4 ms·

Julia frustrates me. I was in a mathematical modelling sphere, so I learned the language basics years ago and immediately fell in love (multiple dispatch, optio

by drej 9y ago

Julia frustrates me. I was in a mathematical modelling sphere, so I learned the language basics years ago and immediately fell in love (multiple dispatch, optional types, broadcasting, ...). But I couldn't get approvals at work to push it, because it was immature. Then a month later, all my code broke. I rewrote it using new APIs. It broke again. I know they reserved the right to make breaking changes up until 1.0, but they are hindering adoption because of all this. Sorry, guys, but I'll have to stick to Python, Go and (sigh) Fortran. That being said, I applaud getting good funding for a project that's actually more than beyond MVP, it has happy users/customers and it serves a purpose. That's not all that common these days. Good luck, guys.

smt88 9y ago
My girlfriend does incredibly complex modeling (she's a computational geneticist). From my limited understanding, it sounds like you may have some of the same challenges that she does. Do you think Python and Go are more suited to that domain than R or Julia?
- flavio81 9y ago
  R is very, very slow compared to all the other programming platforms (including Ruby, a nice language but with some of the slowest implementations out there). R syntax is very unlike most programming languages so jumping to other languages can feel strange. Python with Numpy, SciPy and Pandas, is the current widespread alternative to R and a good starting point. Recommended!!
  - nnm 9y ago
    For all the cases I worked on, R is not slow at all. At home, I wrote a R package to do deep learning (include most common layers, but only conv layer was implemented in C++) from scratch, and I also wrote one in python (using numpy + numba). My R version is same fast as python on the MNIST dataset (without conv layers, hence all code is in plain R or python).
    - flavio81 9y ago
      What this means is that for your purposes R is fast enough. But in the absolute sense (or in any case, relative to other programming platforms), R is one of the slowest ever. R is usually slower than Python, and Python in some cases can be 100x slower than C. Performance, assuming identical hardware, depends on the program itself (that is, on what you are trying to calculate or compute or perform), but if you choose many different programs you can have a comparison between platforms ("platform" in this case means combination of programming language plus compiler or interpreter.) Here is a graphic comparing Julia to other platforms: https://julialang.org/benchmarks/ https://julialang.org/benchmarks/ You can see that according to that graphic, R can be up to 400x slower than Julia (or C). There is another independent benchmark here. It does not include R but it includes Python, and it's very interesting. http://benchmarksgame.alioth.debian.org/ http://benchmarksgame.alioth.debian.org/ But the best comparison graphs are here, this is worth a look: http://blog.gmarceau.qc.ca/2009/05/speed-size-and-dependability-of.html http://blog.gmarceau.qc.ca/2009/05/speed-size-and-dependabil...
      - ferdterguson 9y ago
        > Python in some cases can be 100x slower than C Yeah, the Python implemented version is. But people doing serious computing in Python that requires speed are doing it with NumPy or even Cython or just straight up calling C/Fortran libraries in Python.
        fourthark 9y ago
        It's the same in R - you can have good performance of you use vectorized routines, because those routines are written in C/C++/Fortran.
- throwaway7645 9y ago
  Go isn't meant for computational science, but business code as services. Python is really good in this area when combined with Matplotlib, Numpy, & ScyPy.
VHRanger 9y ago
Biggest changes were from 0.3->0.5. Even from 0.6->1.0 it's not expected to break nearly as much code as before. It's easier to argue to use Julia now since the New York fed uses it for their full scale model of the US economy
flavio81 9y ago
An alternative would be getting your bosses approve for you to use Common Lisp. It is mature (30+ years in use, rock solid ANSI standard), many compilers available for many plataforms, highly portable code, multiple dispatch (CLOS is arguably the most powerful object oriented system available), you can also very easily call C libraries with CFFI and there is portable support for threading... ...and well written CL code will approach Fortran and C in speed, at least with SBCL which is a free, popular compiler. You can also compile to the JVM and to the LLVM, and if you need pro support for it, there is LispWorks and Franz Inc. Ah, and with ABCL you can also easily call java libraries, if you need to.
- throwaway7645 9y ago
  Yes Common Lisp is freaking amazing, but it isn't as good of an out of the box experience as a lot of these products. How do you do plotting besides pushing to GNUPlot? The REPL is great, but can you specify types? Yes it can reach near C/Fortran speeds, but how much time will I spend optimizing it? Are the built-in linear algebra, optimization...etc libraries good, or must I toil with FFI? How much time is really saved over dealing with all that and learning SBCL, slime, emacs...etc? Data scientists just don't have the time.
  - flavio81 9y ago
    I was assuming that the use case cited above by drej was doing a complex, big system requiring numerical computation/analysis. It was mentioned that Julia was rejected on grounds of being immature technology, so I assume we're talking about a big project, perhaps destined to run as a server or as a clustered system. The use case you imply is different, it is more like a scientist working alone at a workstation for interactive calculations, in which case perhaps Python with Numpy and Scipy is just perfect. In any case, for the previous use case (complex, big system), i'd use CL for the calculations themselves, for the visualizations I could do interactive web graphics using the extremely nice Bokeh library for Python. I'd use Bokeh under Python (using Flask as a web framework) for the presentation layer, and the Lisp system for the actual calculations. BTW, CL does have many libraries for plotting: http://cliki.net/plotting http://cliki.net/plotting > Are the built-in linear algebra, optimization...etc libraries good, or must I toil with FFI? You can use the GNU Scientific Library also within Lisp, to do everything you mention and more: https://www.gnu.org/software/gsl/ https://www.gnu.org/software/gsl/ https://common-lisp.net/project/gsll/ https://common-lisp.net/project/gsll/ > Yes it can reach near C/Fortran speeds, but how much time will I spend optimizing it? I guess you are aware that for maximizing speed in C you need to be very careful on how you write your code. In this sense there is no easy way out in any language. The good part is that Lisp, unlike C, is garbage collected, so working on a piece of code can be easier since there is no need to deal with pointers, allocation, and deallocation. It also supports arbitrary number precision and fractions and complex numbers out of the box, no need to do anything in particular to optimize their performance nor rely on external libs to operate on them. It is particularly good (speedwise) with big integers and arbitrary precision numbers. As for Common Lisp itself, besides that observation -which applies to all languages-, you need to add type declarations. This improves the performance a lot. You are going to add type declarations on a statically typed language (C, C++, Julia, Java, etc) anyways, so i fail to see how this could take more time than in other languages.

