ogblsc2022

log

• 2022-7-18 model: GIN-Virtual {'Train': 0.0948409306068622, 'Validation': 0.11321253329515457} Best validation MAE so far: 0.10685155540704727 {'Train': 0.10280203166671775, 'Validation': 0.1292327642440796} Best validation MAE so far: 0.12899625301361084

Workflow

• Tiger: GNN简化，EGAT等
• Leo: 基于3D的GNN构建，Pretrain任务
• Ruifeng: 官能团信息的引入

Distributed Parallel

You should add the following code in .bashrc, to keep Pytroch multiGPU works.

New Model

Dataset pipeline:

• Train Set (3D Geometry original + RDKIT extracted, 2D Graph, Node & Edge Attr.)
• Valid/Test Set (3D Geometry extracted by RDKIT, 2D Graph, Node & Edge Attr.)
• Train Split

Model pipeline:

• Geometry learning (Random Sampling from train set)
• ComENet Based + Edge Attr. + Attention mechanism

export NCCL_SHM_DISABLE=1