DeepDriveMD-F (DeepDriveMD-pipeline)

DeepDriveMD-F: Deep-Learning Driven Adaptive Molecular Simulations (file-based continual learning loop)

Details can be found in the documentation. For more information, please see our website.

How to run

Running DeepDriveMD requires the use of virtual environment. At this point we distinguish different stage runs of DeepDriveMD using different virtual environments to alleviate package compatibility with associated dependencies across different stages.

For instance, below is a list of Python versions used by different virtual environments:

• RCT env: Python 3.7.8
• OpenMM env: Python 3.7.9
• pytorch (AAE) env: Python 3.7.9
• keras-cvae (CVAE) & rapids-dbscan: Python 3.6.12

Setup

Stage: molecular_dynamics

1. Install deepdrivemd into a virtualenv with a Python virtual environment:

python3 -m venv env
source env/bin/activate
pip install --upgrade pip setuptools wheel
pip install -e .

Or with a Conda virtual environment:

. ~/miniconda3/etc/profile.d/conda.sh
conda create -n deepdrivemd python=3.7.9
conda activate deepdrivemd
pip install --upgrade pip setuptools wheel
conda install scipy (this step is needed if a failure of installing scipy is observed)
pip install -e .

2. Install OpenMM:

• by source code (for Linux ppc64le, e.g., Summit) https://gist.github.com/lee212/4bbfe520c8003fbb91929731b8ea8a1e

• by conda (for Linux x86_64, e.g., PSC Bridges)

module load anaconda3
module load cuda/9.2
source /opt/packages/anaconda/anaconda3-5.2.0/etc/profile.d/conda.sh
conda install -c omnia/label/cuda92 openmm

3. In some places, DeepDriveMD relies on external libraries to configure MD simulations and import specific ML models.

For MD, install the mdtools package found here: https://github.com/braceal/MD-tools

git clone https://github.com/braceal/MD-tools.git
pip install .

For ML (specifically the AAE model), install the molecules package found here: https://github.com/braceal/molecules/tree/main

git clone https://github.com/braceal/molecules.git
pip install .

Stage: machine_learning

1. Install the deepdrivemd virtual environment as above (deepdrivemd is needed in all the virtual environments since each task uses the DDMD_API to communicate with the outputs of other tasks).

2. Install the keras-CVAE model with rapidsai DBSCAN package found here: https://www.ibm.com/docs/en/wmlce/1.6.2?topic=installing-mldl-frameworks

conda config --prepend channels https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/
conda install powerai-rapids

3. Install packages scikit-learn and h5py version 2.10.0:

conda install scikit-learn h5py=2.10.0

4. Install the tensorflow-gpu package (need to compile with CUDA 10.2.89, not compatible with CUDA 10.1.243 and CUDA 11.1.1 or higher versions):

conda install tensorflow-gpu

Generating a YAML input spec:

First, run this command to get a sample YAML config file:

python -m deepdrivemd.config

This will write a file named deepdrivemd_template.yaml which should be adapted for the experiment at hand. You should configure the molecular_dynamics_stage, aggregation_stage, machine_learning_stage, model_selection_stage and agent_stage sections to use the appropriate run commands and environment setups.

Running an experiment

Then, launch an experiment with:

python -m deepdrivemd.deepdrivemd -c <experiment_config.yaml>

This experiment should be launched

Note on input data

The input PDB and topology files should have the following structure:

ls data/sys*

data/sys1:
comp.pdb comp.top

data/sys2:
comp.pdb comp.top

Where the topology files are optional and only used when molecular_dynamics_stage.task_config.solvent_type is "explicit". Only one system directory is needed but an arbitrary number are supported. Also note that the system directory names are arbitrary. The path to the data directory should be passed into the config via molecular_dynamics_stage.initial_pdb_dir.

DeepDriveMD-S (Streaming asynchronous execution with ADIOS)

The streaming version of DeepDriveMD uses the adios2 package.

adios2 is installed with spack:

spack install adios2 +python -mpi

To use adios2 in python, one needs to load the corresponding module, for example, with

module load adios2

or

spack load adios2

and to set up PYTHONPATH to the corresponding subdirectory of the adios2 installation:

export PYTHONPATH=<ADIOS2_dir>/lib/python<version>/site-packages/:$PYTHONPATH

To make a small 30m, 12 simulation, 1 aggregator, test run of DeepDriveMD-S, cd into test/ and run

make run1

To make a large 12h, 120 simulations, 10 aggregators run do

make run2

in DeepDriveMD-pipeline directory.

To watch how one of the aggregation files grows, do, for example

make watch1 d=3	

assuming that the experiment directory is ../Outputs/3.

To watch what happens in one of the simulation task directory, do

make watch2 d=3

To watch the log for task 0014 (for run1 it corresponds to the outlier search log), do

make watch3 d=0014

To clean after the run, do

make clean d=3

The configuration files for the run, including generate.py that is used to create config.yaml, adios xml files for SST streams between simulations and aggregators and for BP files between aggregators and the downstream two components, are in a subdirectory of test/bba, for example, test1_stream (run1) and lassen-keras-dbscan_stream (run2). Yaml files are generated by running ./generate.py > config.yaml or, if you prefer, you can edit config.yaml directly and not use generate.py.

To use multiple input files, put the corresponding pdb files into cfg.initial_pdb_dir. The simulation sorts pdb files from this directory and picks up the one corresponding to its task id modulo the number of pdb files.

Contributing

Please report bugs, feature requests, or questions through the Issue Tracker.

If you are looking to contribute, please see CONTRIBUTING.md.

License

DeepDriveMD has a MIT license, as seen in the LICENSE file.