What language would you choose for a deep neural network?

What language would you choose for a deep neural network?

Other urls found in this thread:

microsoft.com/en-us/research/microsoft-computational-network-toolkit-offers-most-efficient-distributed-deep-learning-computational-performance/
twitter.com/AnonBabble

Visual Basic.

C, already a library for it.

ASM if you need full performance and are insane.

windows

An ok starting point could be Matlab or octave.

statistics

or dont and be cucked by lack of intuition for solving problems with

>There are cucks who really use matlab over numpy

>these shitters do not know about muh fortran

Well, Microsoft's CNTK is written in C++, and has Python bindings.

microsoft.com/en-us/research/microsoft-computational-network-toolkit-offers-most-efficient-distributed-deep-learning-computational-performance/

Haikuu

Erlang was literally made for this.

english

Python

MATLAB is better than numpy for matrix operations
FORTRAN is much better than both

C++ (in the style of C)
Because EVERYTHING should run C [with Classes]
C/++ is only valid programming language for most programs at all

You're an idiot.

Name one application that isn't a script or gadget that wouldn't run amazingly in C/++.

Name one reason why languages shouldn't be universalized like based C/++.

Tell a programmer, who also does web dev why web languages would NOT work better if they were more-so designed after C/++
>WebC would be great. maybe one day...

C++ is fucking trash and comes with all sorts of disadvantages, such as an unstable ABI.
If you're going to use C, use C.

whitespace

Java

Erlang

>tfw i had this idea and someone's already fucking done it

Really depends on how many layers and what kind of learning it's doing

All relevant DNN implementations are in C++.

A neural Network?

Thats easy.

LISP.

>ASM if you need full performance and are insane
...ly good at it and can produce asm that is more optimized than what a c compiler would output.

How the fuck would ASM be remotely feasible? Do you have any idea how hard it is to write dependable ASM even for babby computers?

>C++ is fucking trash

You're a fucking bigot, scumlord.

C++ is a bloated piece of shit.
Most programs could/(should) be written in C, even if in a OO paradigm. Not doing so is only being lazy, unless you are prototyping.

Julia
runs as fast as C/C++
handles matrix operations like R/Matlab
intuitive psuedocode-like syntax like Python/Ruby
handles low level backend I/O very well

>i have no idea what i'm talking about

anyone who uses either C or C++ for number crunching is an idiot. the only way you could use C/C++ for scientific computing is by using libraries that force a 'safe subset' of the language so that it emulates what Fortran gives you out of the box

Julia has really shit startup time though

There is absolutely no limit on what you can do with assembly. Stop being a cry baby.

Python + tensorflow. Why make this hard.

Scala or Java, so it can be used with Spark.

there is no limit to the time required to do it either

A good library like Theano, Tensorflow, Caffe, Leaf, and many others.

For inference, it needs to be as fast as possible on whatever platform you're running on (be it embedded, mobile, etc), so C and if you're using a custom accelerator like a DSP or GPU, whatever runs fastest on your accelerator, so CUDA, OpenCL, and C with the appropriate libraries.

I didn't say there was a limit dumbass

python bindings to C++ fuctions like a normal human being.

Python

>C, already a library for it.
sauce ?

>runs as fast as C/C++
sauce required

...in which case you would definitely not be asking here.

If you're just learning about neural nets you should hand program feedforward/backprop in matlab or numpy first before you do anything else, otherwise you won't understand what's going on.

python+theano or tensorflow if you want multi gpu support, speed and flexibility to try weird architecture and weight constraints
Theano and tensorflow are very similar. theano is more flexible, slightly faster, more mature and has windows support. Tensorflow is more popular, less compile time, easier to use and has less cryptic error messages

lua+torch if you just want the fastest feedforward convnet you can get without having to directly program it in cuda

caffe if you care about deployment at all

keras if you want things to be easy and readable