", multimodal=True)

cc = ChainConsumer()
cc.add_chain(chain_multimodal)
fig = cc.plotter.plot()
fig.tight_layout()

Total running time of the script: ( 0 minutes 1.139 seconds)

Download Python source code: plot_7_multimodal_chains.py

Download Jupyter notebook: plot_7_multimodal_chains.ipynb

Gallery generated by mkdocs-gallery