What happened to plot_hdi() ?

[pranavduraisamy]

Hi peeps,

After a long time, I'm back to sampling new models today, and I just realized a lot has changed since then. I'm happy to see that certain functions have greatly improved, and thanks a lot to the devs for that!!

One problem I faced was with plot_hdi(). I couldn't find any information about it in the migration guide either..,

I’d really appreciate any help or clarification on this. Thanks in advance😁

Traceback (most recent call last):
  File "C:\Users\USER\.conda\envs\ghost\Lib\site-packages\marimo\_runtime\executor.py", line 139, in execute_cell
    return eval(cell.last_expr, glbls)
  Cell 
marimo://studies\code\17-true-glm-brute.py#cell=cell-17

, line 13, in <module>
    az.plot_hdi(_x,_ms)
    ^^^^^^^^^^^
  File "C:\Users\USER\.conda\envs\ghost\Lib\site-packages\arviz\__init__.py", line 118, in __getattr__
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
AttributeError: module 'arviz' has no attribute 'plot_hdi'. Did you mean: 'plot_bf'?

[aloctavodia]

We need to update the migration guide then. Try using "plot_lm", happy to help if you need some assistance with that.

[pranavduraisamy]

Thanks! I tried to create a subplot and just realized a lot of things have changed. Happy to see that customizing these plots got much easier with visuals and aesthetics... Now I'm playing with PlotCollection, reading the docs and trying to figure it out how it works...

[aloctavodia]

Great. Feedback is welcome, including pointing out missing elements in the docs.

[pranavduraisamy]

In previous versions, I could use axes with plot_lm() or anything else... But now I’m struggling a bit to create subplots with different InferenceData objects. Is there any way around this? Can you help me sort it out, please?

Or any example to demonstrate the same could also help...

Currently I'm doing everything manually like i attached., But I noticed when using plot_lm() it smoothens the corridor pretty well.

x_vals   = np.linspace(0, 4, 50)
colors   = ["dodgerblue", "red"]

fig, axes = plt.subplots(6, 2, figsize=(12, 22), sharey='row',sharex=True, constrained_layout=True)

for _a in range(6):
    for _b in range(2):
        ax = axes[_a, _b]
        for _c in range(2):
            with model:
                pm.set_data({
                    "x": x_vals,
                    "c_idx": np.full(50, _c, dtype=int),
                    "b_idx": np.full(50, _b, dtype=int),
                    "a_idx": np.full(50, _a, dtype=int),
                })
                _idata = pm.sample_posterior_predictive(idata, return_inferencedata=True, extend_inferencedata=True, predictions=True, progressbar=False);
            preds = _idata.predictions["obs"].values.reshape(-1, 50)
            ax.fill_between(x_vals, np.quantile(preds, 0.05, axis=0), np.quantile(preds, 0.95, axis=0), color=colors[_c], alpha=0.18, linewidth=0)
            ax.plot(x_vals, preds.mean(axis=0), color=colors[_c], lw=1.5)
plt.show()

[aloctavodia]

The new plotting logic in ArviZ has been inspired by the grammar of graphics. So the mindset is that we typically first build or transform the data (usually a DataTree) that we want to visualize, and then the plotting functions construct the figure by mapping dimensions and coordinates to aesthetics and visual layers. So less looping and more mapping.

For instance, something like this should work.

idata_0 = az.load_arviz_data("centered_eight")
idata_1 = az.load_arviz_data("non_centered_eight")

combined = xr.concat(
    [idata_0, idata_1],
    dim=xr.DataArray(["model_0", "model_1"], dims="__model__")
)

az.plot_autocorr(combined, var_names=["mu"]);

Something like this will not currently work with plot_lm (see arviz-devs/arviz-plots#508). I am also checking systematically when this pattern fails/works for other functions.

Also notice that some functions, like plot_dist, accept a dictionary of DataTrees as the main input, specifically intended for plotting different models.