PolyChord v 1.20.2
Will Handley, Mike Hobson & Anthony Lasenby
arXiv:1502.01856
arXiv:1506.00171
Latest version Released Apr 2020
Users are required to accept the licence agreement given in LICENCE file. PolyChord is free for academic usage
Users are also required to cite the PolyChord papers:
- arXiv:1502.01856
- arXiv:1506.00171
in their publications.
For Python users in a hurry:
pip install git+https://github.com/PolyChord/PolyChordLite@master
wget https://raw.githubusercontent.com/PolyChord/PolyChordLite/master/run_pypolychord.py
python run_pypolychord.py
You can then modify the file run_pypolychord.py to your needs. If you have mpi compilers available, this version can be run in parallel with mpi.
You should make sure that you have gfortran (or equivalent) fortran compilers installed.
If any of the above steps fail (this can in general happen for certain Mac OSX versions), then try installing without pip:
git clone https://github.com/PolyChord/PolyChordLite.git
cd PolyChordLite
python setup.py install
If you do not have sudo access/virtual environments/anaconda, then appending --user to the install command may be necessary.
We recommend the tool anesthetic for post-processing your nested sampling runs. A plot gallery can be found here
https://github.com/williamjameshandley/anesthetic
The code is MPI compatible with openMPI. To disable the MPI parallelization, set MPI=0 in ./Makefile, or compile with
make <target> MPI=0
PolyChord requires no additional libraries to run in linear mode To run with MPI it requires the openMPI library
PolyChord compiles with both gfortran and intel compilers.
Compiler type is chosen in the Makefile with the COMPILER_TYPE flag;
set COMPILER_TYPE = gnu for gfortran compilers (free)
set COMPILER_TYPE = intel for intel compilers (proprietary, much faster)
First, try a couple of quick examples:
- 20 dimensional Gaussian
Run the commands:
$ make gaussian $ ./bin/gaussian ini/gaussian.ini
- Rastrigin
Run the commands:
$ make rastrigin $ ./bin/rastrigin ini/rastrigin.ini
This runs the rastrigin 'bunch of grapes' loglikelihood.
In general, binary executables are stored in the directory ./bin, and ini files are stored in the directory ./ini.
You can create new likelihoods by modelling them on the ones in likelihoods/examples, and triggering them with their own ini files
Alternatively you can take a more "MultiNest" like approach, and manually generate the prior transformations. PolyChord's settings are then modified in the driver files src/drivers.
You should place your likelihood code in the function loglikelihood and your prior code in the function prior, contained in:
./likelihoods/fortran/likelihood.f90
Any setup required (such as reading in input files) should be conducted in the function setup_loglikelihood. In most cases, this will likely just be a call to your own pre-written library.
You should then alter the polychord run-time settings within the driver file:
./src/drivers/polychord_fortran.f90
Your code can be compiled and run with the commands:
$ make polychord_fortran $ ./bin/polychord_fortran
You should place your likelihood code in the function loglikelihood, contained in
./likelihoods/CC/CC_likelihood.cpp
Any setup required (such as reading in input files) should be conducted in the function setup_loglikelihood. In most cases, this will likely just be a call to your own pre-written library.
You should then alter the polychord run-time settings within the driver file:
./src/drivers/polychord_CC.cpp
or use the ini file version:
./likelihoods/CC_ini/CC_ini_likelihood.cpp ./src/drivers/polychord_CC_ini.cpp
Your code can be compiled and run with the commands:
$ make polychord_CC $ ./bin/polychord_CC
or
$ make polychord_CC_ini $ ./bin/polychord_CC_ini ini/gaussian_CC.ini
If you have an additional suggestions to make the c++ wrapper more easy to use, please email Will (wh260@mrao.cam.ac.uk).
Being python, this interface is the most self-explanatory. You can install direct from the git repository using:
pip install https://github.com/PolyChord/PolyChordLite/archive/master.zip
(N.B. PyPi coming soon) or you can install locally with the command:
git clone https://github.com/PolyChord/PolyChordLite.git
cd PolyChordLite
pip install . --user
This has the advantage of using intel compilers if you have them (e.g. on a HPC machine). You may wish to consider installing pypolychord in a virtual environment <https://packaging.python.org/guides/installing-using-pip-and-virtual-environments>, in which case you don't need the --user argument.
Once installed, you can then import pypolychord from anywhere with the lines:
import pypolychord
and check that it's working by running:
$ python run_pypolychord.py
or in MPI:
$ mpirun -np 4 python run_pypolychord.py
If so, the rest of the interface is relatively painless. Follow the example in run_pypolychord.py, and consult the docstring if you need help:
import pypolychord
from pypolychord.settings import PolyChordSettings
help(pypolychord.run_polychord)
help(PolyChordSettings)
There is also a demo python notebook.
PolyChord produces several output files depending on which settings are chosen
Run time statistics
Files for resuming a stopped run. Semi-human readable. This is produced if settings%write_resume=.true. This is used if settings%read_resume=.true.
File containing weighted posterior samples. Compatable with the format required by getdist package which is part of the CosmoMC package. Contains ndims+nderived+2 columns:
weight -2*loglike <params> <derived params>
Refer to the following website in order to download or get more information about getdist: http://cosmologist.info/cosmomc/readme.html#Analysing
If settings%cluster_posteriors=.true. there are additional cluster files in clusters/[root]_<integer>.txt
As above, but the posterior points are equally weighted. This is better for 'eyeballing' the posterior, and provides a natural ~4 fold compression of the .txt file.
Live points in the physical space. This is produced if settings%write_phys_live=.true. This file contains ndims+nderived+1 columns, indicating the physical parameters, derived parameters and the log-likelihood. This is useful for monitoring a run as it progresses.
Points that have been killed off. This is produced if settings%write_dead=.true. This file contains ndims+nderived+1 columns, indicating the loglikelihood, physical parameters, derived parameters and the log-likelihood. This is useful for monitoring a run as it progresses, and for performing alternative calculations and checks on evidence and posterior computations
Parameter names file for compatibility with getdist
These can be used to reconstruct a full nested sampling run, as well as simulate dynamic nested sampling. The format & contents of these two files are as follows: They have has ndims+nderived+2 columns. The first ndims+nderived columns are the ndim parameter values along with the nderived additional parameters that are being passed by the likelihood routine for PolyChord to save along with the ndims parameters. The ndims+nderived+1 column is the log-likelihood value. The ndims+nderived+2 column is the log-likelihood value that the point was born at. They are is identical to the [root]_phys_live.txt and [root]_dead.txt file, except for an additional column including the birth contours
Visualization of PolyChord Output:
[root].txt file created by PolyChord is compatable with the format required by getdist package which is part of the CosmoMC package. Refer to the following website in order to download or get more information about getdist: http://getdist.readthedocs.org/en/latest/
Common Problems & FAQs:
1 Output files ([root].txt & [root]_equal_weights.dat) files have very few (of order tens) points.
These files only become populated as the algorithm approaches the peak(s) of the posterior. Wait for the run to be closer to finishing.
2 MPI doesn't help
- Currently, the MPI parallelisation will only increase speed for 'slow' likelihoods, i.e. likelihoods where the slice sampling step is the dominant computational cost (compared to the organisation of live points and clustering steps).
- Parallelisation is only effective up to ncores~O(nlive).
Most issues are usually one associated with an out-of-date MPI library or fortran compiler. Ideally you should be using:
- gfortran 4.8 or ifort 14
- openMPI 1.6.5 or Intel MPI 4.1