Deep Learning for Collective Variables (DeepCV)

DeepCV is a computer code that implements our developed unsupervised machine learning called deep autoencoder neural network (DAENN) for learning low-dimensional collective variables (CVs) aka slow-mode reaction coordinates of a set of molecules for enhanced sampling.

Website: https://lubergroup.pages.uzh.ch/deepcv/

Main Features

1. Molecular features
   • Internal coordinates
   • SPRINT and eXtended SPRINT coordinates
2. Dense autoencoder neural nets
   • Single and multi-input simple and stacked autoencoder
   • Avoid saturation
   • GPU acceleration
3. Learn CVs in expanded configurational space
   • Customized loss functions with minimaxation technique
     • Primary loss: main loss with regularization
     • Secondary loss: Additional loss to be maximized to expand the CV space
   • Self-directed expansion of configurational space
4. Generative model for generating data
   • Generative adversarial networks (GANs)
   • Variational autoencoder (future work)
5. Can interface with PLUMED and CP2K
6. Input file generator (GUI) for PLUMED and CP2K
7. Analysis tools
   • Feature importance
   • Sampling convergence assessment

Quick installation

• Python codes

  cd deepcv/
  pip install -r requirements.txt
  pip install .

• C++ codes

  cd deepcv/
  export LIB_TF=/usr/local/tensorflow/
  
  g++ -Wall -fPIC -o deepcv.o deepcv.cpp \
      -O3 -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++14 -fPIC \
      -I${LIB_TF}/include/ -L${LIB_TF}/lib \
      -ltensorflow -ltensorflow_cc -ltensorflow_framework
  
  g++ -shared -fPIC -o deepcv.so deepcv.o

  or you can use make

  make CXXFLAGS="-std=c++14 -fPIC"

Usage

The following is an example command for training model of CVs for Diels-Alder reaction using features extracted from reactant trajectory. You can call DeepCV's DAENN via either main.py API or deepcv_daenn command (register entrypoint).

python main.py daenn -i input_ae_DA.json

or

deepcv_daenn -i input_ae_DA.json

---

Development

• Python 3.6 or a newer version
• Use git control: git clone https://gitlab.uzh.ch/lubergroup/deepcv.git
• Please write function docstring and comment for difficult-to-understand code
• Document modules and packages you are developing
• Format codes with Black
• Send pull-request to master with an explanation, for example, what you contribute, how it works, and usefulness

Packages requirements

To install all dependencies packages of DeepCV, you can follow either following way:

1. All at once (for the users)

   • Using PIP
     • pip install --upgrade pip
     • pip install -r requirements.txt
   • Using Conda (recommended)
     • conda update --all -y
     • conda install --file requirements.txt

2. Install packages separately (recommended for the developers)

   • NumPy >= 1.22.2
     • E.g., pip install --upgrade numpy==1.23.4
   • TensorFlow + Keras 2.10.0
     • pip install tensorflow (CPU)
     • conda install tensorflow-gpu (+ CUDA & cuDNN)
   • NVIDIA GPU and CUDA 10.1 (for GPU enabled)
     • https://developer.nvidia.com/cuda-toolkit-archive
   • cuDNN v7.6.4 (September 27, 2019), for CUDA 10.1
     • https://developer.nvidia.com/rdp/cudnn-archive
   • pydot (for keras.utils.vis_utils.plot_model)
     • conda install pydot
   • other important packages are listed in requirements.txt

3. DeepCV C++ makes use of JSON parser

   • https://github.com/nlohmann/json

In Progress

1. Variational autoencoder
2. Improve neural network algorithm for large systems e.g. metal-oxide surface
3. Improve code compatibility between TensorFlow, PLUMED, and CP2K

---

Authors

1. Rangsiman Ketkaew (rangsiman.ketkaew@chem.uzh.ch)
2. Sandra Luber (sandra.luber@chem.uzh.ch)