This code allows the use of models trained using TENSOAP/SOAPFAST; the key parts of this code needed to apply a symmetry-adapted Gaussian process regression (SA-GPR) model are re-implemented in FORTRAN, making them faster and more parallelizable.
I have tested this code with gfortran, version 7.4.0, and ifort, version 18.0.5. In order to compile this code, it should suffice to run the configure script in the topmost directory. This will search for the required libraries, and compile them from scratch if needed; it will then produce a Makefile, so that you can compile the code by running make in the src directory.
The command make install installs the executables and libraries. The default root folder is /usr/local; to use a different directory either run the configuration script with the command SAGPR_PREFIX=/path/to/folder configure or modify the SAGPR_PREFIX variable in the Makefile.
It should be noted that there is no guarantee this code is as optimized as it can be; further, if you already have the required libraries (BLAS, lapack) on your system these are recommended, rather than compiling your own version.
In order to use this code, the first thing needed is a model produced using TENSOAP. There are a couple of requirements:
- Power spectra should be sparsified on spherical harmonic components. This means that there will be two files, with names like
PS_fps.npyandPS_Amat.npyproviding sparsification details. - Power spectra should also be sparsified over environments. There will be a single sparsified power spectra, with a name like
PS.npy - Finally, there should be a weights file, with a name like
weights_0.npy.
To create model files, run /path/to/soapfast_fortran/bin/sagpr_convert -ps PS.npy -sf PS -w weights_0.npy -o fname, where the PS in -sf PS is the prefix to _fps.npy and _Amat.npy. Any additional (lambda-)SOAP hyperparameters are specified via command line arguments (/path/to/soapfast_fortran/bin/sagpr_convert -h gives a list of these). For example, -p -n 5 -l 3 -z 2 is for a periodic system, with nmax=5, lmax=3 and zeta=2.
This will create the file fname.mdl, a binary file containing the training power spectrum, sparsification details, weights and hyperparameters.
To use this model, sagpr_apply is called as /path/to/TENSOAP-FAST/bin/sagpr_apply -f file_name.xyz -m fname.mdl -o prediction.out, which uses the model fname.mdl to make predictions for file_name.xyz, storing them in prediction.out. Note: these models can only be for a single spherical component; a separate prediction must be made for each component.
In the examples subfolder are a few examples that show how different modes of the code work, including sockets, committee models and long-range descriptors. It is recommended that the first-time user begins with the tutorial example, which goes through the process of creating a model with the TENSOAP code, converting it to be used by TENSOAP-FAST and applying the model.
David Wilkins, Andrea Grisafi, Yair Litman
Please contact d.wilkins@qub.ac.uk with any problems.