MatDGL is a neural network package that allows researchers to train custom models for material modeling tasks. It aims to accelerate the research and application of material science. It provides user a series of state-of-the-art models and supports user's innovative researches.
- Easy to installation.
- Three steps to fast testing.
- Flexible and adaptive to user's trainning task.
MatDGL can be installed easily through anaconda! As follows:
-
Create a new conda environment named "matdgl" by command, then activate environment "matdgl":
conda create -n matdgl python=3.8 conda activate matdgl
It's necessary to create a new conda environment to aviod bugs causing by version conflict.
-
Configure dependencies of matdgl:
conda install -c conda-forge tensorflow-gpu
-
Install pymatgen:
conda install --channel conda-forge pymatgen
-
Install other dependencies:
conda install --channel conda-forge mendeleev conda install --channel conda-forge graphviz conda install --channel conda-forge pydot conda install --channel conda-forge sklearn
-
Install matdgl:
pip install matdgl
MatDGL is very easy to use!
Just three steps can finish a fast test using matdgl:
- download test data
Get test datas from https://github.com/huzongxiang/MatDGL/tree/main/datas/
There are four json files in datas: dataset_classification.json, dataset_multiclassification.json, dataset_regression.json
and dataset_pretrain.json. - prepare workdir
Download datas and put it in your trainning work directory, test.py file should also be put in the directoryworkdir │ test.py | └───datas │ dataset_classification.json │ dataset_multiclassification.json │ dataset_regression.json │ dataset_pretrain.json - run command
run command:python test.py
You have finished your testing multi-classification trainning! The trainning results and model weight could be saved in /results and /models, respectively.
You can use matdgl by provided trainning scripts in user_easy_trainscript only, but understanding script will help you custom your trainning task!
-
get datas
Get current work directory of running trainning script, the script will read datas from 'workdir/datas/' , then saves results and models to 'workdir/results/' and 'workdir/models/'from pathlib import Path ModulePath = Path(__file__).parent.absolute() # workdir
-
fed trainning datas
Module Dataset will read data from 'ModulePath/datas/dataset.json', 'task_type' defines regression/classification/multi-classification, 'data_path' gets path of trainning datas.from matdgl.data import Dataset dataset = Dataset(task_type='multiclassfication', data_path=ModulePath)
-
generator
Module GraphGenerator feds datas into model during trainning. The Module splits datas into train, valid, test sets, and transform structures data into labelled graphs and gets three generators. BATCH_SIZE is batch size during trainning, DATA_SIZE defines number of datas your used in entire datas, CUTOFF is cutoff of graph edges in crystal.from matdgl.data.generator import GraphGenerator BATCH_SIZE = 128 DATA_SIZE = None CUTOFF = 2.5 Generators = GraphGenerator(dataset, data_size=DATA_SIZE, batch_size=BATCH_SIZE, cutoff=CUTOFF) train_data = Generators.train_generator valid_data = Generators.valid_generator test_data = Generators.test_generator #if task is multiclassfication, should define variable multiclassifiction multiclassification = Generators.multiclassification
-
building model
Module GNN defines a trainning framework that accepts a series of models. MatDGL provides a series of mainstream models as your need.from matdgl.models import GNN from matdgl.models.gnnmodel import MpnnBaseModel, TransformerBaseModel, CgcnnModel, GraphAttentionModel gnn = GNN(model=MpnnBaseModel, atom_dim=16 bond_dim=64 num_atom=118 state_dim=16 sp_dim=230 units=32 edge_steps=1 message_steps=1 transform_steps=1 num_attention_heads=8 dense_units=64 output_dim=64 readout_units=64 dropout=0.0 reg0=0.00 reg1=0.00 reg2=0.00 reg3=0.00 reg_rec=0.00 batch_size=BATCH_SIZE spherical_harmonics=True regression=dataset.regression optimizer = 'Adam' )
-
trainning
Using trainning function of model to train. Common trainning parameters can be defined, workdir is current directory of trainning script, it saves results of model during trainning. If test_data exists, model will predict on test_data.gnn.train(train_data, valid_data, test_data, epochs=700, lr=3e-3, warm_up=True, load_weights=False, verbose=1, checkpoints=None, save_weights_only=True, workdir=ModulePath)
-
prediction
The simplest method for predicting is using script predict.py in /user_easy_train_scripts.
