v0.2.1

Bug fixes, thanks to @timpara :

• Fixed implementation of the RBF kernel gradient
• Fixed an issue where GaussianProcess.predict(w_pred) did not correctly support sequential execution with varying w_pred values

v0.2.0

• Important Notice (interface change: dy -> cov_y): Full covariance matrices are now supported as part of the input data for the model. This slightly changes the interface when using a single number to globally define errors for all data points contained in the array $y$: instead of specifying the standard deviation $\sigma_y$, replace it with the variance $\sigma_y^2$. Similarly, for individual errors associated with different points $y_i$ (but zero covariance between different $y_i$), the information now needs to be provided as a 1-d array containing all $\sigma_{y_i}^2$ (i.e., the diagonal of the covariance matrix) instead of $\sigma_{y_i}$.

• Refactored the functionality for construction of the model covariance matrices for inference, in order to make it easier for users to define custom linear equality constraints and specify the associated rules for how to combine them with the kernel function. This is now possible by defining custom classes that inherit from covariance.OneSided and covariance.TwoSided and implementing functions following the naming convention def _{Constraint}(self, ...) and def _{Constraint1}_{Constraint2}(self, ...), respectively. E.g., for combining a constraint of type constraints.Integral with one of type constraints.Derivative, the function has to be defined as def _Integral_Derivative(self, ...).

• More data types are now supported for the input data provided to the model, namely dict, list and np.ndarray.

v0.1.0

Initial release

fredipy is now available on pip: pip install fredipy