niujiashu / polychord Goto Github PK

PolyChord v 1.9
Will Handley, Mike Hobson & Anthony Lasenby
[email protected]
arXiv:1502.01856
arXiv:1506.00171
Released April 2016

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
PolyChord Licence
=================

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.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
MPI Support
===========

The code is MPI compatible with openMPI. To disable the MPI parallelization, 
set MPI= in ./Makefile, or compile with
make MPI= <target>


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Additional Libraries  
====================

PolyChord requires no additional libraries to run in linear mode
To run with MPI it requires the openMPI library


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Compilers
=========

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)


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Running PolyChord
=================

Examples
--------
First, try a couple of quick examples:

1) 20 dimensional Gaussian

run the commands:
$  make gaussian
$  ./bin/gaussian ini/gaussian.ini

2) 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.


Fortran likelihoods
-------------------
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



C++/C likelihoods
-----------------
You should place your likelihood code in the function cpp_loglikelihood,
contained in 

./likelihoods/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_fortran.cpp

Your code can be compiled and run with the commands:
$  make polychord_CC
$  ./bin/polychord_CC 

If you have an additional suggestions to make the c++ wrapper more easy to use, 
please email Will ([email protected]).



Python likelihoods (PyPolyChord)
--------------------------------
Being python, this interface is much more self-explanatory. You need to compile
the library with
$  make PyPolyChord

You then need to point python to the library folder ./lib by adding it to
LD_LIBRARY PATH. For example, in bash:
$  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$PWD/lib

You may then need to also amend LD_PRELOAD with your mpi library. This is only
necessary on some systems, and only then if you wish to run it with MPI.  Find
out where your mpi library is kept, and then add it to LD_PRELOAD. For
examples, in bash on Ubuntu 14.04:
$  export LD_PRELOAD=/usr/lib/openmpi/lib/libmpi.so:$LD_PRELOAD


You can then check that this works 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.PyPolyChord as PPC
>   help(PPC.run_nested_sampling)



~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Output files 
=============
PolyChord produces several output files depending on which settings
are chosen


[root].stats
------------
Run time statistics

[root].resume
-------------
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.

[root].txt
----------
File containing weighted posterior samples. Compatable with the format
required by getdist package which is part of the CosmoMC package.
Contains npars+ndims+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 

[root]_equal_weights.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. 


[root]_phys_live.txt
--------------------
Live points in the physical space. This is produced if
settings%write_phys_live=.true.
This file contains npars+ndims+1 columns, indicating the physical
parameters, derived parameters and the log-likelihood. This is useful
for monitoring a run as it progresses. 

[root]_dead.txt
--------------------
Points that have been killed off. This is produced if
settings%write_dead=.true.
This file contains npars+ndims+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

[root].paramnames
------------
Parameter names file for compatibility with getdist


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
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:


Run time Issues
===============
1. PolyChord crashes after segmentation fault.

Try increasing the stack size:
Linux:    ulimit -s unlimited 
OSX:      ulimit -s hard 
& resume your job.
The slice sampling & clustering steps use a recursive procedure. The
default memory allocated to recursive procedures is embarassingly
small (to guard against memory leaks).


2. 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.


3. 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).


Compilation Issues
==================
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