Reproducibility, provenance

Scientific numerical calculations are … scientific calculations.

Hence, like any other kind of calculations, according to the basic principles of science, everyone should be able to reproduce them, reuse or modify them. Therefore, the detailed instructions leading to results or figures should be published along with them. To achieve these goals, in practice we need to be able to do simply the following things:

  • Store along with the data the version of the code used to produced them (or even the code itself!), and the configuration options of this code.

  • Keep with the figures all the instructions (i.e. the script) that have produced it.

  • We want to do that easily at no cost in human time, and hence without adding a new layer of tools (which means new things to learn, which takes time, etc.). Indeed this task is important but admittedly extremely boring for physicists!

Fortunately, python helps solve these issues easily and efficiently.

TRIQS adds very little to the standard python tools here. So this page should be viewed more as a wiki page of examples. TRIQS does not impose any framework on you, it just provides tools and lets you organize your work as you wish.

TRIQS code version

The tiny module triqs.version contains various pieces of information configured automatically at compile time

from triqs.version import *

version    # The version of the TRIQS library
release    # The release number of the library
git_hash   # The git commit used at compilation

# publishing this information may lead to a security issue....
show_machine_info()   # Hostname and login of the compilation

Saving the script in the data archive

It is actually very simple to save the script (hence the parameters) along with the data, simply by putting it in the HDFArchive, e.g.

# ... computation ...
Results = HDFArchive("solution.h5",'w')
Results["G"] = S.G # save the results

import sys, triqs.version as version
Results.create_group("log")
log = Results["log"]
log["code_version"] = version.release
log["script"] = open(sys.argv[0]).read() # read myself !

The script that is currently being executed will be copied into the file solution.h5, under the subgroup /log/script.

In a more complex situation, you may decompose your computation in several scripts, e.g.

  • A script common.py, with some common functions, classes…

  • A little one, computation1.py for each computations.

In such situation, one can simply use the inspect module of the python standard library e.g.

import common
# set parameters
# run...
# save...
# Ok, I need to save common too !
import inspect,sys, triqs.version as version
log = Results.create_group("log")
log["code_version"] = version.release
log["script"] = open(sys.argv[0]).read()
log["common"] = inspect.getsource(common) # This retrieves the source of the module in a string

From the data to the figures

Warning

TO BE WRITTEN