Scientific Python Overview

Python is a widely-used programming language that has merged as a popular environment to perform data analysis and visualization. This popularity is based around a large user community that develops and maintains a large number of free, open-source packages for many different applications in the worlds of science, finance, business, system programming, and more.

Although Python is provided “with batteries included” (it has a comprehensive standard library with built-in tools for many tasks including file I/O and processing, datetime handling, command line tool integration, and much more), its use for scientific visualization and analysis is built on a suite of community-built tools:

State of the Stack, Jake Vanderplas, 2015

State of the “Scientific Python Stack” circa 2015 (courtesy Jake VanderPlas)

n-dimensional Arrays

Out of the box, Python does not have a high-performance multi-dimensional array object. This sort of data structure is critical for numerical analysis. In the Python ecosystem, multi-dimensional arrays are provided by the NumPy library, which contains a high-performance array object and related linear algebra routines. The array “interface” provided by NumPy supports vector programming and is very general, underpinning many of the numerical tools widely used in the Python world. For instance, the companion SciPy package implements many useful functions foor optimization, statistics, interpolation, image processing, and spatial mathematics. Many of these functions work with NumPy arrays having arbitrary size, dimensions, and data types.


The version of Python you will most likely use has been implemented in C, and is designed in such a way that it is very easy to work with other compiled codebases, especially those written in C/C++ or Fortran. In fact, libraries like NumPy heavily rely on optimized, compiled codes in these languages, which greatly improves their performance, reliability, and speed.

Another library which builds on the NumPy array interface is pandas, which extends the array with labeling and a powerful engine for the transformation and analysis of structured datasets. The data structures provided by pandas - particularly the DataFrame - greatly simplify timeseries analysis, split-apply-combine workflows, and other common research processing tasks. The xarray package extends pandas by providing addtional data structures to handle n-dimensional labeled arrays (such as those contained in a NetCDF file). Most importantly, the array and labeling semantics in NumPy, pandas, and xarray are similar if not identical in the majority of cases, and because each package sequentially builds on the others, they can all be used within the same analysis context and data can easily be shuttled back-and-forth to whatever format works best for a given task.


The core visualization library in the Python world is matplotlib. Although it originated as an emulation of the graphics capabilities of MATLAB, matplotlib has grown into the defacto base layer for 2D graphics in Python. Matplotlib provides fine-grained control of graphics, and works natively with both base Python objects and NumPy array derived-types (including pandas Series and DataFrames and xarray DataArrays). In fact, both pandas and xarray provide shim layers to help automate plotting numerical data through matplotlib.

Many libraries extend the core features of matplotlib. For instance, seaborn implements many useful statistical visualizations and leans particularly heavily on pandas to help organize data for plotting. To plot geographical information, one can use the cartopy library, which itself wraps several open-source cartographic libraries and has support for geographical projections, shapefiles, and more. Users coming from R who love ggplot2 should be aware of several upcoming grammar of graphics implementations based on matplotlib, including plotnine, ggplot, and altair.

Domain-Specific Toolkits

Python users can be found throughout the ranks of researchers in the natural sciences, and many contribute specialized toolkits to help with their own niche applications. Here is a short summary of tools that can be useful for different research tasks in Python:

A toolkit implement a wide variety of algorithms for un/supervised machine learning tasks, including regressions, clustering, manifold learning, principal components, density estimation, and much more. It also provides many useful tools to help build “pipelines” for managing modeling tasks such as data processing/normalization, feature engineering, cross-validation, fitting, and prediction.
A module for fitting and estimating many different types of statistical models as well as performing hypothesis testing and exploratory data analysis. It features tools for fitting generalized linear models, survival analyses, and multi-variate statistics. Furthermore, it implements formula-based regression specification similar to R which natively works with pandas data structures.
An image processing library featuring many common operations including convolutional mapping, filtering, edge detection, and image segmentation.
A Python package for reprojecting earth observing satellite data, capable of handling both swath data from polar-orbitting satellites and gridded data from geostationary satellites.
A spatial analysis library which extends Python to work as a fully-featured GIS environmental comparable to commercial software such as ArcGIS.
A full-featured computer algebra system (CAS) similar to Mathematica or Maple. SymPy powers an additional ecosystem of domain-specific tools used in pure mathematics research and which have many applications in physics.
A toolkit for Bayesian statistical modeling and probabilistic programming, including a suite of Markov chain Monte Carlo fitting algorithms.


Python is slower than statically-typed, compiled languages such as C/C++ and Fortran. However, it doesn’t have to be slow. Vectorized programming through NumPy and pandas can dramatically increase the performance of Python in executing numerical analyses and calculations. However, in applications where vectorization is non-trivial or inappropriate, Python’s performance can be dramatically improved by using one of several different approaches.

First, an optimising compiler called Cython is available to compile your code into high-performance, efficient C kernels. Cython will work on your normal Python code with few modifications, and can often times increase its performance by 1-2x. However, by incorporating a special set of static typing directives into your code (similar to what you’d do in Fortran by declaring variable types), Cython can go a step further and yield much more significant performance improvements, often achieving speed-ups to within 50-90% of comparable C-code. It also trivializes the task of wrapping legacy code from C/C++ or Fortran applications.

Alternatively, one can use a just-in-time (JIT) compiler to compile code on-the-fly. One approach in the Python world implementing a JIT is the PyPy project, which is an alternative implementation of Python itself. A drawback to PyPy is that it does not totally support all of the numerical libraries in the scientific Python stack. Instead, one can use Numba to target specific, expensive functions or subroutines in a codebase. Numba-compiled functions can target multi-core or GPU architectures when available.

Niche optimization tools also exist in the Python world. For instance, the PyCUDA package helps to glue together Python with high-performance GPGPU routines written in C/CUDA. Meta-programming libraries for working with tensor mathematics are also widely used, including theano and TensorFlow, which themselves can bootstrap GPGPU kernels for an added performance boost.

Parallel Computing

The most recent versions of Python include modules and infrastructure for asynchronous and coroutine programming in the standard library. The programming model used in this paradigm (using “futures” or other “delayed” objects representing a contract for future results from calculations) is extended by several libraries in the Python ecosystem, although it may not be familiar to most scientific researchers.

As an example, the joblib library implements a very lightweight, easy-to-use set of tools in this programming model. Joblib strives to let you make simple modifications to your single-threaded code to achieve parallelism when and only where it becomes necessary, usually by re-factoring an inner loop of a program to a stand-alone kernel and parallelizing its application.

An alternative approach is dask, which provides a similar API but works natively with array-like and DataFrame-like structures from NumPy and pandas. Dask abstracts the parallel programming model one step further, tracking a graph representing your computations; it optimizes this graph before any calculations are actually performed, which allows it to optimize the amount of data held in memory at any given time, or scale to arbitrary resources as they become available. An advantage of dask is its ability to work with virtually any abritrary data processing task or pipeline.

Traditional MPI tools also exist in the Python ecosystem, although these tend to be very low-level and “un-Pythonic.” If you’re writing C or Fortran in Python, then you’re probably doing something wrong.