scipy still doesn't have a good submodule for doing manipulations with multivariate gaussians, which is a HUGE part of machine learning applications.
I'm referring to the analog of R's mvtnorm library.
Yes, you can use rpy2. Yes, you can roll your own dmvnorm (but be careful about the degenerate situations with zero eigenvalues that always arise in high dimensional problems with real data). It is significantly more of a pain to roll your own pmvnorm, because you are now talking about efficient computation of high dimensional integrals, which always seems to involve a trip to Numerical Recipes or the like (see the Genz and Miwa references when you do "?pmvnorm" in R).
The definition of "basics" depends on your domain. For the fluid dynamics and geological data that Enthought (the sponsor and primary developer of SciPy) works with, SciPy and NumPy provides all the basics.
One could argue that if you're going to call something "SciPy", it ought to cover all sorts of scientific computation...but, it is Open Source under a very open license. Contributing is very easy, and the team are very friendly to outsiders sending patches, and it's not hard to no longer be considered an "outsider". So, if you need it and SciPy doesn't provide it, why not develop it and contribute?
What I'm trying to say is that complaining about missing functions you need from an Open Source project just wastes your time and annoys the pig.
Hey, without going into details I have contributed a fair amount of code to R.
Scipy (and the whole Python numerical computation stack) has nontrivial competition out there and here's what it needs to do to compete...if its developers care about adoption, which most do.
Definitely not saying scipy == crap, just that for stats/machine learning (which is a big percentage of a lot of applications today) it is not mature.
I'm referring to the analog of R's mvtnorm library.
Yes, you can use rpy2. Yes, you can roll your own dmvnorm (but be careful about the degenerate situations with zero eigenvalues that always arise in high dimensional problems with real data). It is significantly more of a pain to roll your own pmvnorm, because you are now talking about efficient computation of high dimensional integrals, which always seems to involve a trip to Numerical Recipes or the like (see the Genz and Miwa references when you do "?pmvnorm" in R).
Basically this is a huge thing to lack.
PS: yes, if you grep the codebase, you can find this: http://www.scipy.org/doc/api_docs/SciPy.stats.kde.gaussian_k...
But you need to write wrappers because it doesn't provide a basic interface to the multivariate gaussian density.
I could go on about other parts of scipy -- it has some advanced stuff but lacks a lot of basics...