Sample returns different dimension result after first iteration

I’m creating a simple HMM with gaussian emission. Modified from example from: Example: Hidden Markov Models — Pyro Tutorials 1.8.1 documentation

I got two issues:

  1. sample function always return [0, 1, 2] after first epoch.
  2. parameter p_mu merely change
def model(observations, num_state):

    with poutine.mask(mask = True):

        p_mu = pyro.sample("p_mu",
                            dist.Normal(0, 10000).expand([num_state]).to_event(1))
        

        p_tau = pyro.sample("p_tau",
                            dist.Exponential(1).expand([num_state]).to_event(1))

    
        
        p_transition = pyro.sample("p_transition",
                            dist.Dirichlet((1 / num_state) * torch.ones(num_state, num_state)).to_event(1))

    current_state = 0
    
    counter = 0
    for t in pyro.markov(range(len(observations))):
        counter += 1
        current_state = pyro.sample("x_{}".format(t),
                                    dist.Categorical(Vindex(p_transition)[current_state]),
                                    infer = {"enumerate" : "parallel"})

        print("counter: " + str(counter))
        print(current_state)
        pyro.sample("y_{}".format(t),
                    dist.Normal(Vindex(p_mu)[...,current_state], Vindex(p_tau)[...,current_state]),
                    obs = observations[t])

I use TraceTMC_ELBO
I have 200 observations, the value of current_state during first svi.step() is correct. p_transition is a 2d matrix
After the second time of svi.step(), the value of current_state is always [0, 1, 2].

counter: 198
tensor(0)
counter: 199
tensor(1)
counter: 200
tensor(0)
counter: 1
tensor([[0],
        [1],
        [2]])
counter: 2
tensor([[[0]],

        [[1]],

        [[2]]])
counter: 3
tensor([[0],
        [1],
        [2]])
counter: 4
tensor([[[0]],

        [[1]],

        [[2]]])

The first issue is due to enumeration. Actually the [[0],[1],[2]] shapes are returned every iteration. What you’re seeing first is a trial run of the guide to determine shapes. For explanation see the Tensor shapes tutorial.

So is it proper for my model? Since I expect to only sample one number at a time.

Yes, these tensor shapes are expected and proper. Each of Pyro’s inference algorithms evaluates your model in a different nonstandard interpretation. Trace_ELBO is straightforward and simply evaluates with differentiable PyTorch tensors; TraceEnum_ELBO evaluates with specially shaped tensors that enumerate discrete values over all possible values, then later marginalize out those values; TraceTMC_ELBO evaluates with batches of samples at each site. These nonstandard evaluations enable inference.

Thanks for replying. I changed the ELBO for my HMM into Trace_ELBO, and something very wired happened.

Here’s my current model.

def model(observations, num_state):
    
    assert not torch._C._get_tracing_state()

    with poutine.mask(mask = True):
  
        p_transition = pyro.sample("p_transition",
                                   dist.Dirichlet((1 / num_state) * torch.ones(num_state, num_state)).to_event(1))
        
    p_mu = pyro.param(name = "p_mu",
                      init_tensor = torch.randn(num_state),
                      constraint = constraints.real)

    p_tau = pyro.param(name = "p_tau",
                       init_tensor = torch.ones(num_state),
                       constraint = constraints.positive)


    current_state = 0
    
    with pyro.plate('obs_plate', len(obs)):

      for t in pyro.markov(range(len(observations))):

          current_state = pyro.sample("x_{}".format(t),
                                      dist.Categorical(Vindex(p_transition)[current_state, :]),
                                      infer = {"enumerate" : "parallel"})
          

          pyro.sample("y_{}".format(t),
                      dist.Normal(Vindex(p_mu)[current_state], Vindex(p_tau)[current_state]),
                      obs = observations[t])

I fit it by

device = torch.device("cuda:0")
obs = torch.tensor(obs)
obs = obs.to(device)

torch.set_default_tensor_type("torch.cuda.FloatTensor")

guide = AutoDelta(poutine.block(model, expose_fn = lambda msg : msg["name"].startswith("p_")))

Elbo = Trace_ELBO
elbo = Elbo(max_plate_nesting = 1)


optim = Adam({"lr": 0.001})
svi = SVI(model, guide, optim, elbo)

while 1:


    loss = svi.step(observations = obs, num_state = 3)

The true value of my parameter mu are [-10, 1, 10]
True value of sigma are [1,2,1]

But as the model trains, the ELBO is

mu and sigma looks like:

Do you have any thoughts about how I could fix the problem?