• Install DeePMD-kit
  • Install tensorflow's Python interface
  • Install tensorflow's C++ interface
  • Install xdrfile
  • Install DeePMD-kit
  • Install Lammps' DeePMD-kit module
• Use DeePMD-kit
  • Prepare data
  • Train a model
  • Freeze the model
  • Run MD with Lammps
  • Run path-integral MD with i-PI
  • Run MD with native code
• Code structure
• License

The installation of the DeePMD-kit is lengthy, but do not be panic. Just follow step by step. Wish you good luck..

A docker for installing the DeePMD-kit on CentOS 7 is available here.

There are two ways of installing the Python interface of tensorflow, either using google's binary, or installing from sources. When you are using google's binary, do not forget to add the option -DTF_GOOGLE_BIN=true when building DeePMD-kit.

Firstly get the source code of the tensorflow

cd /some/workspace
git clone https://github.com/tensorflow/tensorflow tensorflow

The DeePMD-kit works with tensorflow r1.4 -- r1.6. Now taking r1.4 for example:

cd tensorflow
git checkout r1.4

Please make sure you have the Bazel higher than version 0.5.4, otherwise, please install it.

DeePMD-kit is compiled by cmake, so we need to compile and integrate tensorflow with cmake projects. The rest of this section basically follows the instruction provided by Tuatini. Now execute

./configure

You will answer a list of questions that help configure the building of tensorflow. It is recommended to build for Python3. You may want to answer the question like this:

Please specify the location of python. [Default is /usr/bin/python]: /usr/bin/python3

The library path for Python should be set accordingly.

Now build the shared library of tensorflow:

bazel build -c opt --verbose_failures //tensorflow:libtensorflow_cc.so

You may want to add options --copt=-msse4.2, --copt=-mavx, --copt=-mavx2 and --copt=-mfma to enable SSE4.2, AVX, AVX2 and FMA SIMD accelerations, respectively. It is noted that these options should be chosen according to the CPU architecture. If the RAM becomes an issue of your machine, you may limit the RAM usage by using --local_resources 2048,.5,1.0.

Now I assume you want to install tensorflow in directory $tensorflow_root. Create the directory if it does not exists

mkdir -p $tensorflow_root

Before moving on, we need to compile the dependencies of tensorflow, including Protobuf, Eigen and nsync. Firstly, protobuf

mkdir /tmp/proto
tensorflow/contrib/makefile/download_dependencies.sh
cd tensorflow/contrib/makefile/downloads/protobuf/
./autogen.sh
./configure --prefix=/tmp/proto/
make
make install

Then Eigen

mkdir /tmp/eigen
cd ../eigen
mkdir build_dir
cd build_dir
cmake -DCMAKE_INSTALL_PREFIX=/tmp/eigen/ ../
make install

And nsync

mkdir /tmp/nsync
cd ../../nsync
mkdir build_dir
cd build_dir
cmake -DCMAKE_INSTALL_PREFIX=/tmp/nsync/ ../
make
make install
cd ../../../../../..

Now, copy the libraries to the tensorflow's installation directory:

mkdir $tensorflow_root/lib
cp bazel-bin/tensorflow/libtensorflow_cc.so $tensorflow_root/lib/
cp bazel-bin/tensorflow/libtensorflow_framework.so $tensorflow_root/lib/
cp /tmp/proto/lib/libprotobuf.a $tensorflow_root/lib/
cp /tmp/nsync/lib/libnsync.a $tensorflow_root/lib/

Then copy the headers

mkdir -p $tensorflow_root/include/tensorflow
cp -r bazel-genfiles/* $tensorflow_root/include/
cp -r tensorflow/cc $tensorflow_root/include/tensorflow
cp -r tensorflow/core $tensorflow_root/include/tensorflow
cp -r third_party $tensorflow_root/include
cp -r /tmp/proto/include/* $tensorflow_root/include
cp -r /tmp/eigen/include/eigen3/* $tensorflow_root/include
cp -r /tmp/nsync/include/*h $tensorflow_root/include

Now clean up the source files in the header directories:

cd $tensorflow_root/include
find . -name "*.cc" -type f -delete

The temporary installation directories for the dependencies can be removed:

rm -fr /tmp/proto /tmp/eigen /tmp/nsync

xdrfile is a lib that read, compress and write the MD trajectories. Firstly get the source:

cd /some/workspace
wget ftp://ftp.gromacs.org/pub/contrib/xdrfile-1.1.4.tar.gz

I assume you want to install it in $xdrfile_root, then you will probably do

tar xvf xdrfile-1.1.4.tar.gz
cd xdrfile-1.1.4
./configure --prefix=$xdrfile_root
make 
make install

The DeePMD-kit was tested with compiler gcc >= 4.9.

