How to combine pymc with neural network

Questions v5
1

I’m trying to use neural network as my model for finding input x by means of plain bayes for observation y,like this:


with pm.Model() as multi_objective_model:
    para1=pm.Uniform('para1',lower=[-10,-10,-10],upper=[10,10,10])
    para2=pm.Uniform('para2',lower=[-10,-10,-10],upper=[10,10,10])
    para3=pm.Uniform('para3',lower=[-10,-10,-10],upper=[10,10,10])
    sigmas=pm.HalfCauchy('sigmas',[1,1,1])
    y_observed=pm.Normal(
        "y_observed",
        mu=neural network(X,para1,para2,para3),
        sigma=sigmas,
        observed=Y,
    )
    
    prior = pm.sample_prior_predictive()
    posterior=pm.sample(draws=4000,tune=2000,target_accept=0.9,chains=4)

I don’t know how to combine neural network built by pytorch or tensorflow with pymc, after all, both tensorflow/pytorch and pytensor have strict shape requirements. Is there any similar example or interface? Is there any similar example or interface?

2

You cannot directly use external packages like pytorch or tf in a pymc model, any more than you could define a pytorch model using tensorflow layers.

Your options are:

  1. Wrap the NN in a Pytensor Op. This blog post shows how to do it with Jax/Flax.
  2. Write the NN yourself directly in pytensor. This example might be helpful.

If the architecture is simple, (2) is probably a much easier option.

1 Like
3

Thank you very much, I might try the second method, the neural network I am using is a specially designed fully connected neural network that is difficult to implement with pytensor. Any more examples of similar applications?