Tags

0.8.1

Hotfix to improve GFR numerical stability when adjacency is nearly zero.

0.8

Many updates since 0.7

0.8a17

- Fixed a QM9 downloader issue.
- Fixed a bug with the Maximin metric when some hyperparameters are fixed.

0.8a15

- Downloader for QM9.

0.8a14

0.8a13

- A new and experimental calling convention to allow evaluations of the
  graph kernels at a list of specific indices.

0.8a12

- A sequence-based rewriter for Monte Carlo tree search.
- Convert any kernel into a metric via `KernelInducedDistance`.
- Convert any norm into a kernel via `KernelOverMetric`.

0.8a11

- Improvements to the active learning hierarchical drafter.

0.8a10

- Performance optimization for GFR leave-one-out cross validation gradients.

0.8a9

- Leave-one-out cross validation for Gaussian field regressor.

0.8a8

- Normalized the MaxiMin graph distance metric to [0, 1].
- More data downloaders: METLIN SMRT, AMES, and a custom downloader.

0.8a7

- Gradient evaluation for the MaxiMin graph metric.
- Gradient evaluation for Gaussian field regressor prediction loss.

0.8a6

- Optimized the evaluation of the gradient of the loss function for the
  Gaussian field regressor.
- Implemented a finite-difference based graph kernel nodal gradient.

0.8a4

Funtionality enhancements and bug fixes for the GPR models.

0.8a3

Added an experimental Monte Carlo tree search model.

0.8a2

Enabled Low-rank GPR (Nystrom) training with missing target values.

0.8a1

allow missing labels in GPR training set.

0.7

This version formalizes the inclusion of new features introduced from 0.7a1
to 0.7b2. An (incomplete) list of features include:
- A redesigned active learning module (`graphdot.model.active_learning`).
- The PBR graph reordering algorithm for graph kernel acceleration
  (`graphdot.graph.reorder.pbr`).
- LOOCV predictions using the low-rank approximate GPR.
- Significant improvement to the robustness of the training methods of GPR
  and Low-rank GPR models.
- Allow kernel/microkernel hyperparameters to be declared as 'fixed' via the
  `*_bounds` arguments.
- Added a `DotProduct` microkernel for vector-valued node and edge features.
- Added a `.normalized` attribute to all elementary and composite microkernels.
- Graph representation string can now be directly deserialized using `eval`.
- New atomic adjacency options such as alternative bell-shaped compact
  adjacency functions (`compactbell[a,b]`), and new length scale choices using covalent radiu etc.
- Perform value range check for the node and edge kernels during graph
  kernel creation.
- Added a `to_networkx()` method to `graphdot.Graph`.
- Enhanced the readability of the string representations of kernel
  hyperparameters using an indented print layout.
- Various performance and bug fixes.

0.7b2

- Added a `DotProduct` microkernel for vector-valued node and edge features.
- Added a `.normalized` attribute to all elementary and composite microkernels.
- Perform value range check for the node and edge kernels during graph kernel
  creation.

0.7b1

Beta release for 0.7 with all new features introduces since 0.7a to 0.7a13.