Firstly clone the DeePMD-kit source code

cd /some/workspace
git clone https://github.com/deepmodeling/deepmd-kit.git deepmd-kit

If one downloads the .zip file from the github, then the default folder of source code would be deepmd-kit-master rather than deepmd-kit. For convenience, you may want to record the location of source to a variable, saying deepmd_source_dir by

cd deepmd-kit
deepmd_source_dir=`pwd`

Then goto the source code directory and make a build directory.

cd $deepmd_source_dir/source
mkdir build 
cd build

I assume you want to install DeePMD-kit into path $deepmd_root, then execute cmake

cmake -DXDRFILE_ROOT=$xdrfile_root -DTENSORFLOW_ROOT=$tensorflow_root -DCMAKE_INSTALL_PREFIX=$deepmd_root ..

If you are using google binary for tensorflow python interface, then you need to specify

cmake -DXDRFILE_ROOT=$xdrfile_root -DTENSORFLOW_ROOT=$tensorflow_root -DCMAKE_INSTALL_PREFIX=$deepmd_root -DTF_GOOGLE_BIN=true ..

If the cmake has executed successfully, then

make
make install

If everything works fine, you will have the following executables installed in $deepmd_root/bin

$ ls $deepmd_root/bin
dp_frz  dp_ipi  dp_mdnn  dp_test  dp_train

DeePMD-kit provide module for running MD simulation with Lammps. Now make the DeePMD-kit module for lammps.

cd $deepmd_source_dir/source/build
make lammps

DeePMD-kit will generate a module called USER-DEEPMD in the build directory. Now download your favorite Lammps code, and uncompress it (I assume that you have downloaded the tar lammps-stable.tar.gz)

cd /some/workspace
tar xf lammps-stable.tar.gz

The source code of Lammps is store in directory, for example lammps-31Mar17. Now go into the lammps code and copy the DeePMD-kit module like this

cd lammps-31Mar17/src/
cp -r $deepmd_source_dir/source/build/USER-DEEPMD .

Now build Lammps

make yes-user-deepmd
make mpi -j4

The option -j4 means using 4 processes in parallel. You may want to be use a different number according to your hardware.

If everything works fine, you will end up with an executable lmp_mpi.

The DeePMD-kit module can be removed from Lammps source code by

make no-user-deepmd

In this text, we will call the deep neural network that is used to represent the interatomic interactions (Deep Potential) the model. The typical procedure of using DeePMD-kit is

1. Prepare data
2. Train a model
3. Freeze the model
4. MD runs with the model (Native MD code or Lammps)

One needs to provide the following information to train a model: the atom type, the simulation box, the atom coordinate, the atom force, system energy and virial. A snapshot of a system that contains these information is called a frame. We use the following convention of units:

The frames of the system are stored in two formats. A raw file is a plain text file with each information item written in one file and one frame written on one line. The default files that provide box, coordinate, force, energy and virial are box.raw, coord.raw, force.raw, energy.raw and virial.raw, respectively. We recommend you use these file names. Here is an example of force.raw:

$ cat force.raw
-0.724  2.039 -0.951  0.841 -0.464  0.363
 6.737  1.554 -5.587 -2.803  0.062  2.222
-1.968 -0.163  1.020 -0.225 -0.789  0.343

This force.raw contains 3 frames with each frame having the forces of 2 atoms, thus it has 3 lines and 6 columns. Each line provides all the 3 force components of 2 atoms in 1 frame. The first three numbers are the 3 force components of the first atom, while the second three numbers are the 3 force components of the second atom. The coordinate file coord.raw is organized similarly. In box.raw, the 9 components of the box vectors should be provided on each line. In virial.raw, the 9 components of the virial tensor should be provided on each line. The number of lines of all raw files should be identical.

We assume that the atom types do not change in all frames. It is provide by type.raw, which has one line with the types of atoms written one by one. The atom types should be integers.

