Gcnn Keras Versions Save

• Removed deprecated modules in kgcnn.utils.
• Moved kgcnn.utils.model to kgcnn.model.utils.
• Fixed behaviour for kgcnn.data.base.MemoryGraphDataset.get_train_test_indices to return list of train-test index tuples.
• Updated kgcnn.hyper.hyper.HyperParameter to deserialize metrics and loss for multi-output models.
• Added trajectory_name in summary.py and history.py.
• Fixed kgcnn.layers.geom.PositionEncodingBasisLayer
• Removed deprecated kgcnn.layers.conv.attention and kgcnn.layers.conv.message
• Updated setup.py for requirements.
• HOTFIX: Error in kgcnn.scaler.mol.ExtensiveMolecuarScaler, where scale was not properly applied. However, this was only present in development (master) version, not in release.
• Added kgcnn.layers.relational with relation dense layer.
• Added first draft of kgcnn.literature.HDNNP2nd.
• Added append, update and add to MemoryGraphList.
• Fixed behaviour of GraphDict , which now is not making a copy of arrays.
• Added explainable GNN from visual_graph_datasets.
• Ran training for train_force.py
• Changed backend to RDkit for QMDatasets.
• Added kgcnn.md and kgcnn.xai .
• Added further documentation.

• Removed kgcnn.graph.adapter and switched to completed kgcnn.graph.preprocessor. The interface to MemoryGraphList and datasets does not change. How to update:

from kgcnn.data.base import GraphDict
GraphDict().apply_preprocessor("sort_edge_indices")  # Instead of GraphDict().sort_edge_indices()
GraphDict().apply_preprocessor("set_edge_weights_uniform", value=0.0) # Instead of GraphDict().set_edge_weights_uniform(value=0.0)
# Or directly using class.
from kgcnn.graph.preprocessor import SortEdgeIndices
SortEdgeIndices(in_place=True)(GraphDict())

• Add kgcnn.literature.MEGAN model.
• Add kgcnn.literature.MXMNet model.
• Fixed error in ClinToxDataset label index.
• Reworked kgcnn.graph.adj.get_angle_index with additional function arguments. Default behaviour remains identical. For periodic system an additional allow_reverse_edges=True is now required.
• Added input embedding for edges in kgcnn.literature.EGNN. Debugged model.
• Reworked kgcnn.literature.PAiNN to simplify normalization option and add equivariant initialization method.
• Refactored kgcnn.data.qm including all qm7-qm9 datasets. Improved documentation. If error occurs, please run QM9Dataset(reload=True).
• Refactored kgcnn.data.moleculenet. Interface and behaviour does not change.
• Renamed kgcnn.mol.graph_babel and kgcnn.mol.graph_rdkit and move conversion into kgcnn.mol.convert.
• Added kgcnn.data.datasets.MD17Dataset and kgcnn.data.datasets.MD17RevisedDataset
• Refactored kgcnn.scaler module to follow sklearn definitions. Changed input naming and order for scaler. Add config and weights functionality.
• Changed kgcnn.training.scheduler.LinearWarmupExponentialLearningRateScheduler to take correctly lifetime parameter.
• Reworked kgcnn.data.datasets.QM9Dataset to offer atomization energies and uncharacterized molecules. Please run reload=True.
• Improved docs for kgcnn.training.

• Remove reserved properties form MemoryGraphList, please use set/get methods.
• Removed deprecated kgcnn.selection module.
• Added history score summary in kgcnn.training.history.
• Rework training. Having plots takes up more memory. Prefer summary table of benchmarks.
• Changed kgcnn.data.datasets.PROTEINSDataset to binary graph labels.
• Add kgcnn.literature.CMPNN model.
• Add kgcnn.literature.EGNN model.
• Merge set_attributes into read_in_memory for MoleculeNetDataset and make set_attributes alias of read_in_memory.
• Fix error of node updates in kgcnn.literature.GAT. Rerunning training.
• Fix learning rate scheduler kgcnn.training.scheduler.LinearLearningRateScheduler min. learning rate if trained beyond epoch argument.
• Removed kgcnn.layers.casting.ChangeIndexing at it was not used.
• Added kgcnn.layers.casting.CastEdgeIndicesToDenseAdjacency.
• Merged kgcnn.layers.mlp.MLP and kgcnn.layers.mlp.GraphMLP, but kept GraphMLP as alias. Change in kwargs for "normalization_technique".
• Moved kgcnn.layers.conv.message to kgcnn.layers.message.
• Refactored kgcnn.layers.conv.attention into kgcnn.layers.conv.gat_conv and kgcnn.layers.conv.attentivefp_conv.
• In MoleculeNetDataset and QMDataset changed the shape of 'edge_number' to be (N, ) instead of (N, 1). To agree with 'node_number' shape.
• Removed kgcnn.layers.conv.sparse as it was not used and added its content to kgcnn.layers.conv.gcn_conv and kgcnn.layers.casting
• Started with kgcnn.graph.preprocessor.

