pairplot: documentation, more options and tutorial

[janfb]

The pairplot function is our main function for visualizing posteriors.

https://github.com/sbi-dev/sbi/blob/c3c2b6e142fb4c57d5599effc1f01f8222a37c57/sbi/analysis/plot.py#L280

It would be great to improve it bit:

• adding more documentation
• defaults that produce a decent default figure, even when plotting several distributions, e.g., prior vs posterior
• more options, see https://github.com/sbi-dev/sbi/blob/c3c2b6e142fb4c57d5599effc1f01f8222a37c57/sbi/analysis/plot.py#L328-L330 (use bins="auto" as default)
• a tutorial on how to use all the options 🥇

[famura]

I plan to have a look at it during the hackathon

[Matthijspals]

Also happy to help out here during the hackathon!

[Ziaeemehr]

It would be nice to have joint_plot for arbitrary pair of parameters:

def plot_joint(x, limits, ax, cmap="hot", label=None, points=[],
               xlabel="", ylabel="", add_corr=True):
    '''
    plot joint distribution of given samples
    
    Parameters
    ----------
    x : 2d-array
        samples from the distribution
    limits : list of tuples
        limits of the distribution
    '''
    
    density = gaussian_kde(x[:, [1, 0]].T, bw_method='scott')
    col = 1
    row = 0
    X, Y = np.meshgrid(
        np.linspace(limits[col][0], limits[col][1], 50,),
        np.linspace(limits[row][0], limits[row][1], 50))
    positions = np.vstack([X.ravel(), Y.ravel()])
    Z = np.reshape(density(positions).T, X.shape)
    # normalize Z
    Z = Z / np.max(Z)
    im = ax.imshow(Z, cmap=cmap,
              extent=(limits[col][0], limits[col][1],
                      limits[row][0], limits[row][1]),
                    origin="lower",
                    aspect="auto")
    divider = make_axes_locatable(ax)
    cax = divider.append_axes("right", size="5%", pad=0.04)
    plt.colorbar(im, cax=cax, ax=ax, ticks=[0, 1])

    if len(points) > 0:
        ax.scatter([points[1]], [points[0]], s=150, color='#6cf086', marker='*')
    ax.set_xlabel(xlabel)
    ax.set_ylabel(ylabel)
    ax.set_xticks([0,1])
    ax.set_yticks([0,1])
    if add_corr:
        corr = np.corrcoef(x[:, 0], x[:, 1])[0, 1]
        ax.text(0.6, 0.88, r"$\rho=$"+f"{corr:.1f}", fontsize=25,
                transform=ax.transAxes, color="white")

[michaeldeistler]

what exactly should joint_plot do?

[Ziaeemehr]

it actually extract the joint plot implemented on pairplot, give an option to user to select arbitrary pair of parameters. I had issue for preparing image for publications to have arbitrary panels arrangements.

[michaeldeistler]

we have pairplot(..., subset=[0, 2, 3]), does this work?

[Ziaeemehr]

I am not sure, think it still give a triangle plot right? including diagonal (marginal) and offdiagonal (joint plots). What if user only need one panel of this triangle plot (joint plot of parameter i, j).

[michaeldeistler]

ah, I see your point. I am not 100% convinced we need this, maybe we just leave this to users to implement themselves if really needed?

[Ziaeemehr]

no worries, that's just a suggestion. 👍