CODEWITHSUNDEEP
pytorchpyro
Lay González
Lay González

Reputation: 2985

Using Categorical samples in Pyro as integers

I am defining several categorical samples as follows:

probs = torch.ones(K) / K
object_class = []
for a in pyro.plate('objects', A):
    object_class = pyro.sample('object_classes_{}'.format(a),
                              dist.Categorical(probs))
    object_classes.append(object_class)

Then, what I want to do but just can't figure out is to use those samples as if they were integers. When running inference, each object_class is an integer, but after some iterations, instead of being an integer, it is a tensor instead with all the possible values for the categorical variable and therefore I can't no longer use it as an integer. I tried the following, but it doesn't seem to be working:

torch.tensor([Vindex(cmm)[oject_classes[a]] for a in range(A)])

But I get an error saying that "only one element tensors can be converted to Python scalars".

Is it really possible to use categorical samples as integers or do I need to work with an additional dimension throughout the model?

Upvotes: 0

Views: 20

Answers (0)

Related Questions