Adding weights to a Bernoulli distribution

I would like to create a weighted Bernoulli function in pyro (with known weights). As some of my observations are more trustworthy than others. The log-likelihood for LR: image

My question is how do I alter the likelihood of such Bernoulli function:

with pyro.plate('data'):
    pyro.sample('y', dist.Bernoulli(logits=p), obs=y)

Could I simply add to the power of the weights as shown below:

  with pyro.plate('data'):
        pyro.sample('y', dist.Bernoulli(logits=torch.sign( p ) * torch.pow(torch.abs( p ), w)), obs=y)

(with w being the weights)

My example runs, it only does not give me the desired results… Could someone please explain to me why this does not work or how I should implement it instead?

Hi @Helena.H,

I think you want Bernoulli(logits=w * p). Note that p is kind of a confusing name for the logits arg since p so often denotes a probability as in Bernoulli(probs=p), but logits is a log likelihood ratio:

probs = exp(logits) / (1 + exp(logits))
      = 1 / (1 + exp(-logits))
logits = log(probs / (1 - probs))

Good luck!

1 Like

You can also use

with pyro.plate('data'), pyro.poutine.scale(scale=w):
    pyro.sample('y', dist.Bernoulli(logits=p), obs=y)
2 Likes