Skip to content

GPJax v1.1.0

Latest

Choose a tag to compare

@github-actions github-actions released this 10 Nov 13:06
41930a9

Added

  • gpjax.xarray input transforms. from_xarray(..., transforms=[...])
    turns the named inputs into the columns of X, and the GridSpec applies the
    same fitted transforms to every new grid. Standardise scales inputs with the
    training mean and standard deviation. UnitSphere maps latitude and longitude
    to a point on the unit sphere, so a stationary kernel is valid on the whole
    globe. Cyclic encodes a periodic input, such as the seasonal cycle, as a
    point on a circle. GridSpec.columns names the resulting columns.
  • GridSpec.predict: chunked prediction on large grids.
    spec.predict(lambda x: posterior(x, covariance="diagonal"), grid) gives the
    predictive mean and variance as an xr.Dataset. The model gets at most
    chunk_size cells at a time and is compiled once. If the grid holds Dask
    arrays, the result is lazy and is predicted block by block, along the Dask
    chunks of the grid.
  • New example: Infilling Global Surface Temperature. It reads a netCDF
    file of the 2024 temperature anomaly from the NCEP-NCAR Reanalysis 1, which is
    complete, and removes cells to test the infill against the truth. Cells removed
    at random are filled well, and joint samples give a global mean whose interval
    holds the truth. Cells removed because they are warm show how a GP is biased,
    and overconfident, when data are missing not at random.
  • The xarray introduction is now Working with Gridded Data, under Getting
    started.
    It uses Standardise and GridSpec.predict, and shows
    UnitSphere and Cyclic. Its URL is unchanged.

Fixed

  • TruncatedPeriodic and its state-space SDE no longer return zeros or
    NaNs for small Bessel arguments.
    The Miller downward recurrence in
    _stable_scaled_ive formed unnormalised values that grow like
    $(2k_{\max}/c)^{k_{\max}}$, with $c = 1/(4\ell^2)$. These overflowed in
    float64 for a large lengthscale and a high truncation order (for example
    $\ell > 12$ with truncation_order=25), and in float32 already for the
    default kernel (lengthscale=1.0, truncation_order=6), whose Gram matrix
    was all NaN. The recurrence now runs on the ratios $I_k / I_{k-1}$, which stay
    in $(0, 1)$.
  • PoweredExponential.power is now constrained to $(0, 2]$
    (#813).
    The kernel is
    positive definite in every dimension only in this interval, but any value was
    accepted, and power=3.0 gave a Gram matrix with negative eigenvalues. A
    float in $(0, 2)$ is now wrapped as a trainable SigmoidBounded parameter,
    so fit learns the power and cannot move it outside the interval. Before
    this change a float power was a plain Python float, which fit did not
    train. power=2.0 (the RBF case) stays fixed, and a value outside the
    interval raises ValueError. The type of the power leaf changes, so code
    that reads it should use val(kernel.power).
  • Split BlockDiag input vectors with static input block sizes, so matrix-vector
    products work under JIT and with rectangular blocks.
  • Read the observation count from y for unsupervised Dataset objects, so
    n and full_size work when X is absent.
  • Apply the Gaussian covariance transform to each sample for multidimensional
    GaussianDistribution.sample shapes, including shapes with empty axes.

Changed

  • Allow JAX and JAXlib 0.11 in downstream environments by removing the
    <0.11 dependency bounds (#801).

Full diff: v1.0.0...v1.1.0