On the Zero-inflated Negative Binomial distribution, the ZINBDecoder() class computes p = theta / (theta + mu) where theta is the dispersion and mu is the mean (or location) parameters of the Negative Binomial distribution. This is wrong, as the true value for p = mu / (theta + mu). This isn't a major issue (during inference is the same), but for consistency it should be flipped and tested again in the demo.
On the Zero-inflated Negative Binomial distribution, the ZINBDecoder() class computes p = theta / (theta + mu) where theta is the dispersion and mu is the mean (or location) parameters of the Negative Binomial distribution. This is wrong, as the true value for p = mu / (theta + mu). This isn't a major issue (during inference is the same), but for consistency it should be flipped and tested again in the demo.