DNetPRO in Python

Up to now we have focused on the algorithm performances, leaving out the usability of the DNetPRO algorithm for the (research) community. Despite the C++ is one of the most efficient and older programming language¹, the number of Python-users is growing in the last few years. Python is becoming leader in scientific research publications and the large part of Machine Learning analyses are performed using Python libraries (in particular scikit-learn library). So, we have to reach a compromise between performances and usability of the new codes and a reasonable solution is given by the Cython [behnel2010cython] framework.

Cython “language”² allows an easy interface between C++ codes and Python language. With a relatively simple wrapping of the C++ functions, they can be used inside a pure Python code, preserving as much as possible the computational performances of a pure C++ version. In this way, we have written a simple Python object which performs the full set of DNetPRO steps and, moreover, which is compatible with functions provided by other machine learning libraries.

With this purposes we have chosen to operate a “double wrap” of the C++ functions to separate as much as possible the C++ components from Python³. The Python object was written considering a full compatibility with the scikit-learn library to allow the use of the DNetPRO feature selection as an alternative component of other Machine Learning methods.

next »

1. Still in common use in scientific research groups. ↩

2. It is not a real programming language since it is based on Python. However it has its own syntax and keywords which are different either from Python either from C++. In the end it needs a compiler to run and it is certainly different from Python. ↩

3. Cython wraps are very powerful tools for C++ integration into Python code but, by experience, they are difficult to manage by pure-Python-users. A simple workaround is to perform a first wrap of the C++ functions inside a Cython object, adding a second wrap of it into a pure-Python class. This two-steps wrap certainly gets worse the computational performances, but it allows a complete separation between the compiled part of the code (Cython) and the interpreted (Python) one. Moreover, we can leave back all the checks on input parameters of the C++ function since they can be performed at run time by the Python wrap. ↩