I have replaced myself with a simple neural network



  • So, I've been toying with that card-generating RNN I've posted before. But, since you can look at Magic cards only so much, I decided to use another corpus. And what better corpus is there than my infinite wisdom, as exported from these very forums?

    So I let it churn overnight, generated a few samples, and well, it does need some better training... Some snippets:


    "I didn't even have the whole thread about this game - and it's silly embedding. Can you people have pretty much simple tables for much code condempast, they pinge stuff."

    (x*:notest((blad.jonsorm(""1000011""- ""accalia title"", ""#"";```

    I'm not even sure what the key, colling programs it is ""call"". And if you don't, the times it treat software as it seems like Discourse iso...


    The guy on the internet - you have a shitty code and aren't that bad. It's also fairly quickly who at least sent an instance and useful megian where a children and the about Polish model is that the outer Light as a rent in an option of Jeff can view,


    That gets you beyond guy who crap.


    since it seems Malling and Miscourse as hell fuck are now than the forums, and have a really joke, and all the realist counteren due to the comparent."


    Yeah, I thought ""we've else the sheet-leaded pages""."
    Well, until he fans-quote meta.d. I never really need at all be a


    Yes, that's me notification on your body cards off. What did you think ""if you just want to pass, unless in honest connection"" or so, with a bottle experience (no, if you write it to now because you can't get gold pop up on a consolving Alon Do AND from the thing standall, though."


    Maybe it was one of those babs-getting parady of indicators with his or guys themselves over in the function return for that.


    Besides, it should be ones for some settings, with the whole ""idiot return 0 with the conscented in There is a defined event, strings removing plays it to test


    Rame non-email inition, I'm so nice team in the women. I know about it - anything is a diplotical andievical shifting play! I can know which I've just really a stream.



  • I think you may have needed more training time.



  • @riking said:

    I think you may have needed more training time.

    A bigger training set is more likely. The loss kept increasing with every cycle through the data starting with like cycle 5, so.

    I should tweak it a bit, but I can't jump the fact that my wisdom is just 1MB total.


    Filed under: now, if I fed it Blakey's set...



  • @Maciejasjmj said:

    That gets you beyond guy who crap.

    Sounds pretty good to me.



  • I don't see any improvement over a simple trigram Markov chain. At what level do you generate, BTW?



  • @Hanzo said:

    At what level do you generate, BTW?

    What do you mean by "level" in this context? AFAIK, the network is technically character-level, but keeps some memory of previous states. I Am Not An AI Specialist, though.

    I'm using this network (a nice explanation on how it works), with a 500-neuron/3-layer setting.


  • Discourse touched me in a no-no place

    @Maciejasjmj said:

    Miscourse

    QFF



  • @riking said:

    I think you may have needed more training time.

    So I gave it 100 more epochs. Didn't help much...

    Still, Games I'm a retard.


    Personally, I'm arranged I had profized.
    And I'm weird nuck works on your must called community and maintainable.


    My biggest gripe is that it was basically the reports or something like that."


    All of your social zeroes that I have no prople which I have yet to fire it without heavy


    That breaks, so we help.


    Not getting into the whole API. A lendthing you can try to use Discourse!"


    You have the way in a first nation, because most of the people would make beneficing a VMs, software somewhere in the first place.


    Especially if the libration is basically a pirate standard top on, and then we wouldn't get the fuck up and have valuation.


    So probably cross-origin-problem for printers?


    "Status: very well."


    The problem is that you can move the designer or a cig


    My C64 screwing around here!


    This might, you're in fucking routing, more bad defaults


    The point is, we don't seem that bad, but they're fucking.


    You want to know what the fuck.


    What's the point of what you do was fucking shit."


    1. Come on, that's not even @­accalia, so I don't think so.

    That's not a synonymoon.

    It's definitely a while.


    Hour and I'm not a girl."


  • ♿ (Parody)

    Nobody networks knowledge better than this!


