Whenever someone first begins taking any substantial steps into using Python for math/science/engineering, they face a confounding situation relating the packages SciPy, NumPy, Matplotlib, PyLab. Everything I say here is documented in detail on SciPy.org, or other places that further digging can uncover. If you want a more in depth understanding, that's the place to go. My goal here is to clarify what's going on when you look at example programs in terms of when to use what package, or more precisely, what steps you can take to determine what package/s to use for your purposes.
First, let me point out that many of the codes you will see on the web will likely have redundant calls, or will not suite your style or needs. In forums where feedback is enabled, you can often hunt for the best version. Nothing should be taken as gospel while you develop an understanding. If something works, well, it works. It may not be the best way for you in the long run, but it is a way that will work in the short term.
Instead of an accurate history of development, I'll piece together a reasonable assessment of the relative roles.
NumPy, the Numerical Python package, forms much of the underlying numerical foundation that everything else here relies on. If you don't do a lot of sophisticated math, this might just be enough for you. The odds are against it simply by the fact you are taking time to read this blog. NumPy adds N-dimensional array capabilities and some linear algebra, Fourier analysis, and random number capabilities. It also adds some tools for connecting to compiled languages and applying functions to arrays (broadcasting) similar to Matlab®. For a quick overview, a please look at a table comparing NumPy commands to equivalent Matlab® commands. A brief summary: NumPy allows Python to do some substantial math.
import numpy as np dir(np) # This will provide a seemingly overwhelming # list of the functions of NumPy help(np) # Help is always available for NumPy
Note that when we import a package in this form (import numpy as np) we are requiring our code to call functions with the lead np.. For instance
import numpy as np np.log2(8)
returns \(3\), as \(2^3=8\). For this form, log2(3) will return an error, as log2 is unknown to Python without the np..
We could have, less advisedly, used from numpy import *, which means to bring everything in to the top level name space. The name space is simply the available named commands. To illustrate this, quit Python, rerun it, and try
from numpy import * log2(8)
which returns \(3\). However, np.log2(3) will no longer work.
Why does Python do all of this (as compared to Matlab®)? It keeps the list of available variable names under control, something that can falter when more and more functions are defined.
Note that we can also
import numpy as np help(np.log2) # a function, help(np.ma) # or a subpackage within numpy
from numpy import ma # Import only the masked array package
The best practice when coding is to import only what you need. It enables codes to start up faster (less time loading libraries) and leaves names available for your use as variable names. Further, using import numpy as np keeps those NumPy commands separate from your workspace. An example that happens in other languages is that i = sqrt(-1) by definition. However, you can write over this with i=1, perhaps in a loop. Elsewhere, when you go to use i in an expression, it no longer carries the expected value of sqrt(-1). The more complicated the code, the more likely this becomes.
When working interactively, it is often fine to take a softer (expedient) approach and simply from np import *.
SciPy, Scientific Python, adds substantial capabilities to NumPy. For Matlab® users, it's very much like many of the core toolboxes. If you import scipy as sp, you have also by default imported the core capabilities of NumPy, making importing NumPy almost redundant. The
import scipy as sp dir(sp) # This will provide a seemingly overwhelming # list of the functions of NumPy help(sp) # Help is always available for NumPy, help(sp.log2) # Some NumPy functions are repeated in # availability from NumPy. This should be # redundant, but there seems to be some # exceptions. When in doubt, try it both ways.
Be aware that if you import scipy as sp, but don't also import numpy as np, you will have to use sp.function to call function from numpy. I've recently come to the conclusion for my needs that using import numpy as np is pointless, and that simply importing SciPy's and accessing all of NumPy's capabilities from SciPy is simpler and more consistent.
This is the most popular plotting (data visualization) routine package for Python. Since this is an introductory document, I won't do anything other than suggest you start here. If it doesn't fit your needs, you can use a more advanced (and possibly harder to use) package. I'm not advocating against them... unless you are new. Start here. Many users with Matlab experience will be capable of taking the step to using import matplotlib.pyplot as plt as the commands are sufficiently similar to Matlab® that the challenges are modest with Google at hand. One unfortunate oddity is that the author of Matplotlib, John Hunter, wrote a substantial amount of computational tools and embedded them within Matplotlib. This can provide a bonus, as you don't need to then import them from elsewhere. However, they are typically inferior to SciPy versions, and so best to be avoided in substantial circumstances. The debate about what to do about this overlap continues, and although I never knew Dr. Hunter, I sense that he was a great enough person to recognize that at this point they should likely be deprecated and eventually removed. That is actually a hard thing to do, as many codes already rely on these functions, and it would take a lot of re-writing/corrections/testing to extricate from everywhere, so I empathize with those working towards this.
PyLab is actually embedded inside Matplotlib and provides a Matlab®-like experience for the user. It imports portions of Matplotlib and NumPy. Many examples on the web use it as a simpler Matlab®-like experience, but it is not recommended anymore as it doesn't nurture understanding of Python itself, thus leaving you in a limited environment. Of course, you aren't forced to, but it is much like a paper box. It is not a bad place to start, and can simplify your learning curve, but you may eventually decide that you were better off not using it. It is a matter of philosophy, and I'll leave that debate to others. The thing to be aware of when using PyLab is that it is importing many functions directly into the namespace. If you later decide you want to code with a cleaner namespace, you are likely to need to change your function calls from, for example, linspace to sp.linspace.
My time is running out, so my conclusion will have to wait. However, hopefully this gives you some idea of how these pieces fit together, and now you can understand a bit of how these code snippets you see on the internet behave.
This can be an overwhelming amount of material, so I suggest skimming for awareness (and an occasional refresher), then referencing as needed. Sample codes/notebooks are certainly an easier way to get started than trying to write code from scratch.
Commentscomments powered by Disqus