pair plot with 2 variables looks different than >2

[jecampagne]

Hello,
Here is a snippet

!pip install matplotlib==3.5.0
import matplotlib as mpl
#print(mpl.__version__)   #verif 3.5.0 
import matplotlib.pyplot as plt
import arviz as az
#print(az.__version__) #verif 0.12.1
import numpy as np

def generator(dim):
  mean = [0]*dim
  A = np.random.rand(dim, dim)
  cov = np.dot(A, A.transpose())
  out = np.random.multivariate_normal(mean, cov, 5000).T
  data = {}
  for i in range(len(out)):
    data[f"var_{i}"] = out[i]
  return data

data = generator(dim=2)
axs= az.plot_pair(
            data,
            kind="kde",
            figsize=(5,5),
            marginal_kwargs={"plot_kwargs": {"c": "b"}},
            kde_kwargs={
                "hdi_probs": [0.68, 0.9],
                "contour_kwargs":{"colors":('r', 'green', 'blue'), "linewidths":3},
                "contourf_kwargs":{"alpha":0},
            },
            point_estimate_kwargs={"lw": 3, "c": "b"},
            marginals=True, textsize=10, point_estimate="median",
        );

gives

while with

data = generator(dim=3)

the pair_plotlooks like

Q1: I would like to get the "dim=3" drawing of the marginals for the "dim=2" use case, is-it possible?
Q2: As a side remark, I do not manage to use az.convert_XYZ function to get something working for az.pair_plot it is wht I have used a home-made "data" generator. It would be nice if there exists a simple example to get it right that could be nice.

Thanks.

[OriolAbril]

It is currently not possible to avoid rotating the marginals when there are only two of them. It is hardcoded in https://github.com/arviz-devs/arviz/blob/main/arviz/plots/backends/matplotlib/pairplot.py#L162-L171 in case you'd be interested in sending a PR for this.

I believe I can help with the second question though. When using az.convert_to_inference_data or az.from_dict, ArviZ needs a convention in order to name each of the available dimensions. That convention is:

• 1D array -> draw
• 2D array -> chain, draw
• 3+D array -> chain, draw, *shape. The *shape part means taking as many extra dims as needed and naming them {var_name}_dim_0, {var_name}_dim_1... Moreover, if no var name is provided (aka, passing an array to convert_to_inference_data), x is used as variable name.

I imagine your code didn't work initially because you were giving a 2d array but it had draw, dim_0 dimensions, not chain, draw. Converting that to a dict of 1D arrays fixed it because now draw matched the ArviZ convention. You could update your code to use a 3D array:

def generator(dim):
   mean = [0]*dim
   A = np.random.rand(dim, dim)
   cov = np.dot(A, A.transpose())
   return np.random.multivariate_normal(mean, cov, size=(1, 5000))

so when you do az.convert_to_inference_data(data) (or pass data straight to ArviZ functions) you will get:

<xarray.Dataset>
Dimensions:  (chain: 1, draw: 5000, x_dim_0: 2)
Coordinates:
  * chain    (chain) int64 0
  * draw     (draw) int64 0 1 2 3 4 5 6 7 ... 4993 4994 4995 4996 4997 4998 4999
  * x_dim_0  (x_dim_0) int64 0 1
Data variables:
    x        (chain, draw, x_dim_0) float64 0.2053 0.1567 ... 1.253 1.416
Attributes:
    created_at:     2023-01-25T00:07:24.536681
    arviz_version:  0.14.0

and plots will work.

Personally, I would recommend defining the variable name yourself and converting explicitly, for example with `from_dict({"var": data})