Size of label 'd' for operand 1 (4) does not match previous terms (2)

Misc.
1

Hi~

My code behaves strangely, but i have no idea where the error comes from, please help me.

the code is as follows

if you run the code, it will prompt an error:Size of label ‘d’ for operand 1 (4) does not match previous terms (2), however, if you comment out the ‘prompt’ variable, it will give a Assertion error.
I guess there are conflicts between the two markovs, but why?
Thanks~

# %%
import pyro
import torch
import pyro.distributions as dist
from pyro.infer import SVI, Trace_ELBO, TraceEnum_ELBO, config_enumerate, infer_discrete, JitTraceEnum_ELBO
from pyro.ops.indexing import Vindex

class Agent:
    def __init__(self,
                 id
                 ):
        self.id = id

    @config_enumerate
    def generative_model(self):

        for T in pyro.markov(range(1)):
            cor_level_policy = pyro.sample('{}_cor_level_policy_{}'.format(self.id, T), dist.Categorical(
                probs=torch.ones(4)))
            cp=pyro.deterministic('current_policy',cor_level_policy)

        for round in pyro.markov(range(1)):
            prompt = pyro.sample('{}_prompt_{}'.format(self.id, round), dist.Categorical(
                probs=torch.ones(2)))
            comm_type = pyro.sample('{}_comm_type_{}'.format(self.id, round), dist.Categorical(
                probs=torch.ones(2)))
            comm = pyro.sample('{}_comm_{}'.format(self.id, round), dist.Categorical(
                    probs=Vindex(torch.eye(4))[cp]))

    def guide(self):
        pass

    def comm(self):
        adam = pyro.optim.Adam({'lr': 0.01})
        elbo = pyro.infer.TraceEnum_ELBO(num_particles=1)
        svi = pyro.infer.SVI(self.generative_model, self.guide, adam, elbo)
        losses = []
        for step in range(10):
            loss = svi.step()
            logging.info("Elbo loss: {}".format(loss))
            losses.append(loss)

agents = {}
for id in range(2):
    agents[id] = Agent(id=id)
for id in range(2):
    foci_agent = agents[id]
    foci_agent.comm()
2

I did some research, and found that the problem occurs because i was using an enumerated varible inside a markov context as an index in the downstream code.
But i really need some specific time slice for determine downstream variable,
how can i avoid this problem?
Thanks~

3

Hi @yaow, As with your other models, I see no observation statements. Could you revise this as a more idiomatic Pyro model with both latent variables and observed variables, corresponding to the prior and likelihood, respectively?

4

Thanks!
My original model have real world meanings and observations, but to make the code easy to read i removed all irrelevant codes without interfereing the error, so i guess lacking observation node is not the cause of the problem.
I reexamined my code and found that it is not caused by conflict between two markov. I can reproduce the error but because we have an enumerated sample site which is set inside a markov context, and use it out side this sample site, which somewhat violate rule 2 in the enumeration tutorial, if we see markov context as a special case of parallel plate.
I guess that to use a markov contexted discrete sample site out of that context, you should use guide side sequential enumeration. By doing so i solved this error. Would you please tell me whether i’m right?