- Date: 2021.08.04 ~ 2021.09.27 18:00
- Task: Molecular property prediction (The gap between S1 and T1 energy)
- Result: 21st place / 220 teams
Generating node-level feature / edge type
- Using only atom type (13 dim)
- Each atom is embedded into 256 dimensional vector by a simple linear transformation
- There are 6 edge type : the number of combinations of (
bond type
,bond stereo
)
Relational Graph Convolutional Network (RGCN)
- Total 8 RGCN layers each of which has 256 channels
- Skip-connections
- Using the sum of node representations followed by one hidden layer MLP as the graph representation
Readout phase
- Multi layers perceptron with 2 hidden layers (1,024 dim, 512 dim)
- Dropout with p=0.3
- Directly predicting
ST1 gap
10-Fold ensembling
- Taking the simple average of 10 models
- Data preparation
dir_data
: the directory where train.csv, dev.csv, and test.csv are storeddir_output
: the directory where the preprocessedtgm.data.Data
files will be stored.
python gnn_preprocess.py --dir_data './data' --dir_output './outputs/rgcn'
- Training a single model and predicting test data
- Modify
TrainConfig
inconfig.py
- It gives public score about 0.127
python train.py