MCMC Sampler¶

Since Lenstool v8, the ZeroMQ library offers a mechanism to make Python communicate with Lenstool. This lightweight messaging inter-process library opens a communication channel to transfer several types of information, such as computing a likelihood given input parameters values between 0 and 1, write a line in the burning or in the bayes.dat file, get the number of free parameters in a Lenstool model.

In this implementation, Lenstool is considered as a server and the Python script is a client. There can be several Lenstool servers running in parallel, but only one Python client.

In the current implementation, the EMCEE Ensemble sampler has been chosen to test the method. However, any MCMC sampler adapted to communicate with Lenstool using the ZMQ library should work.

EMCEE single process¶

In the single process mode, Lenstool is started as a background process on the command line, and the Python EMCEE script is started afterwards. The communication url between Lenstool and the Python script must match.

In Lenstool, the URL is specified in the runmode section with the inverse keyword

    inverse 6 ipc:///tmp/lenstool_azerty

In the EMCEE script, the URL is given as an argument to the initialisation function of the sm_client class

import sm_client
sm = sm_client.client("ipc:///tmp/lenstool_azerty")

Here is a shell script to launch the whole thing

lenstool test.par -n &
python emcee_EnsembleSampler.py

and the following should appear on the console

$> lenstool nfw.par -n &
[1] 10517
$> You are running openMP version of lenstool with 24 threads
You can change number of threads by set environment variable OMP_NUM_THREADS
INFO: prepare arclet 3470/3470
INFO: Waiting for clients here url = ipc:///tmp/sm_lenstool.test

$> python emcee_EnsembleSampler.py
INFO: Connected to clients here url = ipc:///tmp/sm_lenstool.test
INFO: Number of free parameters :  6
Choose initial values for the 100 walkers
burning:  100
Mean acceptance fraction: 0.02
Autocorrelation time: 0.84 steps
min/max prob= -6499.184082952672 -1440.5218257264935
Ask server to write in burnin.dat

and you can check using the top command, that everything runs fine

  PID USER      PR  NI    VIRT    RES    SHR S  %CPU %MEM     TIME+ COMMAND                                                                         
10517 ejullo    20   0 1514164  13416   3252 R 899,0  0,0  30:44.08 lenstool nfw.par -n   
10528 ejullo    20   0 1290580  65672  23448 S  45,6  0,1   0:12.61 python emcee_EnsembleSampler.py 

Be aware that in EMCEE the switch between burnin and sampling phases is arbitrary. Therefore, the chains in bayes.dat might not have fully converged. You should check with care their behaviour before estimating error bars and contours.

EMCEE pool process¶