Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
coffeetable
Feb 5, 2006

TELL ME AGAIN HOW GREAT BRITAIN WOULD BE IF IT WAS RULED BY THE MERCILESS JACKBOOT OF PRINCE CHARLES

YES I DO TALK TO PLANTS ACTUALLY

vodkat posted:

For a super hardcore markov bot you make a per character neutral network like this guy http://karpathy.github.io/2015/05/21/rnn-effectiveness/ but that is way beyond my lovely programming skills or my MacBooks integrated graphics.
ive got torch n all that installed (neural nets are part of my dayjob). someone give me an input .txt file and some prompts i'll train one up n see how it compares to ngram

also if anyone else wants to do this, the successor to karpathy's library is

https://github.com/jcjohnson/torch-rnn

coffeetable fucked around with this message at 19:00 on Feb 20, 2016

Adbot
ADBOT LOVES YOU

coffeetable
Feb 5, 2006

TELL ME AGAIN HOW GREAT BRITAIN WOULD BE IF IT WAS RULED BY THE MERCILESS JACKBOOT OF PRINCE CHARLES

YES I DO TALK TO PLANTS ACTUALLY

vodkat posted:

I made a huge corpus (2 million +) word corpus for markov botman from stupid silicon valley VC but he's still not generating very good tweets. I'm gonna try and write a better tweet generator over the next few days and report back with the results.

post the file

coffeetable
Feb 5, 2006

TELL ME AGAIN HOW GREAT BRITAIN WOULD BE IF IT WAS RULED BY THE MERCILESS JACKBOOT OF PRINCE CHARLES

YES I DO TALK TO PLANTS ACTUALLY

Snapchat A Titty posted:

e: are the hashes literally generated by the markov chain too? like those arent real hashes from the commits you fed it?

it's an rnn, not a mc. mc wouldn't be able to indent things consistently (context isn't long enough)

see

http://karpathy.github.io/2015/05/21/rnn-effectiveness/
https://github.com/jcjohnson/torch-rnn

coffeetable
Feb 5, 2006

TELL ME AGAIN HOW GREAT BRITAIN WOULD BE IF IT WAS RULED BY THE MERCILESS JACKBOOT OF PRINCE CHARLES

YES I DO TALK TO PLANTS ACTUALLY

vodkat posted:

I'm playing around with these at the moment but I still haven't got anything I'm happy with. Charter rnn seems to work fine but takes ages to run since I don't have a gpu. Torch-rnn is much faster and efficient but the output so far has mostly been unusable.
what training/test loss do you get from each? & check the default settings for torch-rnn vs char-rnn. they both implement the same architecture, so you should be able to get very similar results out of both

if it's too slow, make the network smaller. io aside, runtime is quadratic in the number of neurons and linear in the sequence length and number of layers. performance should degrade fairly gracefully as you cut it down to size

coffeetable
Feb 5, 2006

TELL ME AGAIN HOW GREAT BRITAIN WOULD BE IF IT WAS RULED BY THE MERCILESS JACKBOOT OF PRINCE CHARLES

YES I DO TALK TO PLANTS ACTUALLY

probably based on this paper

http://arxiv.org/pdf/1506.06714.pdf

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply