machine_learning_example

This repository could be copied to oscer with git clone git@github.com:cc-ats/machine_learning_example.git. It would be more than welcomed to submit any bugs and feature request in the issue section. I recommend you to set up your own branch by,

git branch -M junjie
git push -u origin junjie

Raw data

This is a machine learning example, we start from the files inside raw_data:

raw_data:
coord.npy  energy.npy  force.npy

The data could be read with numpy.load, for example,

python -c '
from numpy import load
coord_data  = load("./raw_data/coord.npy")
force_data  = load("./raw_data/force.npy")
energy_data = load("./raw_data/energy.npy")

print("coord_data.shape  = ",  coord_data.shape)
print("force_data.shape  = ",  force_data.shape)
print("energy_data.shape = ", energy_data.shape)
'

we have,

coord_data.shape  =  (1000, 96, 3)
force_data.shape  =  (1000, 96, 3)
energy_data.shape =  (1000,)

all the energies are in atomic unit $E_\mathrm{h}$ , the coordinates are in $\text{\normalfont\AA}$ , and the forces are in $E_\mathrm{h}/\text{\normalfont\AA}$ . However, the working units of DeePMD-kit are, ( $1E_\mathrm{h} = 27.211386245988\mathrm{eV}$ )

Property	Unit
Time	$\mathrm{ps}$
Length	$\text{\normalfont\AA}$
Energy	$\mathrm{eV}$
Force	$\mathrm{eV}/\text{\normalfont\AA}$
Pressure	$\mathrm{Bar}$

Build system

We can use ./script/build_system.py to build the system:

python script/build_system.py . 4

Train the model

To set up the training

mkdir train
cd train
cp ../src/inp.json     .
cp ../src/train_cpu.sh .
sbatch train_cpu.sh

please make sure you have carefully read the inp.json file and the train_cpu.sh before submitting it.

Summary

You can do all the thing by,

sbatch ./script/run.sh