The second format is the data sets of numpy binary data that are directly used by the training program. User can use the script $deepmd_source_dir/data/raw/raw_to_set.sh to convert the prepared raw files to data sets. For example, if we have raw files that contains 6000 frames,

$ ls 
box.raw  coord.raw  energy.raw  force.raw  type.raw  virial.raw
$ $deepmd_source_dir/data/raw/raw_to_set.sh 2000
nframe is 6000
nline per set is 2000
will make 3 sets
making set 0 ...
making set 1 ...
making set 2 ...
$ ls 
box.raw  coord.raw  energy.raw  force.raw  set.000  set.001  set.002  type.raw  virial.raw

It generates two sets set.000, set.001 and set.002, with each set contains 2000 frames. The last set (set.002) is used as testing set, while the rest sets (set.000 and set.001) are used as training sets. One do not need to take care the binary data files in each of the set.* directories. The path containing set.* and type.raw is called a system.

The method of training is explained in our DeePMD paper. With the source code we provide a small training dataset taken from 400 frames generated by NVT ab-initio water MD trajectory with 300 frames for training and 100 for testing. An example training parameter file is provided. One can try with the training by

$ cd $deepmd_source_dir/examples/train/
$ $deepmd_root/bin/dp_train water.json

$deepmd_root/bin/dp_train is the training program, and water.json is the json format parameter file that controls the training. The components of the water.json are

{
    "_comment": " model parameters",
    "use_smooth":	false,
    "sel_a":		[16, 32],
    "sel_r":		[30, 60],
    "rcut":		6.00,
    "axis_rule":	[0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 1, 0],
    "_comment":	" default rule: []",
    "_comment":	" user defined rule: for each type provides two axes, ",
    "_comment":	"                    for each axis: (a_or_r, type, idx)",
    "_comment":	"                    if type < 0, exclude type -(type+1)",
    "_comment": "                    for water (O:0, H:1) it can be",
    "_comment": "                    [0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 1, 0]",
    "n_neuron":		[240, 120, 60, 30, 10],

    "_comment": " training controls",
    "systems":		["../data/water/"],
    "set_prefix":	"set",    
    "stop_batch":	1000000,
    "batch_size":	4,
    "start_lr":		0.001,
    "decay_steps":	5000,
    "decay_rate":	0.95,

    "start_pref_e":	0.02,
    "limit_pref_e":	8,
    "start_pref_f":	1000,
    "limit_pref_f":	1,
    "start_pref_v":	0,
    "limit_pref_v":	0,

    "seed":		1,

    "_comment": " display and restart",
    "_comment": " frequencies counted in batch",
    "disp_file":	"lcurve.out",
    "disp_freq":	100,
    "numb_test":	100,
    "save_freq":	100,
    "save_ckpt":	"model.ckpt",
    "load_ckpt":	"model.ckpt",
    "disp_training":	true,
    "time_training":	true,

    "_comment":		"that's all"
}

The option rcut is the cut-off radius for neighbor searching. The sel_a and sel_r are the maximum selected numbers of fully-local-coordinate and radial-only-coordinate atoms from the neighbor list, respectively. sel_a + sel_r should larger than the maximum possible number of neighbors in the cut-off radius. sel_a and sel_r are vectors, the length of the vectors are same as the number of atom types in the system. sel_a[i] and sel_r[i] denote the selected number of neighbors of type i.

The option axis_rule specifies how to make the axis for the local coordinate of each atom. For each atom type, 6 integers should be provided. The first three for the first axis, while the last three for the second axis. Within the three integers, the first one specifies if the axis atom is fully-local-coordinated (0) or radial-only-coordinated (1). The second integer specifies the type of the axis atom. If this number is less than 0, saying t < 0, then this axis exclude atom of type -(t+1). If the third integer is, saying s, then the axis atom is the sth nearest neighbor satisfying the previous two conditions.

The option n_neuron is an integer vector that determines the shape the neural network. The size of the vector is identical to the number of hidden layers of the network. From left to right the members denotes the size of each hidden layers from input end to the output end, respectively.

The option systems provide location of the systems (path to set.* and type.raw). It is a vector, thus DeePMD-kit allows you provide multiple systems. DeePMD-kit will train the model with the systems in the vector one by one in a cyclic manner.

