One distribution over weights

[cerebrou]

When building a network (and possibly calling random_module) is it possible to specify a distribution over a few weights instead of one? Meaning that in the forward step, weights would be sampled from one distribution and we update the parameters of that distribution accordingly in the backward step according to the loss obtained from the samples from that one distribution?

[jpchen]

if you do it directly (without random_module) then you can do whatever you want - just sample each layer from the same distribution. if you use random_module, if you want to sample every layer from the same dist my_dist, you can just pass that dist into random_module:

my_dist = dist.Normal(...)
lifted_module = pyro.random_module('random nn', nn_module, my_dist)

if you only want certain layers lifted, this should work:

priors = {'layer_1': dist_1, 'layer_2': dist_1, 'layer_3': dist_2}
lifted_module = pyro.random_module('random nn', nn_module, priors)

[cerebrou]

Ok, that is fair enough. Would you be able though to help me with a minimum working example without using the random_module?

[jpchen]

sure there's an example tutorial as well.