Check if the distribution is a dirac

[Vayel]

@MaxHalford for ImageExplainer, it's always the case:

Both quantiles are equal to zero.

Is it expected?

[MaxHalford]

Do you mean that bias_low == bias_high even when n_samples > 1?

[Vayel]

I mean that bias(tau = -1) == bias(tau = 1). But it's because I only checked black pixels...

It still raises the question of raising a warning when we have a dirac distribution (which is the case for pixels that are always black across samples).

[MaxHalford]

Yeah indeed, there is no variation whatsoever if the pixels are black.

[Vayel]

To avoid having the warning for images, we can just have it for local bias. Actually it makes sense because if the whole dataset is a dirac, we can consider it's the user's mistake. But for local bias, we sample a subset of the dataset so we need to make sure we do it properly.

[MaxHalford]

We should issue a warning when a variable has only one unique value, because typically that is an edge-case. However, for images this case arises quite regularly, and so we shouldn't issue a warning.

It seems that we can catch warnings in Python and suppress them. See this StackOverflow post. I think that we can use the category parameter in the warnings.simplefilter.