Gcnn Keras Versions Save

• Removed unused layers and added manual built in scripts and training functions, since with keras==3.0.5 the pytorch trainer tries to rebuild the model, even if the model is already built and does it eagerly without proper tensor input, which causes crashes for almost every model in kgcnn.
• Fix Error in ExtensiveMolecularLabelScaler.transform missing default value.
• Added further benchmark results for kgcnn version 4.
• Fix error in kgcnn.layers.geom.PositionEncodingBasisLayer
• Fix error in kgcnn.literature.GCN.make_model_weighted
• Fix error in kgcnn.literature.AttentiveFP.make_model
• Had to change serialization for activation functions since with keras>=3.0.2 custom strings are not allowed also causing clashes with built-in functions. We catch defaults to be at least as backward compatible as possible and changed to serialization dictionary. Adapted all hyperparameter.
• Renamed leaky_relu and swish in kgcnn.ops.activ to leaky_relu2 and swish2.
• Fix error in jax scatter min/max functions.
• Added kgcnn.__safe_scatter_max_min_to_zero__ for tensorflow and jax backend scattering with default to True.
• Added simple ragged support for loss and metrics.
• Added simple ragged support for train_force.py
• Implemented random equivariant initialize for PAiNN
• Implemented charge and dipole output for HDNNP2nd
• Implemented jax backend for force models.
• Fix GraphBatchNormalization .
• Fix error in kgcnn.io.loader for unused IDs and graph state input.
• Added experimental DisjointForceMeanAbsoluteError

Completely reworked version of kgcnn for Keras 3.0 and multi-backend support. A lot of fundamental changes have been made. However, we tried to keep as much of the API from kgcnn 3.0 so that models in literature can be used with minimal changes. Mainly, the "input_tensor_type"="ragged" model parameter has to be added if ragged tensors are used as input in tensorflow. For very few models also the order of inputs had to be changed. Also note that the input embedding layer requires integer tensor input and does not cast from float anymore.

The scope of models has been reduced for initial release but will be extended in upcoming versions. Note that some changes are also stem for keras API changes, like for example learning_rate rate parameter or serialization. Moreover, tensorflow addons had to be dropped for keras 3.0 .

The general representations of graphs has been changed from ragged tensors (tensorflow only, not supported by keras 3.0) to the disjoint graph representation compatible with e.g. PyTorchGeometric. Input can be padded or (still) ragged input. Or direct disjoint representations with batch loader. (See models chapter in docs).

For jax we added a padded_disjoint parameter that can enable jit'able jax models but requires a dataloader, which is not yet thoroughly implemented in kgcnn . For padded samples it can already been tested, but the padding of each sample is a much larger overhead than padding the batch.

Some other changes:

• Reworked training scripts to have a single train_graph.py script. Command line arguments are now optional and just used for verification, all but category which has to select a model/hyperparameter combination from hyper file. Since the hyperparameter file already contains all necessary information.
• Train test indices can now also be set and loaded from the dataset directly.
• Scaler behaviour has changed with regard to transform_dataset. Key names of properties to transform has been moved to the constructor! Also be sure to check StandardLabelScaler if you want to scale regression targets, since target properties are default here.
• Literature Models have an optional output scaler from new kgcnn.layers.scale layer controlled by output_scaling model argument.
• Input embedding in literature models is now controlled with separate input_node_embedding or input_edge_embedding arguments which can be set to None for no embedding. Also embedding input tokens must be of dtype int now. No auto-casting from float anymore.
• New module kgcnn.ops with kgcnn.backend to generalize aggregation functions for graph operations.
• Reduced the models in literature. Will keep bringing all models of kgcnn<4.0.0 back in next versions and run benchmark training again.

• Added flexible charge for rdkit_xyz_to_mol as e.g. list.
• Added from_xyz to MolecularGraphRDKit .
• Started additional kgcnn.molecule.preprocessor module for graph preprocessors.
• BREAKING CHANGES: Renamed module kgcnn.layers.pooling to kgcnn.layers.aggr for better compatibility. However, kept legacy pooling module and all old ALIAS.
• Repair bug in RelationalMLP .
• HyperParameter is not verified on initialize anymore, just call hyper.verify().
• Moved losses from kgcnn.metrics.loss into separate modul kgcnn.losses to be more compatible with keras.
• Reworked training scripts especially to simplify command line arguments and strengthen hyperparameter.
• Started with potential keras-core port. Not yet tested or supported.
• Removed get_split_indices to make the graph indexing more consistent.
• Started with keras-core integration. Any code is WIP and not tested or working yet.

