An example where distribution is extended

[cranmer]

I assume that we can add a distribution without modifying the carl package, but I haven't checked.
If so, an example notebook with that would be nice.

[glouppe]

Yes this is possible and quite straightforward (i really need to write documentation... !).

For the likelihood-free part (approximating ratios), the only requirement is that the object that represents the generator exposes a rvs(n_samples) method that returns an array of samples.

For implementing an actual distribution, the requirement is to implement the carl.distributions.DistributionMixin interface. And for nice interactions with the theano world, one should inherit from carl.distributions.TheanoDistribution.