numgp

A port of the gp module of PYMC3 for NumPyro.

Free software: MIT license
Documentation: https://numgp.readthedocs.io.

Features

TODO

Credits

numgp's People

Stargazers

Watchers

numgp's Issues

Ice core example

Hi there,

Thanks for the great library, it was exactly what I was looking for! I am attempting to port the Mauna Lao CO2 model (example notebook here) but was running into difficulties when using the predict method, the model is basically.

def model(X, y, Xnew):

    T = X.shape[0]

    η = numpyro.sample('η', dist.HalfNormal(5))
    ℓ = numpyro.sample('ℓ', dist.Gamma(4, 2))
    α = numpyro.sample('α', dist.Gamma(3, 1))  
    cov = η**2 * npg.cov.RatQuad(1, α, ℓ)

    gp = npg.gp.Marginal('f', cov_func=cov)

    t_diff = numpyro.sample('t_diff', dist.Normal(0., 0.02), sample_shape=(T,))
    t_uncert = X - t_diff

    σ = numpyro.sample('σ', dist.HalfNormal(5))
    
    y_ = gp.marginal_likelihood(X=t_uncert[:, None], y=y, noise=σ)
    
    _ = gp.conditional(t_uncert[:, None], Xnew)

Train methods:

svi = SVIHandler(model, AutoDiagonalNormal(model), num_epochs=10000)
svi.fit(t_n, y_train, None, True, return_sites=('ℓ', 'η', 'α', 'σ', 'f'))

Predict methods:

Xnew = np.linspace(-100, 2150, 2000) * 0.01

marg = MarginalConditional(model, 'f')
mu, var = marg.conditional_from_guide(
    svi.guide, 
    svi.params, 
    X,
    y_train, 
    Xnew[:, None]
)

However, when it comes to make predictions based on the new data, it only returns nan is there something I'm missing?

 with numpyro.handlers.seed(rng_seed=34):
    pred = numpyro.sample('pred', dist.MultivariateNormal(loc=mu, covariance_matrix=var).expand([200]))

pred
>>> Array([[nan, nan, nan, ..., nan, nan, nan],
           [nan, nan, nan, ..., nan, nan, nan],
           [nan, nan, nan, ..., nan, nan, nan],
           ...,
           [nan, nan, nan, ..., nan, nan, nan],
           [nan, nan, nan, ..., nan, nan, nan],
           [nan, nan, nan, ..., nan, nan, nan]], dtype=float32)

Recommend Projects