The architecture of this project is inspired by https://github.com/open-mmlab
There are GraphSAGE, GAT models. Other models will be added soon. Stay tuned!
GraphSAGE
Inductive Representation Learning on Large Graphs (William L. Hamilton et al.)
GAT
Graph Attention Networks (Yoshua Bengio et al.)
- Python==3.6+
- tensorflow==2.3.0
- tensorflow-tensorboard==2.3.0
- numpy==1.18
- networkx==2.4
- tensorflow_addons==0.9.1
Config file:
- In model dict define the model architecture
- Define learning rate schedule and optimizer params
- In train_cfg and test_cfg define needed settings for training/testing phase that are used in model class.
- Define dataset_type for training/validation/testing
- Define data loader type and its chain operations that are exist in tf.data API
- In work_dir set up folder name for saving logs and models
- restore_model_path for restoring the needed model
- To set training and validation pipelines use workflow variable: type of mode (train/valid) and number of times for repeating the given mode
Training:
For training run main_train.py file
In API folder train.py contains whole high-level logic for training, in runner.py all details of training process.
Valid PPI dataset
| Mean Aggregator | Max Pool Aggregator | RNN Aggregator |
|---|---|---|
 |
 |
 |
| Node embeddings by ground truth label #23 | ||
 |
 |
 |
| Node embeddings by labels with the highest score | ||
 |
 |
 |