Concatenate samples

Misc.
1

Hi everyone,

I’m new to pyro and I am trying to understand how I can use the previous output of a distribution to my next node.

Essentially, I want to ‘stack’ outputs such as:

vs = []

for i in range (n_nodes):
    v = pyro.sample(f"v{i}", dist.Bernoulli(torch.tensor([0.3])
    vs.append(v)

This works but the issue is that ‘vs’ is a list of pyro objects and I cannot apply pyro functions afterwards. Therefore, I struggle to understand how I could build such a concatenation of outputs and use the values without breaking my graph dependencies.

For example doing something like this:

vs = torch.zeros((n_nodes)) 

for i in range (n_nodes):
    v = pyro.sample(f"v{i}", dist.Bernoulli(torch.tensor([0.3])
    vs[i] = v

breaks the flow of my graph as ‘vs’ is a torch.Tensor and not a pyro object anymore.

Any help welcomed!

2

there is no such thing as a “pyro object”: the return values that come from pyro.sample are torch.Tensors and so as long as you don’t do something to break things (tensor.detach(), tensor.numpy(), etc.) you can freely use operations like torch.stack to combine those tensors

3

Thanks, yes it does work!