Getting NaNs when running plot_trace

Describe the bug I’m getting a ValueError: cannot convert float NaN to integer when running plot_trace on my NumPyro sampled data. As far as I can neither my data nor samples contain any NaNs.

To Reproduce

I’m not able to share any code and data however I’ve done some debugging that could lead to where the error is coming from:

Full Stacktrace:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<timed exec> in <module>

/project/venv/lib/python3.9/site-packages/arviz/plots/traceplot.py in plot_trace(data, var_names, filter_vars, transform, coords, divergences, kind, figsize, rug, lines, circ_var_names, circ_var_units, compact, compact_prop, combined, chain_prop, legend, plot_kwargs, fill_kwargs, rug_kwargs, hist_kwargs, trace_kwargs, rank_kwargs, labeller, axes, backend, backend_config, backend_kwargs, show)
    252 
    253     plot = get_plotting_function("plot_trace", "traceplot", backend)
--> 254     axes = plot(**trace_plot_args)
    255 
    256     return axes

/project/venv/lib/python3.9/site-packages/arviz/plots/backends/matplotlib/traceplot.py in plot_trace(data, var_names, divergences, kind, figsize, rug, lines, circ_var_names, circ_var_units, compact, compact_prop, combined, chain_prop, legend, labeller, plot_kwargs, fill_kwargs, rug_kwargs, hist_kwargs, trace_kwargs, rank_kwargs, plotters, divergence_data, axes, backend_kwargs, backend_config, show)
    286                     aux_plot_kwargs = dealiase_sel_kwargs(plot_kwargs, compact_prop_iter, sub_idx)
    287                     aux_trace_kwargs = dealiase_sel_kwargs(trace_kwargs, compact_prop_iter, sub_idx)
--> 288                     ax = _plot_chains_mpl(
    289                         ax,
    290                         idy,

/project/venv/lib/python3.9/site-packages/arviz/plots/backends/matplotlib/traceplot.py in _plot_chains_mpl(axes, idy, value, data, chain_prop, combined, xt_labelsize, rug, kind, trace_kwargs, hist_kwargs, plot_kwargs, fill_kwargs, rug_kwargs, rank_kwargs, circular, circ_var_units, circ_units_trace)
    486             aux_kwargs = dealiase_sel_kwargs(plot_kwargs, chain_prop, chain_idx)
    487             if not idy:
--> 488                 axes = plot_dist(
    489                     values=row,
    490                     textsize=xt_labelsize,

/project/venv/lib/python3.9/site-packages/arviz/plots/distplot.py in plot_dist(values, values2, color, kind, cumulative, label, rotated, rug, bw, quantiles, contour, fill_last, figsize, textsize, plot_kwargs, fill_kwargs, rug_kwargs, contour_kwargs, contourf_kwargs, pcolormesh_kwargs, hist_kwargs, is_circular, ax, backend, backend_kwargs, show, **kwargs)
    210 
    211     plot = get_plotting_function("plot_dist", "distplot", backend)
--> 212     ax = plot(**dist_plot_args)
    213     return ax

/project/venv/lib/python3.9/site-packages/arviz/plots/backends/matplotlib/distplot.py in plot_dist(values, values2, color, kind, cumulative, label, rotated, rug, bw, quantiles, contour, fill_last, figsize, textsize, plot_kwargs, fill_kwargs, rug_kwargs, contour_kwargs, contourf_kwargs, pcolormesh_kwargs, hist_kwargs, is_circular, ax, backend_kwargs, show)
     91         legend = label is not None
     92 
---> 93         ax = plot_kde(
     94             values,
     95             values2,

/project/venv/lib/python3.9/site-packages/arviz/plots/kdeplot.py in plot_kde(values, values2, cumulative, rug, label, bw, adaptive, quantiles, rotated, contour, hdi_probs, fill_last, figsize, textsize, plot_kwargs, fill_kwargs, rug_kwargs, contour_kwargs, contourf_kwargs, pcolormesh_kwargs, is_circular, ax, legend, backend, backend_kwargs, show, return_glyph, **kwargs)
    259                 bw = "experimental"
    260 
--> 261         grid, density = kde(values, is_circular, bw=bw, adaptive=adaptive, cumulative=cumulative)
    262         lower, upper = grid[0], grid[-1]
    263 

