Wrapper class for samples

[aagrande]

This adds a simple wrapper class for the main output of Treeffuser: the samples.

Subclassing and custom methods

EDIT: see update below by @ANazaret . Samples subclasses ndarray. Samples:

• uses view casting to cast an existing standard ndarray as an instance of Samples;
• ensures downstream compatibility with the NumPy API by preserving the same function signatures of existing methods (details here);
• adds methods to simplify the computation of sample statistics (e.g. sample_mean) and custom methods to compute correlation matrices, confidence intervals, range, etc..

See NumPy's documentation on subclassing for further reference.

Usage

The goal is to facilitate the user experience when computing estimates from Treeffuser samples, such as mean, quantiles, correlation matrices, modes, etc.

samples = model.sample(x, n_samples=100)

samples = Samples(samples)
samples.sample_mean() # conditional mean
samples.sample_quantile(q=[0.05, 0.95])  # conditional quantiles
samples.sample_kde(bandwidth="scott) # conditional KDEs
...

[ANazaret]

Thanks @aagrande ! For the record, we chose to not define Samples as a subclass of np.array because the numpy inherited functions were enabling the user to transform a Samples object into a corrupted state: example taking the transpose was returning a non-sensical Samples object.