Using predict_data funciton to predict.gnn.predict_datas(test_data, workdir=ModulePath) # predict on test datas with labels y_pred_keras = gnn.predict(datas) # predict on new datas without labels
-
preparing your custom datas
If you have your structures (and labels), the Dataset receives pymatgen.core.Structure type. So you should transform your POSCAR or cif to pymatgen.core.Structure type.import os from pymatgen.core.structure import Structure structures = [] # your structure list for cif in os.listdir(cif_path): structures.append(Structure.from_file(cif)) # for POSCAR too # construct your dataset from matdgl.data import Dataset dataset = Dataset(task_type='my_classification', data_path=ModulePath) # task_type could be my_regression, my_classification, my_multiclassification dataset.prepare_x(structures) dataset.prepare_y(labels) # if you have labels used to trainning model, labels could be None in prediction on new datas without labels # alternatively, you can construct dataset as follow dataset.structures = structures dataset.labels = labels # save your structures and labels to dataset in dataset_my*.json dataset.save_datasets(strurtures, labels) # for prediction on new datas without labels, Generators has not attribute multiclassification, should assign definite value Generators = GraphGenerator(dataset, data_size=DATA_SIZE, batch_size=BATCH_SIZE, cutoff=CUTOFF) # dataset.labels is None Generators.multiclassification = 5 multiclassification = Generators.multiclassification # multiclassification = 5
-
models provided by matdgl
We provide GraphModel, MpnnBaseModel, TransformerBaseModel, MpnnModel, TransformerModel, DirectionalMpnnModel, DirectionalTransformerModel and CGCNN model according to your demends. TransformerModel, GraphModel and MpnnModel are different models. TransformerModel is a graph transformer. MpnnModel is a massege passing neural network. GraphModel is a combination of TransformerModel and MpnnModel. MpnnBaseModel and TransformerBaseModel don't take directional informations of crystal into count so them run faster. MpnnBaseModel is the fastest model but accuracy is enough for most tasks. TransformerModel can achieve the hightest accuracy in most tasks. The CGCNN model is the crystal graph convolution neural network model. The GraphAttentionModel is the graph attention neural network.from matdgl.models import GNN from matdgl.models.gnnmodel import MpnnBaseModel, TransformerBaseModel , DirectionalMpnnModel, DirectionalTransformerModel, MpnnModel, TransformerModel, GraphModel, CgcnnModel, GraphAttentionModel
-
custom your model and trainning
The Module GNN provides a flexible trainning framework to accept tensorflow.keras.models.Model type customized by user. Yon can custom your model and train the model according to the following example.from tensorflow.keras.models import Model from tensorflow.keras import layers from matdgl.layers import MessagePassing from matdgl.layers import PartitionPadding def MyModel( bond_dim, atom_dim=16, num_atom=118, state_dim=16, sp_dim=230, units=32, message_steps=1, readout_units=64, batch_size=16, ): atom_features = layers.Input((), dtype="int32", name="atom_features_input") atom_features_ = layers.Embedding(num_atom, atom_dim, dtype="float32", name="atom_features")(atom_features) bond_features = layers.Input((bond_dim), dtype="float32", name="bond_features") local_env = layers.Input((6), dtype="float32", name="local_env") state_attrs = layers.Input((), dtype="int32", name="state_attrs_input") state_attrs_ = layers.Embedding(sp_dim, state_dim, dtype="float32", name="state_attrs")(state_attrs) pair_indices = layers.Input((2), dtype="int32", name="pair_indices") atom_graph_indices = layers.Input( (), dtype="int32", name="atom_graph_indices" ) bond_graph_indices = layers.Input( (), dtype="int32", name="bond_graph_indices" ) pair_indices_per_graph = layers.Input((2), dtype="int32", name="pair_indices_per_graph") x = MessagePassing(message_steps)( [atom_features_, edge_features, state_attrs_, pair_indices, atom_graph_indices, bond_graph_indices] ) x = PartitionPadding(batch_size)([x[0], atom_graph_indices]) x = layers.BatchNormalization()(x) x = layers.GlobalAveragePooling1D()(x) x = layers.Dense(readout_units, activation="relu", name='readout0')(x) x = layers.Dense(1, activation="sigmoid", name='final')(x) model = Model( inputs=[atom_features, bond_features, local_env, state_attrs, pair_indices, atom_graph_indices, bond_graph_indices, pair_indices_per_graph], outputs=[x], ) return model from matdgl.models import GNN gnn = GNN(model=MyModel, atom_dim=16, bond_dim=64, num_atom=118, state_dim=16, sp_dim=230, units=32, message_steps=1, readout_units=64, batch_size=16, optimizer='Adam', regression=False, multiclassification=None,) gnn.train(train_data, valid_data, test_data, epochs=700, lr=3e-3, warm_up=True, load_weights=False, verbose=1, checkpoints=None, save_weights_only=True, workdir=ModulePath)
You can set edge as your model output.
from matdgl.layers import EdgeMessagePassing def MyModel( bond_dim, atom_dim=16, num_atom=118, state_dim=16, sp_dim=230, units=32, message_steps=1, readout_units=64, batch_size=16, ): atom_features = layers.Input((), dtype="int32", name="atom_features_input") atom_features_ = layers.Embedding(num_atom, atom_dim, dtype="float32", name="atom_features")(atom_features) bond_features = layers.Input((bond_dim), dtype="float32", name="bond_features") local_env = layers.Input((6), dtype="float32", name="local_env") state_attrs = layers.Input((), dtype="int32", name="state_attrs_input") state_attrs_ = layers.Embedding(sp_dim, state_dim, dtype="float32", name="state_attrs")(state_attrs) pair_indices = layers.Input((2), dtype="int32", name="pair_indices") atom_graph_indices = layers.Input( (), dtype="int32", name="atom_graph_indices" ) bond_graph_indices = layers.Input( (), dtype="int32", name="bond_graph_indices" ) pair_indices_per_graph = layers.Input((2), dtype="int32", name="pair_indices_per_graph") x = EdgeMessagePassing(units, edge_steps, kernel_regularizer=l2(reg0), sph=spherical_harmonics )([bond_features, local_env, pair_indices]) x = PartitionPadding(batch_size)([x[1], bond_graph_indices]) x = layers.BatchNormalization()(x) x = layers.GlobalAveragePooling1D()(x) x = layers.Dense(readout_units, activation="relu", name='readout0')(x) x = layers.Dense(readout_units//2, activation="relu", name='readout1')(x) x = layers.Dense(1, name='final')(x) model = Model( inputs=[atom_features, bond_features, local_env, state_attrs, pair_indices, atom_graph_indices, bond_graph_indices, pair_indices_per_graph], outputs=[x], ) return model
The Module GNN has some basic parameter necessary to be defined but not necessary to be used:
class GNN: def __init__(self, model: Model, atom_dim=16, bond_dim=32, num_atom=118, state_dim=16, sp_dim=230, batch_size=16, regression=True, optimizer = 'Adam', multiclassification=None, **kwargs, ): """ pass """
MatDGL
We list currently supported GNN models:
- GCN from Kipf and Welling: Semi-Supervised Classification with Graph Convolutional Networks (ICLR 2017)
- GAT from Veličković et al.: Graph Attention Networks (ICLR 2018)
- GN from Battaglia et al.: Relational inductive biases, deep learning, and graph networks
- Transformer from Vaswani et al.: Attention Is All You Need (NIPS 2017)
Zongxiang Hu
Please contact me if you have any questions.
Mail: [email protected]
Wechat: voodoozx2015