WIP - please use with care.
Online kinetics trains a deep kinetic models of biomolecular simulations (using the VAMPnets
using Deeptime package as inspiration) in an online manner:
i.e., using 1 epoch of training.
I gave a presentation on this at ScotChem in 2022, you can view the slides here
The motivation for this idea is that unbiased simulations used to create kinetic models can be quite large and potentially prohibitively expensive to train, especially in automated adaptive sampling methods (e.g., Casalino et. al.)
The intended use case is that approximate models are trained online for EDA purposes and then more accurate models
trained in a batch / multi-epoch manner.
Alternatively, one could use an approximate online model as a pre-processing step in creating an approximate discrete
Markov state model (for an e.g., see here)
for use with adaptive sampling.
The current method uses Hedged Back Propagation: Sahoo, D. et al. (2017) There are tentative plans to use the self-expanding neural networks of Mitchel et. al. (2023)
celerity- a package for training (online and batch) VAMPnet models of molecular kinetics.notebooks- analysis of models and data.tests- unit and integration tests will go here eventually.
Assuming a linux OS with CUDA drivers 11.3 - see Pytorch.org for details for other distros/CUDA versions.
conda create -n onlinekinetics python==3.9 -y
conda activate onlinekinetics
conda env update -f environment.yaml
pip install -qe .
Create a virtual environment your favourite way and then:
pip install -r requirements.txt
pip install -qe .