A simple hack to make figures reproducible

In my previous posts I explained why reproducibility is an important topic that is probably underappreciated in the weather and climate sciences at the moment. Here, I just wanted to present a simple hack which makes figures created with the Python library matplotlib more reproducible.

The basic idea, based on Damien Irving’s inspiring paper, is to save a log file along with each figure. The file has the same name as the figure file but with the ending .log instead of .pdf or whatever format the figure is in. The log file contains information about:

• When the script was called.
• Which script was called to produce the figure and which arguments were used (using the sys.argv command).
• The git commit at the time (using the package GitPython, which you might need to install additionally). For this to work changes should be committed regularly of course.
• Which Anaconda environment was used.

Here is an example of such a log file:

Time: 2017-08-17T11:55:15

Executed command:

python verif_fc_prec.py --expid DA_REF DA_PSPv2 DA_REF_TL500 DA_PSPv2_TL500 noDA_PSPv2 --ana det --date_start 20160526000000 --date_stop 20160609000000 --composite

In directory: /home/s/S.Rasp/repositories/dwd_scripts

Git hash: ba7f57f8a55b63af0c9eb53e9231e58da3d33e79

Anaconda environment: py_env                *  /home/s/S.Rasp/anaconda2/envs/py_env

From this information it should be easy to reproduce the respective figure. The only change required is to replace the usual savefig call with a call to the function save_fig_and_log. A similar method could be used to trace intermediate steps, such as pre-processing data.

Certainly there are more elegant solutions. But hopfully this code snippet can provide a good starting point for implementing reproducibility and serve as inspiration for better solutions. As always I am very interested to hear your opinions.