Training of OpenNMT-based RXN models
Project description
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:
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
- 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"
- Prepare the data (standardization, filtering, etc.)
rxn-prepare-data --input_data $DATA_1 --output_dir $DATA_DIR_1
- Preprocess the data with OpenNMT
rxn-onmt-preprocess --input_dir $DATA_DIR_1 --output_dir $PREPROCESSED --model_task $MODEL_TASK
- 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.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file rxn-onmt-models-1.1.0.tar.gz
.
File metadata
- Download URL: rxn-onmt-models-1.1.0.tar.gz
- Upload date:
- Size: 25.6 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.11.6
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | b2fb6c8b72340c51f380cb339c40ff97a7ad72e388afd0bcfa3f6d94970667ab |
|
MD5 | e6fc3c0f1fd2b7698fba2cbedac78ae5 |
|
BLAKE2b-256 | 1659f53a51b7cb9221f69b49360699926cfe7b759b0e6f891001f31c1f4f4757 |
File details
Details for the file rxn_onmt_models-1.1.0-py3-none-any.whl
.
File metadata
- Download URL: rxn_onmt_models-1.1.0-py3-none-any.whl
- Upload date:
- Size: 29.5 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.11.6
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 62975185ac0fb8572bda9b6e1bcd59adcab31b6c33fb2980bed0ea56ede8a72e |
|
MD5 | a54d561129829c226c06543fa790b9bd |
|
BLAKE2b-256 | 23c07074652131ff547d8e3d0d96159069617767b857a1c6dbfa811c926394d9 |