Truncated distribution of Guide (reject negative samples from guide)

1

Is there any method to reject negative samples from guide. For example, here is a custom guide “normal_guide”. If the pyro.sample produces a negative sample, we should reject this negative sample and generate a new sample.

def normal_guide(data):
    # register the two variational parameters with Pyro
    mean = pyro.param("mean", torch.ones(1,Num),
                         constraint=constraints.interval(0, 10))
    cov = pyro.param("covariance", torch.rand(Num, Num),
                            constraint=constraints.positive_definite)    
    # sample latent_variables from the variational distribution    
    pyro.sample("latent_variables", dist.MultivariateNormal(mean, cov))

I tried to realize it by the following code, but got wrong with it: RuntimeError: Multiple sample sites named ‘latent_variables’.

def normal_guide(data):
    # register the two variational parameters with Pyro
    
    mean = pyro.param("mean", torch.ones(1,Num),
                         constraint=constraints.interval(0, 10))
    cov = pyro.param("covariance", torch.rand(Num, Num),
                            constraint=constraints.positive_definite)
    
    # sample latent_variables from the variational distribution    
    sample = pyro.sample('latent_variables', dist.MultivariateNormal(mean, cov))
    
    mask=sample<0
    while (True in mask):
        sample=pyro.sample("latent_variables", dist.MultivariateNormal(mean, cov))
        mask=sample<0
    
    return sample