Add attention melody checkpoint.

AFAICT, attention-based models (#10) are using encodings not currently supported by magenta-js converters:

• Melody attention_rnn uses a KeyMelodyEncoderDecoder with attack, ascending/descending, lookback, and a binary clock.
• Drums RNN uses a one-hot multi-drum encoding with a binary clock

Adding appropriate DataConverters for these seems straightforward enough, but how can we also have the appropriate information in the model converter.jsons so the right ones will be picked?

[adarob]

I see your point. One option could be to add a pipe character to the type field in the spec and add an new ConverterArgs type that is composed of 2 different args.

The other is just to add these two specific converters and be done with it since I don’t foresee us adding others like this in the future.

On Sat, Apr 7, 2018, 8:46 AM Tero Parviainen [e-mail] wrote:

If you don’t think it’s hard to implement, then you could go for it. We can just change the converter.json file.

I’ll try with the drums one as I’m already familiar with it. Just needs the 2^9 one-hot encoder and then the binary counter.

How about we add something like nextInput to the DataConverters that concats all the additional info. Would this work? I haven’t had a chance to look back at the old code.

I think it should be possible to compose DataConverters similar to how encoders are composed in Python. Will need to try. Not sure how to express this in the converter.json though. Would need to specify either an array of converters, or then just introduce new types for each composed version e.g. MultiDrumOneHotBinaryCounterConverter 😃

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/tensorflow/magenta-js/issues/14#issuecomment-379478805, or mute the thread https://github.com/notifications/unsubscribe-auth/ABCa6B3Bb959gkPjeSFrgdvfkPG0fpu2ks5tmN8xgaJpZM4TLCJ1 .

[adarob]

Covered by #11