by drej 9y ago

smt88 9y ago

My girlfriend does incredibly complex modeling (she's a computational geneticist). From my limited understanding, it sounds like you may have some of the same challenges that she does. Do you think Python and Go are more suited to that domain than R or Julia?

flavio81 9y ago
R is very, very slow compared to all the other programming platforms (including Ruby, a nice language but with some of the slowest implementations out there). R syntax is very unlike most programming languages so jumping to other languages can feel strange. Python with Numpy, SciPy and Pandas, is the current widespread alternative to R and a good starting point. Recommended!!
- nnm 9y ago
  For all the cases I worked on, R is not slow at all. At home, I wrote a R package to do deep learning (include most common layers, but only conv layer was implemented in C++) from scratch, and I also wrote one in python (using numpy + numba). My R version is same fast as python on the MNIST dataset (without conv layers, hence all code is in plain R or python).
  - flavio81 9y ago
    What this means is that for your purposes R is fast enough. But in the absolute sense (or in any case, relative to other programming platforms), R is one of the slowest ever. R is usually slower than Python, and Python in some cases can be 100x slower than C. Performance, assuming identical hardware, depends on the program itself (that is, on what you are trying to calculate or compute or perform), but if you choose many different programs you can have a comparison between platforms ("platform" in this case means combination of programming language plus compiler or interpreter.) Here is a graphic comparing Julia to other platforms: https://julialang.org/benchmarks/ https://julialang.org/benchmarks/ You can see that according to that graphic, R can be up to 400x slower than Julia (or C). There is another independent benchmark here. It does not include R but it includes Python, and it's very interesting. http://benchmarksgame.alioth.debian.org/ http://benchmarksgame.alioth.debian.org/ But the best comparison graphs are here, this is worth a look: http://blog.gmarceau.qc.ca/2009/05/speed-size-and-dependability-of.html http://blog.gmarceau.qc.ca/2009/05/speed-size-and-dependabil...
    - ferdterguson 9y ago
      > Python in some cases can be 100x slower than C Yeah, the Python implemented version is. But people doing serious computing in Python that requires speed are doing it with NumPy or even Cython or just straight up calling C/Fortran libraries in Python.
      - fourthark 9y ago
        It's the same in R - you can have good performance of you use vectorized routines, because those routines are written in C/C++/Fortran.
throwaway7645 9y ago
Go isn't meant for computational science, but business code as services. Python is really good in this area when combined with Matplotlib, Numpy, & ScyPy.

