Nested plates of different sizes

I have essentially the same question as was asked here. However, that thread didn’t end up answering the question. I am implementing a variant of a mixture model, where each latent discrete variable has multiple observations. This leads me to use nested plates, like so:

with pyro.plate('groups', N_groups):
        z = pyro.sample('z', dist.Categorical(theta),)
        with pyro.plate('samples', N_samples_per_group):
            x = pyro.sample('x', dist.MultivariateNormal(mus[z], 

However, in actuality I have differing numbers of continuous observations (x) for each discrete variable (z). I could perhaps use a for-loop, but this will be very slow. Are there any alternatives?

Hi @bantin,
one trick I often use when batching over time series of differing lengths is to pack all my time series into a single 2D tensor and use poutine.mask to specify the raggedness over one of the dimensions; here’s an example. Take care to use numerically safe values for padding though, e.g. pad extra covariances with torch.eye() and extra means with a finite number; if you accidentally pad with garbage PyTorch might error when looking at the padding, since it doesn’t know you’ll be throwing away those results.