Hello

I have my own model+data to infer parameters related with multinomial distribution but that doesn’t work.

So I tried to solve the SVI Part1 problem with multinomial to see if that’s better but it’s like that doesn’t work neither so I’m trying to guess what’s wrong.

```
data=[]
for i in range(2000):
data.append(pyro.sample("obs_{}".format(i),pdist.Multinomial(10,probs=torch.Tensor([0.6,0.4]))))
def model(data):
alpha0 = 10
alpha1 = 10
f = pyro.sample("latent",pdist.Beta(alpha0, alpha1))
ok = torch.Tensor([f,1-f])
for i in range(len(data)):
pyro.sample("obs_{}".format(i), pdist.Multinomial(10,probs=ok),obs=data[i])
def guide(data):
qalpha0 = pyro.param("qalpha0",torch.Tensor([15]),constraint=constraints.positive)
qalpha1 = pyro.param("qalpha1",torch.Tensor([15]),constraint=constraints.positive)
pyro.sample("latent",pdist.Beta(qalpha0, qalpha1))
adam_params = {"lr": 0.001, "betas": (0.9, 0.999)}
optimizer = pyro.optim.Adam(adam_params)
svi = SVI(model, guide, optimizer, loss=Trace_ELBO())
n_steps = 4000
for step in range(n_steps):
svi.step(data)
if step % 100 == 0:
print(step)
print(pyro.param("qalpha0")/(pyro.param("qalpha0")+pyro.param("qalpha1")))
```

Basically I’m doing the same thing as in the tutorial except that the shape of the data and then the distribution used change (multinomial instead of bernoulli). The problem is that infered f given by the last line is roughly equal to 0.5 instead of 0.6. I’ve already though about the size of the data but with len(data)=2000 that infers perfectly 0.6 when using bernoulli distribution like in the SVI, plus the computation time is pretty high.

What’s wrong ?