VHRanger 9y ago

Biggest changes were from 0.3->0.5. Even from 0.6->1.0 it's not expected to break nearly as much code as before. It's easier to argue to use Julia now since the New York fed uses it for their full scale model of the US economy

flavio81 9y ago

An alternative would be getting your bosses approve for you to use Common Lisp. It is mature (30+ years in use, rock solid ANSI standard), many compilers available for many plataforms, highly portable code, multiple dispatch (CLOS is arguably the most powerful object oriented system available), you can also very easily call C libraries with CFFI and there is portable support for threading... ...and well written CL code will approach Fortran and C in speed, at least with SBCL which is a free, popular compiler. You can also compile to the JVM and to the LLVM, and if you need pro support for it, there is LispWorks and Franz Inc. Ah, and with ABCL you can also easily call java libraries, if you need to.

throwaway7645 9y ago
Yes Common Lisp is freaking amazing, but it isn't as good of an out of the box experience as a lot of these products. How do you do plotting besides pushing to GNUPlot? The REPL is great, but can you specify types? Yes it can reach near C/Fortran speeds, but how much time will I spend optimizing it? Are the built-in linear algebra, optimization...etc libraries good, or must I toil with FFI? How much time is really saved over dealing with all that and learning SBCL, slime, emacs...etc? Data scientists just don't have the time.
- flavio81 9y ago
  I was assuming that the use case cited above by drej was doing a complex, big system requiring numerical computation/analysis. It was mentioned that Julia was rejected on grounds of being immature technology, so I assume we're talking about a big project, perhaps destined to run as a server or as a clustered system. The use case you imply is different, it is more like a scientist working alone at a workstation for interactive calculations, in which case perhaps Python with Numpy and Scipy is just perfect. In any case, for the previous use case (complex, big system), i'd use CL for the calculations themselves, for the visualizations I could do interactive web graphics using the extremely nice Bokeh library for Python. I'd use Bokeh under Python (using Flask as a web framework) for the presentation layer, and the Lisp system for the actual calculations. BTW, CL does have many libraries for plotting: http://cliki.net/plotting http://cliki.net/plotting > Are the built-in linear algebra, optimization...etc libraries good, or must I toil with FFI? You can use the GNU Scientific Library also within Lisp, to do everything you mention and more: https://www.gnu.org/software/gsl/ https://www.gnu.org/software/gsl/ https://common-lisp.net/project/gsll/ https://common-lisp.net/project/gsll/ > Yes it can reach near C/Fortran speeds, but how much time will I spend optimizing it? I guess you are aware that for maximizing speed in C you need to be very careful on how you write your code. In this sense there is no easy way out in any language. The good part is that Lisp, unlike C, is garbage collected, so working on a piece of code can be easier since there is no need to deal with pointers, allocation, and deallocation. It also supports arbitrary number precision and fractions and complex numbers out of the box, no need to do anything in particular to optimize their performance nor rely on external libs to operate on them. It is particularly good (speedwise) with big integers and arbitrary precision numbers. As for Common Lisp itself, besides that observation -which applies to all languages-, you need to add type declarations. This improves the performance a lot. You are going to add type declarations on a statically typed language (C, C++, Julia, Java, etc) anyways, so i fail to see how this could take more time than in other languages.