Do tricks with distribution vectors

Is it possible to estimate values from following model?

def generator(x):

    # values to estimate
    to_be_need =  3
    to_remember = 4

    preds = torch.zeros(len(x)).long()

    for d in range(len(x)):

        # priors
        leads = dist.Uniform(100, 200).sample()

        # likelihood
        needs = dist.Poisson(to_be_need*torch.ones(int(leads))).sample()
        memos = dist.Poisson(to_remember*torch.ones(int(leads))).sample()

        deals = needs[needs >= memos]
        deals = deals[deals < len(x) - d].unique(return_counts=True)

        preds[deals[0].long() + d] += deals[1]

    return preds

x = torch.arange(0, 100, 1)
y = generator(x)

I tried to build naive model, but render shows some problems with dependencies:

def model(x, y):

    to_be_need = pyro.sample("to_be_need", dist.Uniform(1,7))
    to_remember = pyro.sample("to_remember", dist.Uniform(1,7))

    sigma = pyro.sample("sigma", dist.Uniform(0, 1))

    preds = torch.zeros(len(x)).long()

    for d in pyro.plate("days", len(x)):

        leads = pyro.sample(f"leads_{d}", dist.Uniform(100, 200))

        needs = pyro.sample(f"needs_{d}", dist.Poisson(to_be_need*torch.ones(int(leads))))
        memos = pyro.sample(f"memos_{d}", dist.Poisson(to_remember*torch.ones(int(leads))))

        deals = Vindex(needs)[needs >= memos]
        deals = deals[deals < len(y) - d].unique(return_counts=True)
        preds[deals[0].long() + d] += deals[1]

        pyro.sample(f"obs_{d}", dist.Normal(preds[d], sigma), obs=y[d])
pyro.render_model(model, model_args=(x, y))

And inference cause an error:

mcmc = MCMC(NUTS(model), num_samples=1000, warmup_steps=200), y)
ValueError: Continuous inference cannot handle discrete sample site 'needs_0'. Consider enumerating that variable as documented in . If you are already enumerating, take care to hide this site when constructing an autoguide, e.g. guide = AutoNormal(poutine.block(model, hide=['needs_0'])).

Hint from the error advice to use enumerating, but I have no idea how to apply that

Hi @Svetozar, Pyro doesn’t have a great way to infer Poisson latent variables. I’d recommend thinking about continuous relaxation of your model. Re: pyro.render_model() it looks like our rendinging is being thwarted by then in-place modification of preds. Feel free to file a bug report with your model() code. Again I’d recommend trying to continuously relax your model and avoid in-place modifications if possible.

Hi @fritzo, I have a question related to your last comment. Could you please explain why and when in-place modifications should be avoided in Pyro?

  1. In-place modifications can break automatic differentiation in some circumstances. This is a general PyTorch issue. This is such a difficult problem that JAX made the design choice to completely disallow in-place modifications anywhere. PyTorch is more permissive, but the rules are complex and you’ll need to see what works. A simple necessary rule is: no element of a tensor can depend on another element of the same tensor, in the sense of AD graph dependency.
  2. In-place modifications break Pyro’s provenance tracking that is used in pyro.render_model(). This affects rendering, but also the occasional inference algorithm that uses provenance tracking, such as the AutoGaussian guide and the 2nd generation of TraceGraph_ELBO.

Thank you for answer, I’d decided better to change the model and avoid this challenge
But I still believe that I can feel how to use pyro and apply that in practice