TensorFlow for R

Surely we're better off if statisticians and software engineers work together. That we're worse off when the two groups come together doesn't seem probable ...

This is a key point, well said. However, we need more people who are proficient in both "worlds" or at least, as another user points out below, more people willing to collaborate.

As long as the model is well described I'm all for more statistician experimenting with deep learning. I don't mind if initial code isn't great from a software perspective. If it is needed in production someone else would have to reimplement it.

>Great news for R users who largely have been left behind by the deep learning community.

No one often mentions that TensorFlow lacks native windows support.

    >  code that is not very elegant nor easy to understand

That was my impression when I started learning R while learning statistics - again, but this time with an actual interest in the subject and not just for the grade :)

On the other hand, the whole R-learning experience, which was aided by having to do practical "statistical" things in the language and working with actual data, gave me a new perspective on programming, a much more data-oriented perspective if I may call it that.

Before (learning R and statistics) I was interested in the code and what it could do, afterwards (but only for those kinds of applications) I was interested in the data and what I could do with it, the "how" was secondary. From that perspective I understood a lot of the (IMHO) "hacky" nature of R and some of its libraries.

I recently started writing some R code and found it much more elegant than previous code. R has a bunch of functions for handling data and statistics related stuff, that really reduces complexity. But also it has a lot nice iteration functions that make it possible to do a lot of stuff with a single line. I only needed to use a for loop once. I never needed to do anything complex with data structures or nested loops or branching code, everything seemed to reduce to a few functions. It's a really elegant language, at least for its specialized domain.

Unfortunately a lot of R's internal standard library code isn't much better, rife with assumptions of minuscule data, gratuitous copying of input data many times over and pointlessly returning them in results, glacial I/O, grossly inefficient implementations even at the C level, etc. Using R in a production environment is a slow journey through reimplementing R itself.

"Leap" is quite the exaggeration IMO, especially talking about ML. This is improvement wrt to neural networks which had some support in R already (mxnet). In ML, in general, R is quite a bit ahead of the competition.

You have been able to use dplyr on spark dataframes for a while now (at least a year).

I imagine that they did this in order to be able to use more of (or all of) the TensorFlow functionality available in the Python API without having to rewrite a huge amount of code in R.

There are Operations in TensorFlow that don't invoke the C++ layer directly. Instead, they are composed of other Operations in Python, meaning that binding directly to the C++ layer would require finding and rewriting that code.

It still gets sent down to the C++ layer at the end of it, though the stack traces might end up looking pretty ridiculous.