Show HN: RNN that generates Game of Thrones text
github.com
github.com
Because it's the right thing to do in this case? If you define what you want to defend purely based on what a law you don't agree with can be used to do, as opposed to the merits of the actual instance, you're bound to pick a bunch of really bad battles.
It's not like we're talking about situations like cable where I have to pay $100 / month to watch one or maybe two shows I enjoy and have no other way to get access at a reasonable price (and even that is debatable). You can pick up a copy of GRRM's work for < $10 without hassle and contribute back to someone whose work gives you joy and the publishers who got that work into your hands.
The author of this project is riffing on an older creative work to make a new creative work, exactly as artists did for all the millenia of human civilization, until lawyers got involved in the 20th century and fucked it all up.
As far as the project goes, the author should not be bundling the books with everything else (common sense and debatable moral issues.) I have no issue with him using the texts, it's the distribution that's a problem.
A world in which authors of creative works don't have the ability to have a monopoly on those works for limited terms is a world in which people who focus full-time on writing those creative works cannot pay their grocery bills.
It is also a world in which only banks, large corporations, and highly educated folks can get software written for their interests.
I'd imagine it's sufficiently derivative as to be allowed? What if you wrote a model that produced summaries of the book - would that be allowed?
The approach in the repo also uses a word-based approach instead of a character-based approach. Although a word-approach avoids typoes/nonsense, it makes prediction much more difficult since the selection space is harder. (from notebook, error is high as well). There are lots of trade offs in this field.
Training these networks is not for the faint of heart and one of the reasons I developed a tool to train text generators using a pretrained RNN (https://github.com/minimaxir/textgenrnn) on a much more simple network architecture that trains quickly, although it probably would not hit the depth of all of Game of Thrones.
Ours was not nearly as comprehensible, but still funny at points. Some example output:
{{ Stannis runs away and Melisandre walks through the lift door , no smiles) . }}
[ TYRION ]: I wanted him to be a spoiled Greyjoy again , that 's your niece . And you 're a terrible man . My little Give me A_VOICE . .
{{ The HIGH SPARROW sits down on the bucket , DORAN sentences to have waiting to the shoulder , wearing this building . She wargs off the woods and approaches The sound of the riders . BLACK WALDER reaches over his horse beside MARGAERY . }} .
[ EDMURE ]: We 're among The Khaleesi . .
[ JORAH ]: For far soon . .
[ SANSA_STARK ]: He 's travelling a true vow to command until his father 's death.
[ DAVOS ]: I don 't mind , Your Grace . I got to build your place and see rolled women in all of their fanatics And so I getting another conversation , watcher ships in the matters of the dead as gods of Westeros .
cell = tf.contrib.rnn.MultiRNNCell([drop_cell] * num_layers)
with:
cell = tf.contrib.rnn.MultiRNNCell([create_cell() for _ in range(num_layers)])
otherwise you'll be using the same weights for each LSTM layer and your model is less expressive.
Nonetheless, I don't think GRR Martin has to worry about competition from RNN Martin anytime soon, but this is still quite fun, and surprisingly almost cogent at times.
Edit: Nice License
> Copyright (c) 2017 Zack Thoutt
Don't expect too much, though. I was reading an article on it and was excited to watch it, but it's too incoherent, in the usual fashion of today's RNNs.
I actually tried to remove the text within the first few hours of posting (I was not even home or with my computer), but I removed the text in a separate branch and when I merged it into master it created duplicates of every commit instead of removing the text from the git history. Someone put up an issue on Github later that night and I fixed my mistake early the next morning.
I own all of the ASOIAF books both in paperback and ebook and pay for an HBO subscription for the sole purpose of watching GOT for 6-10 weeks per year.
"Raved on fire, he clambered up her pallet in a swatch of blood and crept toward him shudder. She was too well Ghost back behind their crews could not lack for the warg could stand guard. The singers was a fire with our beds they could have won that make yo, tell Joffrey awarded to make repairs and threatening to my sons had entered the flesh. A life and a dim his promises of Eddark Stark, he announced the foundation and set sail cracking their wounded and bucket swayed back room to go go. Har Tormund Well the Bastard s home When the Others followed him a fine rule the better look stupid. She could see the same only empty sockets weeping child he ran with that morning. His nose Wife pleased to the man with winter s notion. My brother will never lifted a scare me wise where."
Maybe not yet a concern for this project.
Found a project ;)