Requirements and Installation

install requirements

create a new environment
conda create --name G2GT python=3.7
conda activate G2GT

pip install pytorch-lightning==1.4.5
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install networkx
pip install tensorboardX==2.4.1
pip install rdkit-pypi==2021.9.3
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.9.0+cu111.html
pip install pympler
pip install --upgrade easy-kubeflow
pip install Cython pip install joblib

Example Usage

#!/usr/bin/env bash
[ -z "${exp_name}" ] && exp_name="uspto"
[ -z "${seed}" ] && seed="0"
[ -z "${arch}" ] && arch="--ffn_dim 1024 --hidden_dim 256 --dropout_rate 0.1 --intput_dropout_rate 0.05 --attention_dropout_rate 0.1 --n_layer 6 --peak_lr 2.5e-4 --end_lr 1e-6 --head_size 12 --weight_decay 0.00 --edge_type one_hop --warmup_updates 3000 --tot_updates 700000"
[ -z "${batch_size}" ] && batch_size="5"
[ -z "${dataset_name}" ] && dataset_name="uspto-50k-split2-split"

echo -e "\n\n"
echo "=====================================ARGS======================================"
echo "arg0: $0"
echo "exp_name: ${exp_name}"      
echo "arch: ${arch}"             
echo "seed: ${seed}"             
echo "batch_size: ${batch_size}" 
echo "==============================================================================="

default_root_dir=$PWD
mkdir -p $default_root_dir
n_gpu=1
export CUDA_VISIBLE_DEVICES=0

python entry.py --num_workers 1 --num_nodes 1 --seed $seed --batch_size $batch_size  --min_epochs 100 --accumulate_grad_batches 1  --sync_batchnorm True  --val_check_interval 1000 \
      --dataset_name $dataset_name --gradient_clip_val 4 \
      --gpus $n_gpu  --accelerator ddp \
      $arch \
      --default_root_dir $default_root_dir --progress_bar_refresh_rate 10  \

Example Inference

bash inference.sh

# Licensed under the MIT License.

#!/usr/bin/env bash
[ -z "${arch}" ] && arch="--ffn_dim 1024 --hidden_dim 256 --dropout_rate 0.1 --intput_dropout_rate 0.1 --attention_dropout_rate 0.1 --n_layer 6 --peak_lr 2.5e-4 --end_lr 1e-6 --head_size 12 --weight_decay 0.00 --edge_type one_hop --warmup_updates 1000 --tot_updates 500000"
[ -z "$ckpt_name" ] && ckpt_name=last.ckpt

export CUDA_VISIBLE_DEVICES=2,4,6
default_root_dir=$PWD
n_gpu=1

python entry.py --num_workers 6 --seed 0 --batch_size 1 \
      --dataset_name uspto-full-distilled-weaklabel \
      --gpus $n_gpu --accelerator ddp  $arch \
      --default_root_dir $default_root_dir \
      --checkpoint_path $default_root_dir/lightning_logs/checkpoints/last.ckpt --test --progress_bar_refresh_rate 1 \
      --inference_path $default_root_dir/results/uspto-full-distilled-weaklabel-sample \
      --beam_size 50 \ #beam size for sampling and beam search
      #--sampling \ # use sampling or beam search

Calculate accuracy

Add multiple foldername if using more than one results (e.g. when combining sampling results with beam search results).
The raw inference outputs are under results folder python score.py --file uspto-50k-sampling uspto-50k-beam --beam 100

Notice

Currently the weak-ensemble tags are hard coded to be 50. Our final results are calculated using a joint set of beam search and weak_enemble sampling