Skip to content

Latest commit

 

History

History
431 lines (263 loc) · 12 KB

CHANGELOG.md

File metadata and controls

431 lines (263 loc) · 12 KB
Raw

Change log of GraphDot

0.8.1 (2021-12-08)

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

0.8 (2021-12-07)

This version formalizes the inclusion of new features introduced from 0.8a1 to 0.8a18. An (incomplete) list of features include:

  • Dataset downloaders (graphdot.dataset)
  • Graph Hausdorff distance metric (graphdot.metric.maximin)
  • Gaussian field regressor (graphdot.model.gfr)
  • Kernel-induced distance metrics (graphdot.metric)
  • Low-rank GPR via Nystrom approximation (graphdot.model.gpr.nystrom)
  • Multiplicative regularization for GPR

0.8a18 (2021-09-30)

  • Maintenance update.

0.8a17 (2021-03-12)

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

0.8a15 (2021-03-12)

  • Downloader for QM9.

0.8a14 (2021-03-09)

  • Minor tweaks to look-ahead rewriter logic.

0.8a13 (2021-03-08)

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

0.8a12 (2021-03-07)

  • 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 (2021-02-25)

  • Improvements to the active learning hierarchical drafter.

0.8a10 (2021-01-14)

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

0.8a9 (2021-01-12)

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

0.8a8 (2021-01-05)

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

0.8a7 (2021-01-02)

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

0.8a6 (2020-12-29)

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

0.8a5 (2020-12-21)

  • Added a downloader for the QM7 dataset.
  • Prototype implementation of a Gaussian field harmonic function regressor.

0.8a4 (2020-12-15)

  • Added an multiplicative regularization option to GPR, which may perform better when the kernel is not normalized.
  • Fixed a linear algebra type error when the GPR kernel matrix is solved with pseudoinverse.

0.8a3 (2020-11-23)

  • Added an experimental Monte Carlo tree search model.

0.8a2 (2020-11-16)

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

0.8a1 (2020-10-15)

  • Enabled GPR training with missing target values.

0.7 (2020-09-21)

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 (2020-09-16)

  • 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 (2020-09-12)

  • Performance improvements to the variance minimizing active learner.

0.7a13 (2020-09-10)

  • Furture improvements to the robustness of the GPR training process.

0.7a12 (2020-09-02)

  • Uses a more robust pseudoinverse algorithm for GPR when the kernel matrix is nearly singular.

0.7a11 (2020-09-02)

  • Added bell-shaped compact adjacency functions.
  • Redesigned the active learning module.

0.7a10 (2020-08-30)

  • Enhanced the readability of the string representations of kernel hyperparameters.
  • New atomic adjacency options.

0.7a9 (2020-08-28)

  • Improved numerical stability tolerance of the GPR and Low-rank GPR models.

0.7a8 (2020-08-27)

  • Added a to_networkx() method to graphdot.Graph.

0.7a7 (2020-08-25)

  • Graph representation string can now be directly deserialized using eval.

0.7a6 (2020-08-23)

  • Optimized GPU gradient evaluation performance
  • predict_loocv now available for the LowRankApproximateGPR model.
  • Unified the fit and fit_loocv method of GaussianProcessRegressor.

0.7a4 (2020-08-18)

  • Fixed a bug related to bounds of kernels contains fixed hyperparameters.

0.7a3 (2020-08-18)

  • Allow kernel/microkernel hyperparameters to be declared as 'fixed' via the *_bounds arguments.

0.7a2 (2020-08-14)

  • Fixed a memory layout issue that slowed down computations using normalized kernels.

0.7a1 (2020-08-12)

0.7a (2020-08-10)

  • Improved the performance of gradient evaluation for the marginalized graph kernel.
  • Introduced a new MaxiMin distance metric between graphs.

0.6.6 (2020-08-10)

  • Added save and load methods to the Gaussian process regressor models.

0.6.5 (2020-08-05)

  • Fixed a bug related to the lmin=1 option of the marginalized graph kernel.

0.6.4 (2020-08-03)

  • Fixed a bug regarding target value normalization in the fit_loocv method of GPR.

0.6.3 (2020-07-30)

  • Fixed a performance degradation due to the inconsistent lexical sorting behavior between numpy.lexsort and numpy.unique.

0.6.2 (2020-07-30)

  • Fixed a bug in computing the gradient of diagonal kernel entries.

0.6.1 (2020-07-29)

  • Fixed a bug in kernel normalization.

0.6 (2020-07-26)

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.

0.6a10 (2020-07-21)

  • Supports graphs with self-looping edges.
  • Made the Graph.from_rdkit method optional in case if RDKit itself is not available.

0.6a9 (2020-07-17)

  • Ensures that graph cookies are not pickled.

0.6a7, 0.6a8 (2020-07-16)

  • Fixed a problem assocaited with converting permuted graphs to octilegraphs.

