can you share an example of icdf with negative binomial and poisson distros?

[jeycervantes]

Hi,

I am looking into this library since it seems to support icdf for these distros but after installing and using in databricks it says these methods aren't implemented?

I see they should be implemented here however:

gluonts/src/gluonts/torch/distributions/negative_binomial.py

Line 57 in 3d05d46

def icdf(self, value: torch.Tensor) -> torch.Tensor:

Sorry if I just missed it, is there some demo snippet of code around this use-case? Thank you!

Here is some sample code below (that does not work)
from gluonts.mx import distribution
import mxnet as mx

Example parameters for Negative Binomial distribution

n = mx.nd.array([10])
p = mx.nd.array([0.5])

Create Negative Binomial distribution

neg_binom_dist = distribution.NegativeBinomial(mu=n, alpha=p)

Generate samples

samples = neg_binom_dist.sample(num_samples=10000).asnumpy()

Compute the inverse CDF (quantile function)

quantiles = neg_binom_dist.icdf(mx.nd.array([0.1, 0.5, 0.9])).asnumpy()

Display the quantiles

display(quantiles)

Best,
J

[jeycervantes]

To clarify, am using the latest version of torch/gluonts

[lostella]

@jeycervantes you appear to be importing from the wrong path: use gluonts.torch instead of gluonts.mx, and contruct the distribution by passing in torch.tensor objects. Here is the negative binomial distribution, with the icdf method: https://github.com/awslabs/gluonts/blob/dev/src/gluonts/torch/distributions/negative_binomial.py#L57