CoolGraph is an easy-to-use Python library with Graph Neural Networks for node classification. The CoolGraph contains several architectures that will help you train a network using two lines of code.
Thus, the parameters for training have already been selected and collected in configs, but you can change them as you wish.
Also, if for some reason the selected parameters do not ok for you, it is possible to use the search for hyperparameters with Optuna.
Moreover, your experiments can be saved in Mlflow and fully tracked.
All you need is graph-structured data.
- Quick start with 2 lines of code
- Good quality of base models, comparable to state of the art
- Heterogeneous graph support
- The best model architecture automatic search via Optuna
- Tracking experiments with MLflow
- User could define the targets count and target weights in the loss function
- Etimating batch size and neighbourhood sampling sizes for the first and second hop via graph metrics calculation
For more details, look at tutorials in folder Notebooks.
conda deactivate
conda create -n cool_graph_env2_py38 python=3.8 cudatoolkit=11.3.1 pytorch=1.12.0=py3.8_cuda11.3_cudnn8.3.2_0 cxx-compiler=1.5.1 pyg=2.2.0=py38_torch_1.12.0_cu113 pyarrow=11.0.0 numpy=1.23.5 pandas=1.4.4 pip=22.3.1 py=1.11.0 mysqlclient=2.0.3 sqlite=3.38.2 psycopg2=2.8.6 optuna=2.10.1 -c nvidia -c pytorch -c conda-forge -c pyg
pip install cool-graph
conda deactivate
conda env create -f environment.yml
pip install cool-graph
You can use CoolGraph in Google Colab without installing the conda, but make sure that the default colab environment matches the required versions for the library. Google Colab
Look at page notebook in Run examples
or you can see the example with open fraud dataset from Yelp at fraud dataset notebook
Here is a graph with 2 groups in nodes Google Drive
Comming soon
In Coolgraph you can use default config structure but also you can change it. See below how to copy config structure to your path, see discovery in configs and run.
coolrun --config <path/to/config>
You can easily override config parameters:
coolrun --config ./cool_graph/config/full.yaml training.n_epochs=5
To copy config structure use command:
get_config --configs <path/to/config/where/you/need/it>
Easy run with Amazon Computers dataset:
# Load Dataset
from torch_geometric import datasets
data = datasets.Amazon(root='./data/Amazon', name='Computers').data
# Train GNN model
from cool_graph.runners import Runner
runner = Runner(data)
result = runner.run()You can override default parameters and/or read parameters from config file
runner = Runner(data,
metrics=['accuracy', ...],
batch_size='auto',
train_size=0.7,
test_size=0.3,
overrides=['training.n_epochs=1', ...],
config_path=...)
result = runner.run() You can run HypeRunner for the best GNN architecture search
# Load Dataset
from torch_geometric import datasets
data = datasets.Amazon(root='./data/Amazon', name='Computers').data
from cool_graph.runners import HypeRunner
runner = HypeRunner(data)
result = runner.optimize_run()For more information look at examples
Graph example from Google Drive
import torch
data = torch.load("sample_of_graph")
from cool_graph.runners import MultiRunner
runner = MultiRunner(data)
result = runner.run()The directory structure of CoolGraph:
├── config <- Config structure
│ ├── data <- Data configs (Datasets on disk)
│ ├── logging <- MLFlow configs
│ ├── metrics <- Metrics configs
│ ├── model_params <- NN model parameters configs
│ ├── training <- Training configs
│ ├── full.yaml <- Main config for CLI
│ ├── in_memory_data.yaml <- Main config for notebook (GraphConv)
│ ├── in_memory_data2.yaml <- Main config for notebook (NNConv)
│
├── cli <- Cli commands
├── data <- Data processing, data loaders, batch sizes
├── logging <- MLflow logging experiments
├── models <- Cool graph models
├── parameter_search <- Sampling model params for Optuna
├── train <- Trainin / eval / metrics code
├── runners.py <- Run training (CLI + Notebook)
add dev dependencies here (or switch to poetry)
Install the package using one of the options described above adding -e flag to pip install.
Run git checkout -b <new_branch_name>.
Introduce changes to the branch.
Test your changes using make test.
Add tests if necessary (e.g. coverage fell below 80%), run make test again.
Ensure that codestyle is OK with make verify_format.
If not, run make format.
After that commit -> push -> PR.
`The MIT License (MIT) Copyright 2023 MTS (Mobile Telesystems). All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.`