0.6a6 (2020-07-16)

  • Fixed a problem with caching graphs on the GPU.

0.6a5 (2020-07-15)

  • Introduced a graph reordering mechanism to improve computational performance on GPUs.
  • The default starting probability of the marginalized graph kernel is now hyperparameterized and will be optimized by default during training.
  • Allow users to specify custom starting probability distributions.
  • Performance improvements due to the in situ computation of starting probabilities instead of loading from memory.
  • Added repeat, theta_jitter and tol options to the Gaussian process regressor.
  • Fixed a normalization bug in GaussianProcessRegressor.fit_loocv.

0.5.1 (2020-06-30)

  • Added a verbose training progress option to the GPR module.
  • The graphdot.kernel.basekernel package has been redesigned and renamed to graphdot.microkernel.

0.5 (2020-06-30)

  • Initial formal release of the Gaussian Process regresion module.

0.5a7 (2020-06-28)

  • Implemented the base kernel exponentiation, i.e. k**a, semantics.
  • Minor docstring fixes.

0.5a6 (2020-06-26)

  • Fixed a regression that causes data frame unpickling errors.

0.5a5 (2020-06-24)

  • Added the leave-one-out cross-validation prediction and training to GPR.

0.5a4 (2020-06-22)

  • Fixed an automatic documentation issue.

0.5a3 (2020-06-20)

  • Added check for the shape of hyperparameter bounds specification to prevent users from unknowingly provide invalid values.

0.5a2 (2020-06-20)

  • Fixed a bug related to Jacobian dimensionality.

0.5a1 (2020-06-17)

  • Added a built-in Gaussian process regression (GPR) module.
  • Fixed an issue that prevented the pickling of graphs.

0.4.6 (2020-06-05)

  • Fixed a minor bug in Graph.from_rdkit.

0.4.5 (2020-05-26)

  • Replaced from_smiles with a more robust from_rdkit function with additional ring stereochemistry features. Thanks to Yan Xiang for the contribution.
  • Added a new Compose method for creating base kernels beyond tensor product base kernels.
  • Fixed a performance degradation issue (#57).

0.4.4 (2020-05-23)

  • Ensure that graphs can be pickled.

0.4.3 (2020-05-23)

  • Ensure graph feature data layout consistency involving a mixture of scalar and variable-length features. Fixes #56.

0.4.2

  • Fixed an integer sign issue introduced with graph type unification.

0.4.1

  • Renamed Graph.normalize_types to Graph.unify_datatype.

0.4.0

  • Now allowing variable-length node and edge features thanks to a redesign of the Python/C++ data interoperation mechanism.
  • Introduced a Convolution base kernel for composing kernels on variable-length attributes using scalar base kernels.

0.3.5

  • Added a dtype option to the MarginalizedGraphKernel to specify the type of returned matrix elements.

0.3.4

  • Specified the minimum version of sympy in installation requirements.

0.3.3

  • Allow M3 metric to use partial charge information.
  • Made the element, bond, and charge parameters adjustable in the M3 metric.

0.3.2

  • Miscellaneous bug fixes.

0.3.1

  • Analytic computation of graph kernel derivatives against hyperparameters.

0.3.0

  • Users can now define new base kernels easily using SymPy expression #45.
  • Better scikit-learn interoperability.

0.2.9 (2019-12-14)

  • Fixed a bug related to atomic adjacency #43.

0.2.8 (2019-11-22)

  • Added an experimental 'M3' distance metric

0.2.7 (2019-11-18)

  • Bug fixes and stability improvements

0.2.6 (2019-10-31)

  • Improved the performance of hyperparameter optimization by enabling lightweight re-parameterization.
  • Implemented a few properties and methods for scikit-learn interoperability.

0.2.5 (2019-10-30)

  • Improved the performance of successive graph kernel evaluations

0.2.4 (2019-10-29)

  • Improved the performance of graph format conversion for the GPU kernel by 3 times.

0.2.3 (2019-10-24)

0.2.1 (2019-10-02)

  • Reduced kernel launch preparation time by 50% to address #28.
  • Fixed a memory leak issue #31.

0.2.0 (2019-09-26)

  • Changed return type of the diag() method of MarginalizedGraphKernel to fix #30.

0.1.9 (2019-09-25)

  • Fixed an edge label consistency issue with graphs generated from SMILES strings.

0.1.8 (2019-09-15)

  • Added a freshly-designed atomic adjacency rule.
  • Significantly accelerated conversion from ASE molecules to graphs.
  • Documentation update.

0.1.7 (2019-08-23)

  • Documentation update.

0.1.6 (2019-08-12)

  • Added the diag() method to Tang2019MolecularKernel.

0.1.5 (2019-08-09)

  • Fixed a regression in the CUDA kernel that caused an order-of-magnitude slowdown
  • Switched to single-precision floating points for edge length in Graph.from_ase
  • Added several performance benchmark code to example/perfbench