/project/venv/lib/python3.9/site-packages/arviz/stats/density_utils.py in kde(x, circular, **kwargs)
    490         kde_fun = _kde_linear
    491 
--> 492     return kde_fun(x, **kwargs)
    493 
    494 

/project/venv/lib/python3.9/site-packages/arviz/stats/density_utils.py in _kde_linear(x, bw, adaptive, extend, bound_correction, extend_fct, bw_fct, bw_return, custom_lims, cumulative, grid_len, **kwargs)
    583         grid, pdf = _kde_adaptive(x, bw, grid_edges, grid_counts, grid_len, bound_correction)
    584     else:
--> 585         grid, pdf = _kde_convolution(x, bw, grid_edges, grid_counts, grid_len, bound_correction)
    586 
    587     if cumulative:

/project/venv/lib/python3.9/site-packages/arviz/stats/density_utils.py in _kde_convolution(x, bw, grid_edges, grid_counts, grid_len, bound_correction, **kwargs)
    704     grid = (grid_edges[1:] + grid_edges[:-1]) / 2
    705 
--> 706     kernel_n = int(bw * 2 * np.pi)
    707     if kernel_n == 0:
    708         kernel_n = 1

ValueError: cannot convert float NaN to integer

The bw parameter from computing the density seems to become NaN at a certain point. From the debugger:

> /project/venv/lib/python3.9/site-packages/arviz/stats/density_utils.py(585)_kde_linear()
    583         grid, pdf = _kde_adaptive(x, bw, grid_edges, grid_counts, grid_len, bound_correction)
    584     else:
--> 585         grid, pdf = _kde_convolution(x, bw, grid_edges, grid_counts, grid_len, bound_correction)
    586 
    587     if cumulative:

ipdb>  bw
nan

It seems that the error comes from the kde_fun call, at this point there are no NaNs anywhere in the arguments:

> /project/venv/lib/python3.9/site-packages/arviz/stats/density_utils.py(492)kde()
    490         kde_fun = _kde_linear
    491 
--> 492     return kde_fun(x, **kwargs)
    493 
    494 

ipdb>  x
array([0.06682166, 0.06682151, 0.0668215 , 0.06682149, 0.06682152,
       0.06682113, 0.0668212 , 0.06682108, 0.06682114, 0.06682119,
       0.06682099, 0.06682099, 0.06682114, 0.06682115, 0.06682112,
       0.06682058, 0.0668207 , 0.06682093, 0.06682101, 0.06682117,
       0.06682105, 0.06682053, 0.06682035, 0.06682046, 0.06682055,
       0.06682087, 0.06682237, 0.06682225, 0.06682158, 0.06682166,
       0.06682136, 0.06682222, 0.06682229, 0.06682215, 0.06682245,
       0.06682227, 0.06682238, 0.06682224, 0.06682215, 0.06682251,
       0.0668226 , 0.06682282, 0.06682256, 0.06682279, 0.06682241,
       0.06682275, 0.06682276, 0.06682317, 0.0668232 , 0.06682297],
      dtype=float32)
ipdb>  kwargs
{'bw': 'experimental', 'adaptive': False, 'cumulative': False}

Expected behavior No NaNs to appear in plot_trace when there are no NaNs in the input data.

Additional context Installed the Arviz development version via GitHub pip install git+git://github.com/arviz-devs/arviz.git

numpyro   : 0.7.2
matplotlib: 3.4.3
sys       : 3.9.7 | packaged by conda-forge | (default, Sep  2 2021, 17:58:34) 
[GCC 9.4.0]
xarray    : 0.19.0
seaborn   : 0.11.2
arviz     : 0.11.3
jax       : 0.2.20

[ahartikainen]

https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance

Because SumSq and (Sum×Sum)/n can be very similar numbers, cancellation can lead to the precision of the result to be much less than the inherent precision of the floating-point arithmetic used to perform the computation. Thus this algorithm should not be used in practice

[peterroelants]

Closed by https://github.com/arviz-devs/arviz/pull/1833