Transfer of Distributions between External Conditions with Normalizing Flows

Code for the paper TRADE: Transfer of Distributions between External Conditions with Normalizing Flows

Installation

First go to https://github.com/noegroup/bgmol and follow the instructions to install the bgmol package, including openMM and mdtraj. Then run the following commands in this github repository:

pip install -r requirements.txt
pip install -e .

Fixes for broken packages:

Currently the package bgmol uses versioneer, which only works for python<3.12.
The package lightning_trainable is missing the requirments einops and requests, which need to be manually installed.
bgmol relies on an old version of mdtraj, which is broken with current cython. A dirty fix is to change the import statements in bgmol/tpl/hdf5.py from
```
from mdtraj.utils.six import PY3
...
from mdtraj.utils.six import string_types
```
to
```
from six import PY3
...
from six import string_types
```
Temperature steerable splines are not provided in the public implementations of bgflow and nflows, we provide forks in requirements.txt with our implementation.
Currently there exist several issues in bgflow, which lead to
1. Ignoring the temperature parameter passed to the Boltzmann generator for sampling
2. Only being able to evaluate the nll of the Boltzmann generator at temperature 1.0
3. Incorrect computation of the energy for augmented normalizing flows
When trying to evaluate models trained with bgflow we therefore advise you to implement metrics and sampling functions yourself or directly using the BGFlowFlow class

Usage

Currently, the lattice model and Gaussian mixture model from Appendix C are run differently from the other datasets. To run the experiments for the lattice model and the two-dimensional Gaussian mixture model, use

python3 train_INN.py --tag <your experiment tag> --config_path ./configs/<data set name>/config_<experiment name>.json

All other experiments can be called via lightning trainable as follows:

Define your configurations in a config.yaml file. You can find various default configs in the configs folder. Then start the experiment using lightning trainable:

python3 -m lightning_trainable.launcher.fit configs/config.yaml

You can also use multiple config files at once. Later configs overwrite the values specified in earlier ones.

Datasets

You can download all datasets from google drive. Alanine dipeptide is downloaded on-demand.

curl -L -o data.zip "https://drive.google.com/uc?export=download&id=10u03uwwoPTN3qAXZh94gTa9JJ-v-H4pB"
unzip data.zip

You can choose a dataset for training by including the following lines in your lightning trainable config.yaml

Double well 2d:

dataset:
  name: double_well_2d

Gaussian mixture model from Flow Annealed Importance Sampling Bootstrap (not Appendix C of our paper):

dataset:
  name: gmm_2d

Multi well 5d:

dataset:
  name: multi_well_5d

Two Moons:

dataset:
  name: two_moons

Alanine Dipeptide 66(/60)d:

dataset:
  name: ala2

Citation

You can cite our work using the following bibtex entry:

@article{wahl2024trade,
  title={TRADE: Transfer of Distributions between External Conditions with Normalizing Flows},
  author={Wahl, Stefan and Rousselot, Armand and Draxler, Felix and K{\"o}the, Ullrich},
  journal={arXiv preprint arXiv:2410.19492},
  year={2024}
}

vislearn/trade

Transfer of Distributions between External Conditions with Normalizing Flows

Installation

Usage

Datasets

Citation