I'm currently translating the C API part of JLink to its python equivalent and I have to say that it's going much smoother than expected

@b3m2a1 Link to that part of J/Link code please?

you mean JLinkNativeLibrary.c

Seamless MMA + python experience :P

Python 3.6.6 |Anaconda custom (64-bit)| (default, Jun 28 2018, 17:14:51)
[GCC 7.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from functools import partial
>>>
>>> import numpy as np
>>> import matplotlib.pyplot as plt
>>>
>>> import wstp
>>> from wstp.hiwrapper import ExpressionSender, PythonicSender, ExpressionReceiver, PythonicReceiver
>>> from wstp.npywrapper import NumpySender, NumpyReceiver
>>> from wstp.symlist import *
>>>
>>> s = wstp.symbol_builder

I think this should be closed as a duplicate of the indicated question: mathematica.stackexchange.com/q/181065/12 I voted too quickly and choose a different reason than duplicate.

This WSTP / MathLink confusion is extremely annoying. Renaming everything to WS just for marketing is absolutely ridiculous.

I see the claim that WSTP is newer/different/does more or something else all the time. It's not true. They're the same except that in some cases MathLink must be used and WSTP isn't even an option.

@Szabolcs I understood this question more in the way that they use LibraryLink and want to use a link to transfer arguments.

@halirutan WSTP and MathLink are exactly the same thing in fact, and are generally interchangeable, except that I don't think that the WS-prefix functions work with LibraryLink. The example you copied from the documentation does show the WS-prefix functons, but I am not actually sure how to get it to work without changing all that that to the ML-prefix.

It may not even be possible.

This just shows how ridiculous this marketing-motivated renaming was. They renamed every ML-thing in the documentation to WS, even when the WS version doesn't actually work!

(BTW I did report this documentation problem to support a long time ago.)

I know sometimes we complain about parts of WL being great on the surface but not that great when you really get down into it, but I have to say the neural networks framework is a masterpiece.

would be great if you could train networks in the cloud on GPUs though

(even though I'm sure that would cost many many cloud credits)

@CarlLange Well said about the neural network framework. It is super convenient to use but I can't train networks fast enough on my computer either.

←(removed)

Personal opinion, I feel like MMA's neural network interface is poorly done.

When I first used Theano in 2014, I was thinking, "wow, this should definitely be part of Mathematica".

Theano gives a symbolic way to define and manipulate tensor expression. And a compiler to turn them into C/machine code.

I've been spending fair amount of time trying to implement a custom gradient in Theano, while in MMA, you can theoretically just define a special rule.

Instead of focusing MMA's strength - symbolic manipulation, the neural network interface turns into a Keras-like script kiddie toolkit.

Yes, it's easy to use but sucks when you wanna dive deeper.

I'm sure there are budget and marketing reasons behind this, but personally, MMA's neural network is a toy, not a serious platform for researching.

@Kh40tiK Interesting to hear you say that. I suppose like many things in MMA it falls apart a bit when you get too far down. But I certainly haven't had any problems with it really! Some small things I'd like to have (mostly slightly clearer documentation) but other than that, I've built probably half a dozen networks from papers in the last week or so and it's been so smooth.