RXN package for OpenNMT-based models

This repository contains a Python package and associated scripts for training reaction models based on the OpenNMT library. The repository is built on top of other RXN packages; see our other repositories rxn-utilities, rxn-chemutils, and rxn-onmt-utils.

For the evaluation of trained models, see the rxn-metrics repository.

Links:

• GitHub repository
• Documentation
• PyPI package

This repository was produced through a collaborative project involving IBM Research Europe and Syngenta.

System Requirements

This package is supported on all operating systems. It has been tested on the following systems:

• macOS: Big Sur (11.1)
• Linux: Ubuntu 18.04.4

A Python version of 3.6, 3.7, or 3.8 is recommended. Python versions 3.9 and above are not expected to work due to compatibility with the selected version of OpenNMT.

Installation guide

The package can be installed from Pypi:

pip install rxn-onmt-models[rdkit]

You can leave out [rdkit] if RDKit is already available in your environment.

For local development, the package can be installed with:

pip install -e ".[dev,rdkit]"

Training models.

Example of usage for training RXN models

The easy way

Simply execute the interactive program rxn-plan-training in your terminal and follow the instructions.

The complicated way

0. Optional: set shell variables, to be used in the commands later on.

MODEL_TASK="forward"

# Existing TXT files
DATA_1="/path/to/data_1.txt"
DATA_2="/path/to/data_2.txt"
DATA_3="/path/to/data_3.txt"

# Where to put the processed data
DATA_DIR_1="/path/to/processed_data_1"
DATA_DIR_2="/path/to/processed_data_2"
DATA_DIR_3="/path/to/processed_data_3"

# Where to save the ONMT-preprocessed data
PREPROCESSED="/path/to/onmt-preprocessed"

# Where to save the models
MODELS="/path/to/models"
MODELS_FINETUNED="/path/to/models_finetuned"

1. Prepare the data (standardization, filtering, etc.)

rxn-prepare-data --input_data $DATA_1 --output_dir $DATA_DIR_1

2. Preprocess the data with OpenNMT

rxn-onmt-preprocess --input_dir $DATA_DIR_1 --output_dir $PREPROCESSED --model_task $MODEL_TASK

3. Train the model (here with small parameter values, to make it fast on CPU for testing).

rxn-onmt-train --model_output_dir $MODELS --preprocess_dir $PREPROCESSED_SINGLE --train_num_steps 10 --batch_size 4 --heads 2 --layers 2 --transformer_ff 512 --no_gpu

Multi-task training

For multi-task training, the process is similar. We need to prepare also the second data set; in addition, the OpenNMT preprocessing and training take additional arguments. To sum up:

rxn-prepare-data --input_data $DATA_1 --output_dir $DATA_DIR_1
rxn-prepare-data --input_data $DATA_2 --output_dir $DATA_DIR_2
rxn-prepare-data --input_data $DATA_2 --output_dir $DATA_DIR_3
rxn-onmt-preprocess --input_dir $DATA_DIR_1 --output_dir $PREPROCESSED --model_task $MODEL_TASK \
  --additional_data $DATA_DIR_2 --additional_data $DATA_DIR_3
rxn-onmt-train --model_output_dir $MODELS --preprocess_dir $PREPROCESSED --train_num_steps 30 --batch_size 4 --heads 2 --layers 2 --transformer_ff 256 --no_gpu \
  --data_weights 1 --data_weights 3 --data_weights 4

Continuing the training

Continuing training is possible (for both single-task and multi-task); it needs fewer parameters:

rxn-onmt-continue-training --model_output_dir $MODELS --preprocess_dir $PREPROCESSED --train_num_steps 30 --batch_size 4 --no_gpu \
  --data_weights 1 --data_weights 3 --data_weights 4

Fine-tuning

Fine-tuning is in principle similar to continuing the training. The main differences are the potential occurrence of new tokens, as well as the optimizer being reset. There is a dedicated command for fine-tuning. For example:

rxn-onmt-finetune --model_output_dir $MODELS_FINETUNED --preprocess_dir $PREPROCESSED --train_num_steps 20 --batch_size 4 --no_gpu \
  --train_from $MODELS/model_step_30.pt

The syntax is very similar to rxn-onmt-train and rxn-onmt-continue-training. This is compatible both with single-task and multi-task.

Saving space by stripping the models

The model files can be quite large (~250 MB) for the default architecture. Roughly two thirds of the size is taken by the optimizer state, which is usually not needed, except maybe for the last checkpoint. You can strip the models with the command

rxn-onmt-strip-checkpoints -m $MODELS