  • :belt_onion:

    Obi-wan and Han on the bridge of the Millennium Falcon seeing the Death star for the first time, with the caption 'That's not a synonymoon'

  • Discourse touched me in a no-no place

    I like how your bot is more vulgar than you are.


  • :belt_onion:

    Obama smoking with the caption 'The problem is that you can move the designer or a cig'




  • @Maciejasjmj said:

    What do you mean by "level" in this context? AFAIK, the network is technically character-level

    That's relatively impressive, then. Characters convey very little information. If you want to improve on the same data, you should change the size of the network: a large network will build an accurate state, but not bring enough variation, while a small network will bring variation, but it will be crappy. It's difficult to tell how much of a change you'll see, but try 250 and 1000.





  • @Hanzo said:

    a large network will build an accurate state, but not bring enough variation, while a small network will bring variation, but it will be crappy. It's difficult to tell how much of a change you'll see, but try 250 and 1000.

    I'm running a 500-neuron/3-layer network on a bigger (15MB or so) corpus currently. Can't really increase it further, since my GPU runs out of memory - and it's already going to take a full two days or so there, so moving it out to CPU would be rather difficult.





  • Awesome, I've always wanted an H1-B mailing list question generator.



  • This is my day sorted then.



  • Reads like @xaade on a good day.


  • 🚽 Regular

    My C64 screwing around here!


  • You can always try to lowering the number of neurons. That usually leads to more freedom, and more noise. It's a tough balancing act.

    What GPU are you running it on? And how long does training take?



  • It's a GTX 760.

    There's a bunch of parameters to balance there, but I don't really know that much about theoretical basics of NNs, and 50 passes (by default) through the corpus take quite a while - that 1MB with my posts was 3 or 4 hours.



  • Ok, that's a seriously slow process then. Interesting, Thanks for the info.



  • Using the combined Likes and Status thread (13,4 MB), 70 passes over the corpus, also tweaked the "temperature", whatever that is.

    It doesn't really make it any more comprehensible, but it seems like there's some improvement...

    But wait, there's something he says he could change into a Jeff-Atwoodica fail.


    I'll give you a like if you carry the guns I can do it.


    I think I didn't start beating your statement in my cabin. You can have it if I was down to try to be an interpreted text, but the system is coming from the believe, you can see that they are cool and hidden. If you want to see, you might be the other way, or in their office you earn any, because they don't respond to the same company as the lack of computer game I stopped liking.


    I don't have a post stream and I should go for it.

    And make a firefly.


    I also are very, very rude like this. You may not be completely unreliable to me:


    I know not what happens. Made of fired performance is not a baby connecting fake alternative to hate them lurkin, but they can log in to his panting.


    Stand by an air puzzle game that I had already been at that point in the first place.


    let me check, quote your post and sudo quickly.


    Indeed. You may need to reconstrume it to multiple proof that your father says "emoji".


    I want it all, I don't have a space meaning. I am not a valid host.

    I don't know what to do


    An Chrome plan for rape and clock this shit in inconsistent idea.


    Too late, Microsoft.
    I guess it's a bug that will overtake me in any jacket.


    Not sure why she said, "You are such a chat platform, see raw problems." That's my obvious CSS for Discourse being able to use (seriously), but whatever.


    The only reason I can guess the forum is not a TV shouting to find it that I don't want to think about the results.


    am I quite computed to true on the interwebs and other failures.


    I know he wasn't.

    I know @coddinghorror didn't make sense.


    So the customer was killing my post about how you can reply to the original topic, and then his creation comes by as they were paid to party...


    I think it was a pop-up Enterprise


    There are some things that hurt they're really hard for people that even make it base you can have a doubt.


    Current status: contacting a turd



  • Time for a duel with translator?


  • :belt_onion:

    @Maciejasjmj said:

    Current status: contacting a turd

    So did it respond?


Log in to reply