MTL-Cov-GCN

A repository for the code for Cov-GCN, MTL-Cov_GCN and related experiments

Folder Introduction

├── [data] # to be initiated
│   ├── data_processing_v2/ # where data should be stored
│   └── data_processing.ipynb # how data is preprocessed
├── [logs] # to be initated
├── fold_linear_layer
│   ├── config.py # where parameters are stored, annotations on parameters are given as comment
│   └── dataset.py # data is repackaged into dataloader object
│   └── linear_layers.py # network
│   └── logger.py # how to save model, tensorboard writer, output tensor and more
│   └── loss_and_optimizer.py # construct loss function and optimizer
│   └── metrics.py # calculate metrics
│   └── MLM.py # masking part of the input, function NOT used in this research
│   └── trainer_single_process.py # main script to run if train for a sinlge model
│   └── trainer.py # main script to run if train for a multiple models in a loop
│   └── utils.py # auxilliary tools, set random seed, calculate sparsity and threshold, etc.
│   └── run.sh # bash script to train or test model, instructions inside
├── fold_trans
├── fold_trans_gnn
│   ├── beta_controller.py # changing beta, function NOT used in this research
│   └── ... 
├── fold_trans_gnn_adp_map
├── shallow_machine_learning_ARIMA
│   ├── param_search_feature_remove.py # search the best parameters for XGBoost, LightGBM, RandomForest and ARIMA
│   └── ... 
└── README.md

How to set up environment?

Install Python 3.8. 0. Preferably with conda environment, run

conda create -n gnn python=3.8 # create virtual environment with name gnn
conda activate gnn # go into virtual env. CAUTION: all installation below should happen inside the env
conda deactivate # exit virtual env

[OPTIONAL] For users from Mainland, China, add channel to speed-up package installation

pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple

Install 1.10.0 torch GPU version with CUDA=11.1, run

pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html

Install torch_scatter, torch_sparse, they are prerequiste for torch_geometric.
- Because they have strict version control w.r.t to torch version and cuda version, we provide path to the wheel file for installation. torch_scatter torch_sparse
- place .whl files inside the .whls/
- For other CUDA version please refer to Documentaion under Installation from Wheels -> Note -> You can look up the latest supported version number here.

pip install ../whls/torch_scatter-2.0.9-cp38-cp38-linux_x86_64.whl
pip install ../whls/torch_sparse-0.6.12-cp38-cp38-linux_x86_64.whl

Install torch_geometric for torch==1.10.0 with CUDA=11.1

pip install torch_geometric==2.0.2

Install other pacakges

pip install -r requirements.txt

Where to find data

You can obtained the well pre-processed datasets from Google Drive or Baidu Drive with four-character password: hmlr . Then place the downloaded data under the folder ./data

How to run experiments?

Run the bash file in each folder to run experiments.

Where is model and metrics saved?

They will be save in the logs/ folder with path assigned in the config.py

What do parameters in configs stands for?

Annotation to parameters are in the config.py.

If I want to use my own data, what parameters should I tune?

# change to your own
'dataset_path'
'pkl_filename'
'sensors'
'input_dim'
'target_dim'
'src_seq_len'
'trg_seq_len'
'past_history_factor'
'max_len'
# need to fine-tune by hand
'threshold'
'sparsity_ratio'
'factor'
'warmup'

SuhanG17 / MTL-Cov-GCN