search

keunwoochoi / lstm_real_book

LSTM source code to generate jazz chord progressions
130 stars 18 forks source link

Chord sentences, one per line #1

Closed loisaidasam closed 7 years ago

loisaidasam commented 7 years ago

instead of using __START__ / __END__ delimiters

keunwoochoi commented 7 years ago

Thanks for PR. What would be it good about this change? I think using delimiters are more common. We can then see what happens when the network meet them. Also the code is written for that.

On 6 Feb 2017, at 20:01, Loisaida Sam notifications@github.com wrote:

instead of using START / END delimiters

You can view, comment on, or merge this pull request online at:

https://github.com/keunwoochoi/lstm_real_book/pull/1

Commit Summary

Chord sentences, one per line File Changes

M chord_sentences.txt (2848) Patch Links:

https://github.com/keunwoochoi/lstm_real_book/pull/1.patch https://github.com/keunwoochoi/lstm_real_book/pull/1.diff — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

loisaidasam commented 7 years ago

Hi Keunwoo,

First off, this is amazing and quite inspiring. Thank you for your hard work.

My initial thoughts were that it'd be helpful for (human) readability, but I see what you mean, that the RNNs actually parse the delimiters as inputs.

I was thinking about other analyses that could be done on these sentences, perhaps by using vector representations of the chords that each sentence contains to do some sort of similarity comparison between sentences themselves.

Would it be possible to get a list of the song titles represented by each "sentence" of chords?

loisaidasam commented 7 years ago

@keunwoochoi

Here's what a CSV of songs w/ titles and associated chords might look like:

https://github.com/loisaidasam/lstm_real_book/commit/695bc6d8e9121f6eda8bb670dd7bd62c28183df0

keunwoochoi commented 7 years ago

Thanks! this folder has the original .lab files, and I think they are just alphabetical order in the merged text file. Vector representation sounds fun. I can't quickly think of good representation that preserve the information in the chord - root and chord - though. Hm.. hm..

loisaidasam commented 7 years ago

Hey @keunwoochoi,

To confirm the ordering, I looked at the .lab files in /more_data_to_play_with/jazz_xlab.zip to try and associate them with the chord "sentences" in chord_sentences.txt. To do this, I wrote this converter.

That worked just fine, but it seems that all of the files included in the repo are in their original keys, not transposed to C. I don't have any easy way of transposing them (I don't have Band-in-a-box or anything). I tried a few naive things based on my knowledge in music theory, but it gets more complicated when tunes introduce more sophisticated chords, which often happens in jazz. For example, see the Am7 chord on the fifth line of Misty, a tune in the key of Eb. The key of Eb doesn't have an A, it has an Ab. This Am7 chord actually serves as the #11 voicing, but how do you express those relationships programatically? I'm assuming there's a good chart somewhere or something. Perhaps a project for another day...

Anyway, what did you do to transpose the tunes to C? Did you use Band-in-a-box? Do you have the transposed .lab files saved anywhere? or would you be willing to generate them again?

In the mean time, I wrote some scripts to do a little automated regex matching to try and match the patterns of chord types against the results in chord_sentences.txt, excluding the actual notes of the chords (something like ([^\s:]+\:maj7)\s([^\s:]+\:maj7)\s([^\s:]+\:min7)\s([^\s:]+\:min7)\s ...), and found that they are ALMOST in alphabetical order, but with a bunch of random items out of order. For example, when listed in alphabetical order, the second listed file 'deed I Do.Mgu.xlab is actually the 611th sentence in chord_sentences.txt.

My writing is getting a bit egregious for a closed pull request, so I'm going to open an issue if you don't mind, and we can continue the discussion there :)

-Sam

PS. Re: the vector representation, I have a few ideas. To get started, I was thinking something along the lines of creating custom embeddings based on the n-grams contained in each "sentence" of chords. I will elaborate on that a bit later once some more of the data wrangling is done.

keunwoochoi commented 7 years ago
keunwoochoi commented 7 years ago

PS. I have been/will be extremely busy for about a month, so please excuse me if I'm not too responsive!