training a neural network to "caption" summary vectors of lines of poetry, then using the network to generate text for summary vectors of arbitrary sentences. this is after seven epochs of training on my laptop last night; source line is first, output is second
the goal here was to be able to put in the vector for (e.g.) "dog" & get back a line about dogs. but it's learning the punctuation and the length of the lines, so putting in single words yields stuff like
abacus ➝ Beginabliny
allison ➝ It is is is is is is ine is ineay
cheese ➝ Great occhanting seaw
daring ➝ The left the lonious courtina
mastodon ➝ shorn born born borner
parrish ➝ the oh
purple ➝ Greath green green green
trousers ➝ To blenting my blank
whoops ➝ Aaann aaas! aaan aaas!
zoo ➝ T
(I know I'm literally just redoing karpathy experiments from "the unreasonable effectiveness of rnns" era but it really is unreasonable how effective this is. also my first nn-training experiment that didn't just immediately overfit and where actually adding more neurons to the hidden layers helped instead of hurt)
christmas, recurrent neural networks Show more
generated text for "Merry Christmas, friends!":
Love! haste! so love! my good soul!
a "sonnet" (14 lines generated from uniformly distributed random vectors)
my silly blyssy slyyly my smlllly ly my sfllllylyly llllyly ky my s
soremorisenoris soreeories sirey seauton oreeose,
For I sfored
That daughteous bloom
stepian steon is shente
Reminds me of finnegans wake.
I like it.
my silly blyssy slyyly my smlllly 😋
@aparrish love that daughteous bloom
@aparrish I like this approach, a sonnet is just 14 of whatever you feel like. 14 knock knock jokes? Sonnet. 14 photographs? Sonnet. 14 bottles of beer on the wall? Sonnet.
@mewo2 as björk said, don't let poets lie to you (about what is and is not a sonnet)
@aparrish I would like to adopt some of these for everyday use, like exclaiming “aaann aas” when I make a blunder and “great occhanting seaw” when I have yummy cheese.
This is also a good explanation for why one would wear trousers.
@aparrish what DOES it say about dogs?
@Ranjit I just ran it to find out and
dogs ➝ earnest heed
which is...!! on a different inference it generated "ear ee" which is also appropriate
@aparrish truly this is unreasonably effective
@aparrish dare we ask it about cats?
@aparrish as a dog partisan I’m hoping cats get something less sweet while feeling slightly guilty for hoping that
@Ranjit right now it generates
cats ➝ eers
kitten ➝ Lose lo lord
so... no idea
@aparrish happy to see both cats and dogs getting eers 🐶🐱
@aparrish I love several things about this
Server run by the main developers of the project It is not focused on any particular niche interest - everyone is welcome as long as you follow our code of conduct!