• Added add_eps to PAiNNUpdate layer as option.
• Rework data.transform.scaler.standard to hopefully now fix all errors with the scalers.
• BREAKING CHANGES: Refactored activation functions kgcnn.ops.activ and layers kgcnn.layers.activ that have trainable parameters, due to keras changes in 2.13.0. Please check your config, since parameters are ignored in normal functions! If for example "kgcnn>leaky_relu" you can not change the leak anymore. You must use a kgcnn.layers.activ for that.
• Rework kgcnn.graph.methods.range_neighbour_lattice to use pymatgen.
• Added PolynomialDecayScheduler
• Added option for force model to use normal gradient and added as option use_batch_jacobian .
• BREAKING CHANGES: Reworked kgcnn.layers.gather to reduce/simplify code and speed up some models. The behaviour of GatherNodes has changed a little in that it first splits and then concatenates. The default parameters now have split_axis and concat_axis set to 2. concat_indices has been removed. The default behaviour of the layer however stays the same.
• An error in layer FracToRealCoordinates has been fixed and improved speed.

• Removed deprecated molecules.
• Fix error in kgcnn.data.transform.scaler.serial
• Fix error in QMDataset if attributes have been chosen. Now set_attributes does not cause an error.
• Fix error in QMDataset with labels without SDF file.
• Fix error in kgcnn.layers.conv.GraphSageNodeLayer .
• Add reverse_edge_indices option to GraphDict.from_networkx . Fixed error in connection with kgcnn.crystal .
• Started with kgcnn.io.file . Experimental. Will get more updates.
• Fix error with StandardLabelScaler inheritance.
• Added workflow notebook examples.
• Fix error in import kgcnn.crystal.periodic_table to now properly include package data.

Major refactoring of kgcnn layers and models. We try to provide the most important layers for graph convolution as kgcnn.layers with ragged tensor representation. As for literature models only input and output is matched with kgcnn .

• Move kgcnn.layers.conv to kgcnn.literature .
• Refactored all graph methods in graph.methods .
• Moved kgcnn.mol.* and kgcnn.moldyn.* into kgcnn.molecule
• Moved hyper into trainig
• Updated crystal .

• Added ACSFConstNormalization to literature models as option.
• Adjusted and reworked MLP . Now includes more normalization options.
• Removed 'is_sorted', 'node_indexing' and 'has_unconnected' from GraphBaseLayer and added it to the pooling layers directly.

• HOTFIX: Changed MemoryGraphList.tensor() so that the correct dtype is given to the tensor output. This is important for model loading etc.
• Added CENTChargePlusElectrostaticEnergy to kgcnn.layers.conv.hdnnp_conv and kgcnn.literature.HDNNP4th .
• Fix bug in latest train_force.py of v2.2.2 that forgets to apply inverse scaling to dataset, causing subsequent folds to have wrong labels.
• HOTFIX: Updated MolDynamicsModelPredictor to call keras model without very expensive retracing. Alternative mode use use_predict=True .
• Update training results and data subclasses for matbench datasets.
• Added GraphInstanceNormalization and GraphNormalization to kgcnn.layers.norm .

• Reworked all scaler class to have separate name for using either X or y. For example StandardScaler or StandardLabelScaler .
• Moved scalers to kgcnn.data.transform . We will expand on this in the future.
• IMPORTANT: Renamed and changed behaviour for EnergyForceExtensiveScaler . New name is EnergyForceExtensiveLabelScaler . Return is just y now. Added experimental functionality for transforming dataset.
• Adjusted training scripts for new scalers.
• Reduced requirements for tensorflow to 2.9.
• Renamed kgcnn.md to kgcnn.moldyn for naming conflicts with markdown.
• In MolDynamicsModelPredictor renamed argument model_postprocessor to graph_postprocessor .

• HOTFIX: Removed tensorflow_gpu from setup.py
• Added HDNNP4th.py to literature.
• Fixed error in ChangeTensorType config for model save.
• Merged pull request for #103 for kgcnn.xai .