ndnng/bt4molgen

BT4MolGen

Back Translation for Molecule Generation.

This project is based on fairseq.
To get help from the official code, see the original README.

Dataset Preparation

1. Molecular Property Improvement Dataset
2. USPTO: Retrosynthesis Prediction Dataset

Environment Preparation

1. Docker image: fyabc/bt4molgen:main

Data Preprocessing

1. Data source (canonical, without extra '@' chars): HERE.

   1. Placed in <./canonical-from-HierG2G>.

2. Data source (non-canonical, with extra '@' chars): HERE.

   1. Placed in <./full-from-JTNN>.

3. For our Transformer translation jobs, we need paired validation and test subsets, so we split train_pairs.txt into new valid and test files.

   1. Sizes: TODO

4. Tokenization

   cd /d/GitProjects/off-hg2g/fy-exp/data
   python /d/GitProjects/fairseq-v0.8.0/run/molecule/preprocessing/tokenize_smiles.py -t re drd2/train.m1 drd2/train.re.m1 --std

Baseline Training

1. HierG2G baseline training

   # Submit a debug job of 'stub-wu2-t-yafan-philly2-torch15-py36-cu101' (docker = 'pytorch1.5-py36-cuda10.1')
   
   conda create -n mol-hg2g python=3.6
   conda activate mol-hg2g
   conda install -y pytorch=1.5.0 torchvision cudatoolkit=10.1 -c pytorch
   conda install rdkit -c rdkit
   python -m pip install networkx
   
   cd /blob/v-yaf/off-hg2g/
   python setup.py develop
   
   for d in drd2 qed logp04 logp06; do
      mkdir -p train_processed/${d}/
      python preprocess.py --train data/${d}/train_pairs.txt --vocab data/${d}/vocab.txt --ncpu 16 < data/${d}/train_pairs.txt
      mv tensor* train_processed/${d}/
   done
   
   # Same for drd2, logp04, logp06
   mkdir -p models/qed/
   nohup python gnn_train.py --train train_processed/qed/ --vocab data/qed/vocab.txt --save_dir models/qed/ > models/qed/train-log.txt 2>&1 &
   
   for d in drd2 qed logp04 logp06; do
      mkdir -p models/$d/translations
      python decode.py --test data/$d/test.txt --vocab data/$d/vocab.txt --model models/$d/model.5 --num_decode 20 > models/$d/translations/translation-$d-pt5-20.txt
   done

2. HierG2G BT Training

   cd /blob/v-yaf/off-hg2g/
   mkdir -p fy-exp/data/reversed-baseline/tmp/
   
   # Prepare reversed data
   for d in drd2 qed logp04 logp06; do
      cut -d ' ' -f 1 data/$d/train_pairs.txt > fy-exp/data/reversed-baseline/tmp/$d-train_pairs-x.txt
      cut -d ' ' -f 2 data/$d/train_pairs.txt > fy-exp/data/reversed-baseline/tmp/$d-train_pairs-y.txt
      paste -d ' ' fy-exp/data/reversed-baseline/tmp/$d-train_pairs-y.txt fy-exp/data/reversed-baseline/tmp/$d-train_pairs-x.txt > fy-exp/data/reversed-baseline/$d-train_pairs.txt
   
      mkdir -p train_processed/$d-reversed/
      python preprocess.py --train fy-exp/data/reversed-baseline/$d-train_pairs.txt --vocab data/${d}/vocab.txt --ncpu 16 < fy-exp/data/reversed-baseline/$d-train_pairs.txt
      mv tensor* train_processed/$d-reversed/
   done
   
   # Train reversed HierG2G models, same for drd2, logp04, logp06
   d=qed
   mkdir -p models/$d-reversed/
   nohup python gnn_train.py --train train_processed/$d-reversed/ --vocab data/$d/vocab.txt --save_dir models/$d-reversed/ > models/$d-reversed/train-log.txt 2>&1 &
   
   # Generate BT data using reversed HierG2G models

Evaluation

BT Training

1. Translation output folder: HERE.
2. Train forward and backward baseline models
3. Generate BT data (x') using backward models
4. Make BT dataset: See make-bt-dataset.py.