Oh that's really awesome, great idea! I'm interested to see this expanded somehow, without using such strange time signatures :) (Btw I didn't even realize it was you sci, until the video ended! I thought it was a random Seth experiment, haha)
no the originals were in pretty standard timesigs iirc. but I've been getting into weird timesig stuff lately. i even gave a list for my last mix (sci2017), though that didn't included oddities like 1/4 during transitions etc. sci2016 has even more weird timesig changes. every prime number between 3 and 13 is included in there. it's fun!
I think they are standard too. Various prime number signatures sound like a fun thing to fool around with! Did you find it challenging to make something you enjoyed? Or was it easy to find inspiration in the constraint? :)
Amazingly, this is from several months ago, and I've since cannibalized the RNN for a different experiment. so....I no longer have the original output or trained model. I can tell you that I actually fed the network several NES tunes, and it just happened to get stuck on Mario 1-1 when generating its output.
What's the NSF file for? Can you explain what software you used to make this? I've been wanting to do something similar for a while but I've only ever used RNN's on text files. Really cool!
sure: an NSF is a way to store NES chiptunes in a way that they can be played back on actual hardware. Software-wise: I used Tensorflow (in a Python environment) for the neural net stuff. RNNs actually work quite well on non-text data, though you may need to increase your sample size etc. In this case, I converted all of the music data into a text file containing the tracker rows. This lowered down the possibility space, since you only needed a note, an octave, and an instrument (NES terms: pulse width) over 4 or 5 channels. After training the network, it still will produce some illegal output (such as 6 channels for a single row, or playing a triangle note in the noise channel), which generally has to be fixed manually. In the case of this particular song, all of the notes have the same length, which lowers down the possibility space even more, resulting in more "coherent" output. hopefully that provided some semblance of an answer to your question. :P
![Mario Kart Wii - 50cc Mushroom Cup Grand Prix (Koopa Troopa Gameplay) [2/4]](http://i.ytimg.com/vi/FLl0glaIgi0/mqdefault.jpg)
Oh that's really awesome, great idea! I'm interested to see this expanded somehow, without using such strange time signatures :)
(Btw I didn't even realize it was you sci, until the video ended! I thought it was a random Seth experiment, haha)
"without using such strange time signatures"
Haha is it because the originals were in weird signatures? I mean if it stays the same during the whole song, wouldn't it be more cohesive?
no the originals were in pretty standard timesigs iirc.
but I've been getting into weird timesig stuff lately. i even gave a list for my last mix (sci2017), though that didn't included oddities like 1/4 during transitions etc.
sci2016 has even more weird timesig changes. every prime number between 3 and 13 is included in there. it's fun!
I was like "At least they're all halves and quarters of 4/4 and not like, some really truly weird time signatures like 7/9 ..."
I think they are standard too.
Various prime number signatures sound like a fun thing to fool around with! Did you find it challenging to make something you enjoyed? Or was it easy to find inspiration in the constraint? :)
well it tried
Please try this again, with all the improvements that came out in the past 8 years!
ha, yeah, I don't think this implementation was even tokenized(!). i dunno if i'll actually revisit it any time soon, though it is a fun thought.
Wicked! Just gotta feel it..
You should post a before/after comparison of earlier ones! If it gave you them that is.
Amazingly, this is from several months ago, and I've since cannibalized the RNN for a different experiment. so....I no longer have the original output or trained model.
I can tell you that I actually fed the network several NES tunes, and it just happened to get stuck on Mario 1-1 when generating its output.
Do the same to the Angry Birds theme
Autistic Mario Bros.
What's the NSF file for? Can you explain what software you used to make this? I've been wanting to do something similar for a while but I've only ever used RNN's on text files. Really cool!
sure: an NSF is a way to store NES chiptunes in a way that they can be played back on actual hardware.
Software-wise: I used Tensorflow (in a Python environment) for the neural net stuff. RNNs actually work quite well on non-text data, though you may need to increase your sample size etc. In this case, I converted all of the music data into a text file containing the tracker rows. This lowered down the possibility space, since you only needed a note, an octave, and an instrument (NES terms: pulse width) over 4 or 5 channels.
After training the network, it still will produce some illegal output (such as 6 channels for a single row, or playing a triangle note in the noise channel), which generally has to be fixed manually.
In the case of this particular song, all of the notes have the same length, which lowers down the possibility space even more, resulting in more "coherent" output.
hopefully that provided some semblance of an answer to your question. :P
This didn’t sound very good. Try Musenet. It’s not exactly the same but it usually sounds good.
I, for one, love the pure prog bullshit that this AI produced
lol