The option batch_size specifies the number of frames in each batch. The option stop_batch specifies the total number of batches will be used in the training. The option start_lr, decay_rate and decay_steps specify how the learning rate changes. For example, the tth batch will be trained with learning rate:

The options start_pref_e, limit_pref_e, start_pref_f, limit_pref_f, start_pref_v and limit_pref_v determine how the prefactors of energy error, force error and virial error changes in the loss function (see the appendix of the DeePMD paper for details). Taking the prefactor of force error for example, the prefactor at batch t is

Since we do not have virial data, the virial prefactors start_pref_v and limit_pref_v are set to 0.

The option seed specifies the random seed for neural network initialization. If not provided, the seed will be initialized with None.

During the training, the error of the model is tested every disp_freq batches with numb_test frames from the last set in the systems directory on the fly, and the results are output to disp_file.

Checkpoints will be written to files with prefix save_ckpt every save_freq batches. If restart is set to true, then the training will start from the checkpoint named load_ckpt, rather than from scratch.

Several command line options can be passed to dp_train, this can be checked with

$ $deepmd_root/bin/dp_train --help

An explanation will be provided

positional arguments:
  INPUT                 the input json database

optional arguments:
  -h, --help            show this help message and exit
  -t INTER_THREADS, --inter-threads INTER_THREADS
                        With default value 0. Setting the "inter_op_parallelism_threads" key for the tensorflow, the "intra_op_parallelism_threads" will be set by the env variable OMP_NUM_THREADS
  --init-model INIT_MODEL
                        Initialize a model by the provided checkpoint
  --restart RESTART     Restart the training from the provided checkpoint

The keys intra_op_parallelism_threads and inter_op_parallelism_threads are Tensorflow configurations for multithreading, which are explained here. Skipping -t and OMP_NUM_THREADS leads to the default setting of these keys in the Tensorflow.

--init-model model.ckpt, for example, initializes the model training with an existing model that is stored in the checkpoint model.ckpt, the network architectures should match.

--restart model.ckpt, continues the training from the checkpoint model.ckpt.

The smooth version of DeePMD can be trained by the DeePMD-kit. An example training parameter file is provided. One can try with the training by

$ cd $deepmd_source_dir/examples/train/
$ $deepmd_root/bin/dp_train water_smth.json

The difference between the standard and smooth DeePMD models lies in the model parameters:

    "use_smooth":	true,
    "sel_a":		[46, 92],
    "rcut_smth":	5.80,
    "rcut":		6.00,
    "filter_neuron":	[25, 50, 100],
    "filter_resnet_dt":	false,
    "n_axis_neuron":	16,
    "n_neuron":		[240, 240, 240],
    "resnet_dt":	true,

The sel_r option is skipped by the smooth version and the model use fully-local-coordinate for all neighboring atoms. The sel_a should larger than the maximum possible number of neighbors in the cut-off radius rcut.

The descriptors will decay smoothly from rcut_smth to the cutoff radius rcut.

filter_neuron provides the size of the filter network (also called local-embedding network). If the size of the next layer is the same or twice as the previous layer, then a skip connection is build (ResNet). filter_resnet_dt tells if a timestep is used in the skip connection. By default it is false. n_axis_neuron specifies the number of axis filter, which should be much smaller than the size of the last layer of the filter network.

n_neuron specifies the fitting network. If the size of the next layer is the same as the previous layer, then a skip connection is build (ResNet). resnet_dt tells if a timestep is used in the skip connection. By default it is true.

The trained neural network is extracted from a checkpoint and dumped into a database. This process is called "freeze" a model. Typically one does

$ $deepmd_root/bin/dp_frz -o graph.pb

in the folder where the model is trained. The output database is called graph.pb.

Run an MD simulation with Lammps is simpler. In the Lammps input file, one needs to specify the pair style as follows

pair_style     deepmd graph.pb
pair_coeff     

where graph.pb is the file name of the frozen model. The pair_coeff should be left blank. It should be noted that Lammps counts atom types starting from 1, therefore, all Lammps atom type will be firstly subtracted by 1, and then passed into the DeePMD-kit engine to compute the interactions.

