Multiple guides for same model

1

I’m trying to run multiple guides with different initialization (random) using a for loop. But it seems the initial value is same each time guide is called. How to I rectify this?

def guide():
       var_mu_1 = pyro.param('var_mu_1', torch.randn(1))
       var_mu_sig_1 = pyro.param('var_mu_sig_1', torch.tensor(1.),
                             constraint=constraints.positive)
       pyro.sample('mu_1', dist.Normal(loc=var_mu_1, scale=var_mu_sig_1))

I want torch.randn to have a different value each time I call guide in SVI

2

are you clearing the param store between using the different guides?

3

What’s your use case for running multiple guides? This will not work because param values have a global scope. The first time pyro.param('var_mu_1', ...) is called, it registers an initial value in the global parameter store keyed by the name ‘var_mu_1’. The next time this statement is called, it will simply return the value from the the param store (without re-initializing). As @jpchen suggested, you could just call pyro.clear_param_store() after training each guide and retrieving the results before starting to train another one in a loop.

4

Thanks guys. I’m working on Variational Inference with different initialization.