- Furture improvements to the robustness of the GPR training process.
- Enhanced the readability of the string representations of kernel hyperparameters. - New atomic adjacency options.
- Improved numerical stability tolerance of the GPR and Low-rank GPR models.
- Added a `to_networkx()` method to `graphdot.Graph`.
- Graph representation string can now be directly deserialized using `eval`.
- Optimized GPU gradient evaluation performance - `predict_loocv` now available for the LowRankApproximateGPR model. - Unified the `fit` and `fit_loocv` method of GaussianProcessRegressor.
- Fixed a bug related to bounds of kernels contains fixed hyperparameters.
- Allow kernel/microkernel hyperparameters to be declared as 'fixed' via the `*_bounds` arguments.
Fixed a memory layout issue that slowed down computations using normalized kernels.
- The PBR graph reordering algorithm is available now.
- Improved the performance of gradient evaluation for the marginalized graph kernel. - Introduced a new `MaxiMin` distance metric between graphs.
- Fixed a bug related to the `lmin=1` option of the marginalized graph kernel.
- Fixed a bug regarding target value normalization in the `fit_loocv` method of GPR.
- Fixed a performance degradation due to the inconsistent lexical sorting behavior between `numpy.lexsort` and `numpy.unique`.
- Fixed a bug in computing the gradient of diagonal kernel entries.
- Fixed a bug in kernel normalization.
This version formally releases the new features as have been introduced in the various 0.6alpha versions, such as: - Nystrom low-rank approximate Gaussian process regressor - Graphs with self-looping edges - Graph permutation and reordering operations for GPU performance boost. - Hyperparameterized and optimizable starting probabilities for the graph kernel.
- Supports graphs with self-looping edges. - Made the `Graph.from_rdkit` method optional in case if RDKit itself is not available.
Bug fix: ensures that graph cookies are not pickled.