 # Transform value vector to count vector (e.g. BetaBinomial to Multinomial)

Dear awesome forum,

Here is my problem: I am sampling from a BetaBinomial multiple times and would like to transform the sampled values into counts, similar to a single draw from a Multinomial. For instance:

``````total_count = 6
n_samples = 4

values = dist.BetaBinomial(torch.tensor([1.]), torch.tensor([2.]), total_count = total_count).sample(sample_shape = (n_samples,)).view(-1)
``````

then transform:

``````values = [2, 2, 0, 4] ## transform this
counts = [1, 0, 2, 0, 1, 0, 0]
## of size total_count, where the bin at index 0 correspond to the count of value 0, index 1 to the count of value 1 and so on
``````

This should then be used such that

``````W = pyro.sample('W', dist.TransformedDistribution(dist.BetaBinomial(torch.tensor([1.,....]), torch.tensor([2.,....], total_count = torch.tensor([vocab_size-1])), [CountTransform(vocab_size)]), obs=docs)
``````

I would like to have a transform for this and tried the following code, which does however not work since I don’t know how to define `log_abs_det_jacobian`.

``````class CountTransform(Transform):

domain = constraints.real_vector
codomain = constraints.real_vector
bijective = True

def __init__(self, total_count: int = 6, normalized: bool = False):
super().__init__(cache_size=1)
self.total_count = total_count
self.normalized = normalized
#self._cache_size = 0
self._cached_logDetJ = None

def __eq__(self, other):
return type(self) == type(other) and self.p == other.p

def _call(self, x):

def _inverse(self, y):
return reverse_count_vector(y)

def log_abs_det_jacobian(self, x, y, intermediates=None):
return x

def to_batched_count_vector(self, values, vocab_size = 6, normalized = True):

if len(values.shape) == 1:
values = values.view(1,-1)
assert len(values.shape) == 2

base = torch.zeros(values.shape, vocab_size, dtype=torch.int64)

if normalized:
count_tensor = torch.nn.functional.normalize(count_tensor.float(), p=1, dim=1)
return count_tensor

def reverse_count_vector(self, count_vector):

count_vector = count_vector.long()
if len(count_vector.shape) == 1:
count_vector = count_vector.view(1,-1)
assert len(count_vector.shape) == 2

values = list()
for i, doc in enumerate(range(0,count_vector.shape)):
notnull_value = torch.nonzero(count_vector[i,:] != 0).view(-1)
notnull_count = count_vector[i,:][notnull_value]
doc_values = list()
for j, word_index in enumerate(notnull_value):
doc_values += [word_index.item()] * notnull_count[j]
values.append(doc_values)

if len(values) == 1:
values = torch.tensor(values)
return values
``````

`TransformedDistribution` is only for continuous-valued distributions.

1 Like

Many thanks for your quick help @martinjankowiak! I am more than happy to explain my high-level goals:

My data are count vectors: [[1,0,3],[0,0,4],[1,2,1]…]. For instance, the first data point can be translated from the count vector [1,0,3] to the value vector [0,2,2,2], because “0” appears 1 time, “1” appears 0 times and “2” appears 3 times. We can model the value vectors using multiple draws from a BetaBinomial.

However, the ordering of the value vector is not only arbitrary [0,2,2,2] = [2,0,2,2] = … but the vector size may also change depending on how many values are observed. Thus, I would like to transform the value vector to a count vector which is of fixed length (the vocabulary of potential values).

This basically brings us to a Multinomial distribution which is however parametrised differently than the BetaBinomial. The Multinomial requires a Dirichlet, while the BetaBinomial only requires the alpha and beta parameter.

are you sure you don’t want the DirichletMultinomial distribution?

@martinjankowiak yes, I am sure! As described above, my model requires the minimalistic parametrisation of the BetaBinomial.

why can’t you do this once outside of the model using e.g. vanilla numpy or torch just like you do would any other data pre-processing?

1 Like

Now, we are getting to the more pyro-related issue, which may be the right direction to solve this.

The problem is that `obs` can be only included within `pyro.sample`, e.g.

``````pyro.sample('W', dist.BetaBinomial(alpha_c, beta_c),total_count=torch.tensor([self.vocab_len - 1])).to_event(1), obs=docs)
``````

Is there a way to sample the value vectors from the BetaBinomial without `obs`, then do the transform to count vectors separately and finally include the `obs` in something like pyro.deterministic? I still want to be able to use MCMC to infer the posterior distribution.

i’m afraid i don’t follow. forget about pyro for a second. what are the modeling assumptions you want to make?

1 Like