The reciprocal space part of the long-range interaction can be calculated by lammps command kspace_style. To use it with DeePMD-kit, one writes

pair_style	hybrid/overlay deepmd graph.pb coul/long 9.0
pair_coeff	* * deepmd
pair_coeff	* * coul/long
pair_modify	pair coul/long compute no
kspace_style	pppm 1.0e-5
kspace_modify	gewald 0.45

In this setting, the direct space part of the long-range interaction is ignored by the pair_modify command, because this part is fitted in the DeePMD model. The splitting parameter gewald is modified by the kspace_modify command.

The i-PI works in a client-server model. The i-PI provides the server for integrating the replica positions of atoms, while the DeePMD-kit provides a client named dp_ipi that computes the interactions (including energy, force and virial). The server and client communicates via the Unix domain socket or the Internet socket. The client can be started by

$ dp_ipi water.json

It is noted that multiple instances of the client is allow for computing, in parallel, the interactions of multiple replica of the path-integral MD.

water.json is the parameter file for the client dp_ipi, and an example is provided:

{
    "verbose":		false,
    "use_unix":		true,
    "port":		31415,
    "host":		"localhost",
    "graph_file":	"graph.pb",
    "coord_file":	"conf.xyz",
    "atom_type" : {
	"OW":		0, 
	"HW1":		1,
	"HW2":		1
    }
}

The option use_unix is set to true to activate the Unix domain socket, otherwise, the Internet socket is used.

The option graph_file provides the file name of the frozen model.

The dp_ipi gets the atom names from an XYZ file provided by coord_file (meanwhile ignores all coordinates in it), and translates the names to atom types by rules provided by atom_type.

DeePMD-kit provides a simple MD implementation that runs under either NVE or NVT ensemble. One needs to provide the following input files

$ ls
conf.gro  graph.pb  water.json

conf.gro is the file that provides the initial coordinates and/or velocities of all atoms in the system. It is of Gromacs gro format. Details of this format can be find in this website. It should be notice that the length unit of the gro format is nm rather than A.

graph.pb is the frozen model.

water.json is the parameter file that specifies how the MD runs. An example parameter file for water NVT simulation is provided.

{
    "conf_file":	"conf.gro",
    "conf_format":	"gro",
    "graph_file":	"graph.pb",
    "nsteps":		500000,
    "dt": 		5e-4,
    "ener_freq":	20,
    "ener_file":	"energy.out",
    "xtc_freq":		20,
    "xtc_file":		"traj.xtc",
    "trr_freq":		20,
    "trr_file":		"traj.trr",
    "print_force":	false,
    "T":		300,
    "tau_T":		0.1,
    "rand_seed":	2017,
    "atom_type" : {
	"OW":		0, 
	"HW1":		1,
	"HW2":		1
    },
    "atom_mass" : {
	"OW":		16, 
	"HW1":		1,
	"HW2":		1
    }
}

The options conf_file, conf_format and graph_file are self-explanatory. It should be noticed, again, the length unit is nm in the gro format file.

The option nsteps specifies the number of time steps of the MD simulation. The option dt specifies the timestep of the simulation.

The options ener_file and ener_freq specify the energy output file and frequency.

The options xtc_file, xtc_freq, trr_file and trr_freq are similar options that specify the output files and frequencies of the xtc and trr trajectory, respectively. When the frequencies are set to 0, the corresponding file will not be output. The instructions of the xtc and trr formats can be found in xtc manual and trr manual. It is noticed that the length unit in the xtc and trr files is nm.

If the option print_force is set to true, then the atomic force will be output.

The option T specifies the temperature of the simulation, and the option tau_T specifies the timescale of the thermostat. We implement the Langevin thermostat for the NVT simulation. rand_seed set the random seed of the random generator in the thermostat.

The atom_type set the type for the atoms in the system. The names of the atoms are those provided in the conf_file file. The atom_mass set the mass for the atoms. Again, the name of the atoms are those provided in the conf_file.

The code is organized as follows:

• data/raw: tools manipulating the raw data files.
• examples: example json parameter files.
• source/3rdparty: third-party packages used by DeePMD-kit.
• source/cmake: cmake scripts for building.
• source/ipi: source code of i-PI client.
• source/lib: source code of DeePMD-kit library.
• source/lmp: source code of Lammps module.
• source/md: source code of native MD.
• source/op: tensorflow op implementation. working with library.
• source/scripts: Python script for model freezing.
• source/train: Python modules and scripts for training and testing.

The project DeePMD-kit is licensed under GNU LGPLv3.0