• Add kgcnn.crystal module, which is still in development.
• Add get_weights and get_config to kgcnn.scaler
• Add get and set alias to GraphDict and MemoryGraphList. Which now can be used to assign and obtain graph properties.
• Refactored GraphDict and adj into kgcnn.graph.
• Add a set_range_periodic function to GraphDict.
• Add make_crystal_model functions to SchNet, Megnet, DimeNetPP.
• Add custom_transform to MoleculeNetDataset.
• Removed add_hydrogen, make_conformer, and optimize_conformer from constructor of MolGraphInterface.
• Added add_hs, make_conformer and optimize_conformer to MolGraphInterface.
• Add normalization option to PAiNN and add make_crystal_model.
• Add kgcnn.literature.CGCNN model with docs.
• Add more list-functionality to MemoryGraphList.
• Add tutorial_model_loading_options.ipynb to notebooks showing different ways to load ragged data.
• Add tutorial_hyper_optuna.ipynb to notebooks

• fix typo to read kgcnn.mol.encoder
• fix bug in GraphDict.from_networkx() for edge attributes.
• Improved docs overall.
• Added ragged node/edge embedding output for TF > 2.8 via "output_to_tensor" model config.
• Added make_function option to training scripts.
• Refactored GraphDict methods into kgcnn.data.adapter.GraphMethodsAdapter.
• Removed kgcnn.layers.modules.ReduceSum as it has not been used and may be problematic.
• Moved kgcnn.utils.data to kgcnn.data.utils.
• Refactored smile to mol generation into kgcnn.mol.convert and renamed kgcnn.mol.gen to kgcnn.mol.external
• fixed bug for GatherEmbedding to have correct concat axis if index tensor happens to be of rank>3 but ragged_rank=1.
• Refactored kgcnn.mol methods into modules and renamed graphRD and graphBabel.
• Continued to work on kgcnn.data.crystal.CrystalDataset.
• Added MatBenchDataset2020 dataset to kgcnn.data.datasets.

Changes, updates and bug fixes

• Continue with dataset serialization. Changed training hyperparameter for dataset serialization and ran training examples.
• Refactored scaler and training utilities with serialization but kept old references.
• Added first draft of HamNet.
• Removed Haste layers, since integration did not work smoothly.
• Simplified MLPBase code.
• Changed MemoryGraphList.clean() to return list of removed indices.
• Replaced explicit keras (python) import with ks = tf.keras

Changes, updates and bug fixes

• Fixed error for MoleculeNetDataset for invalid smiles which stopped at graph size in #44
• Changed name GraphNumpyContainer in GraphDict but kept as alias.
• Added to_networkx and from_networkx to GraphDict
• Changed the name of functional arguments for methods in GraphDict, i.e. removed the "name_" prefix. This does change previous usage if no default arguments are used!
• Added notebook to show usage of GraphDict.
• Fix bug in set_edge_indices_reverse for GraphDict.
• Refactor internal structure of MoleculeNetDataset to support custom_callbacks parameter for attributes. The interface did not change else. Accepted #45
• Added list to initialization of MemoryGraphList.
• Updated "README.md".
• Started with serializaton of datasets.

Realse version 2.0

New version 2 of kgcnn including refactored layers, datasets and training scripts. Changes are not fully compatible with version <2.0.0 of kgcnn. Future updates of version 2.0 are not supposed to introduce breaking changes to the API but add additional functionality and bug fixes. More extensive documentation and tutorials are planned. Some notable changes of version 2.0:

Changes and new features

• More general training scripts for each overall dataset type, like QM, TUDataset. Changes in the hyperparameter config to incorporate future extensions of eg. cross-validation or scaling and encoding of targets with keras-like serialization.
• Renamed dataset classes to dynamically load also dataset module in training script.
• Dataset base class is now a list of individual graphs instead of a collection of lists.
• A graph is represented by a python dictionary of named arrays for indices, node attributes etc.
• Every literature model has now a MLP at the output which defines the model output activation and units.
• Added normalization layers for Graphs following keras naming convention.
• The MLP class now also has dropout and normalization options and therefore had to be split into GraphMLP and MLP, since normalization on graphs is not identical to feed forward networks. Without normalization they are interchangeable.
• Renaming of base layers since in tf.version > 2.7 it is not possible to have a module named keras.py and identical layer names. Most standard keras layers now fully support ragged tensors as for example Dense. We therefore renamed keras.py into modules.py and layers into e.g. LazyAdd, LazyConcatenate and DenseEmbedding for using keras-corresponding layers compatible with tf.version < 2.7.
• MoleculeNetDataset now uses both RDkit and OpenBabel to generate molecular structures from smiles more reliable and optionally generate conformers. Conformer generation is parallelized for multiple molecules to greatly speed up SDF file generation. Note that node/edge or atom/bond attributes still require RDkit and molecules that fail RDkit valence test can not be used if specific attributes are desired. We also started to add the option for external conformer generation programs such as balloon in place of RDkit/OpenBabel. OpenBabel is now also a necessary dependency for MoleculeNetDataset.
• Started to add periodic crystal graph plug-in for the standard GNNs in kgcnn but that is not working yet. Will be possible in future releases.
• Warning and info notifications in modules now uses logging package instead of standard print.

Changes, updates and bug fixes

• Refactored training hyper-parameters to be more keras matched.
• Training paramters are now loaded from python or json/yaml and are visible in training. Removed hyper.datasets.
• Fixed multiple bugs in kgcnn.data.moleculenet.
• Added first version of DMPNN to literature.
• Added examples for how to use Moleculenet and training paramters in jupyter notebooks.

Changes, updates and bug fixes

• New training scripts which are model independent and serializable.
• Trainig scripts with command line parameters to load hyper-parameter dictionary from json-file.
• Modified model input parameters for all models.
• Default hyper-parameters per dataset.
• Updated documentation in several modules.
• Reorganized layers in kgcnn.layers.conv and kgcnn.layers.pool.
• Minor bug fixes.