Lofkin

Aside from the written in python making it more natural and extensible and amenable to messing around with the models, pymc 3 can sample directly from discrete parameters and STAN cannot. Pymc3 has more sampling routines.

More good python resources: http://web.bryant.edu/~bblais/statistical-inference-for-ever...

Harvard Data science class, in python: http://cs109.github.io/2014/

Ultimately Julia combines the strengths of both and much more. It is the future IMO, not R or python.

Python has more consistent stats, time series and programming syntax (for the packages it does have), and better bayesian inference package (pymc 3 > stan).

Python is also better than R for ad hoc statistical modeling and algorithim development (you can write python code on the order of C fast with numba) , general programming, scraping, natural language processing, agent based modeling etc.

Python is also better for GIS, optimization, symbolic math and larger datasets with blaze and dask and pyspark.

R right now is a bit better for visualization, reporting and exploratory data analysis (I think this will change soon though with Bokeh and blaze) and has many more esoteric stats packages.

With statsmodels, pymc3, pandas and scikitlearn etc you can probably do 98% of everything you need in python without needing to dip into the more esoteric packages of R (with some exceptions). For everything else, you can call R packages with Rpy2. With this you get all the advantages of working in one language (not spread too thin) and the advantages that python offers while leveraging R's wealth of packages.

The latter is a bit more difficult through python, but not as hard as trying to remember the syntax of and gluing together a two language workflow.

That is why I chose python... Also I can write excel addins with python (xlwings) instead of using VBA

Why can't it be used for general computing?

It already gets C like performance reasonably predictably. Its not a tracing JIT, like pypy...it restricts the dynamism of python to be able to do type inference.

The thing is, we still don't have the beautiful julia type system, multiple dispatch, macros, better syntax for technical computing easy distribution etc

can you tldr what immoral means in this context? How did he define this and why should we accept his definition?

Julia fills this need. Also python with numba. Julia is the future, takes the best syntax from python, matlab couples that with a beautiful and extensible type system and uses this to drive JIT llvm with speed on the order of C.

toolz is pretty good at lazy fp in python : https://github.com/pytoolz/toolz

Interesting thanks for sharing. Hopefully bokeh might be able to address the visualization space with its upcoming statistical glyphs.

When you say "flexibility to use whatever I wanted to do my work", do you mean data munging and manipulation or statistical modeling packages?

Numba allows multithreaded gil free numerical code.

anaconda solves this

Have you tried data.table?

Thanks for your insight. I have some questions for your, if you don't mind.

What walls and why did you switch? Did you find the shift in syntax and mindset difficult? Have you used rpy2?

There definitely seems to be more innovation in R world, but blaze, numba and bokeh have no R equal.

Also can we have support for Rpy2 or some kind of python -r ffi please????

Numba is faster than pypy and compatible with cpython.

By the way, can you say if we will see better scripting support (python or r) for excel? :p

Sounds super cool! I like the new Microsoft :)

Is this python for data science as well, or just sticking with general programming?

Try bokeh http://bokeh.pydata.org/en/latest/

Have you tried bokeh or vispy? What about seaborn?

This is a job for blaze, out of core pandas and numpy frontend to large datasets. http://blaze.pydata.org/en/latest/

Also numba. It is a multithreaded JIT compiler for a subset of python code. Just need a decorator.

Julia and Python Numba are two fast (almost native) high level languages for multdimensional array programming.

Gotcha..

I wonder if this can help: http://nuitka.net/pages/overview.html?

Have you tried conda, the package manager by continuum analytics? It stores pre compiled instances of packages for easy install cross platform and does some virtual env stuff.

Much easier and more robust than pip:

http://www.continuum.io/blog/conda

Is numba poised to overcome these limitations at some point?

Writing backend agnostic code (CPU and GPU) jitted to fast LLVM is on the roadmap for python's Numba: https://github.com/numba/numba/issues/545

Currently supports both including multithreading. Its just he unified interface that is in the works.