Thank you! about coffemidi HOT 6 OPEN

Hi Alex!

Has been a great surprise to find your message.

CoffeMIDI has been the first job I did while studying data science

The original version, not implemented in web, has the option of loading a MIDI to be compared with the complete dataset. It would be nice to implement it in a future version.

Also, You are right about the LAKH dataset
I have taken hundreds of MIDI files from different places, only filtering for similarity I found a good number of duplicates and I have filtered them and finally I am left with 90k. I think that with it I could also solve the lack of a really descriptive name. Can you tell me more about the dataset and your experience with it ?

I also have seen a little of your work and I found very interesting. For sure it would be good to collaborate. Tell me more about what you have in mind.

Greetings

Patricio

from coffemidi.

Hey Patricio,

You are very welcome :) You did a good job with Coffe MIDI :)

Yes, LAKH dataset is very nice. It has all annotations and everything. It is also de-duped and contains most pub MIDIs. Its about 170k MIDIs.

I already built a very simple and efficient classifier based on LAKH. Check it out:
https://github.com/asigalov61/tegridy-tools/blob/main/tegridy-tools/notebooks/Instant_Classy_MIDI_Classifier.ipynb

It uses a simple match by MIDI notation (start-time, duration, channel, pitch, velocity). This way it is possible to achieve 70% accuracy with minimum compute costs for such a large dataset. Its not as advanced as your matching system, but it is very fast and simple.

RE: Collabing: I mostly wanted to see if I can help you with Coffe MIDI? I can't do web stuff, but I can do Python stuff. And I can also help with hosting I think. I was thinking Pythonanywhere would be a good hosting place for a project like that.

Let me know.

Alex

from coffemidi.

I'm checking Instant_Classy_MIDI_Classifier, thanks !

have you already seen the website?
it is quite simple, visually precarious, but you can see the model API working.
Check finding different gender results.
I am open to returns and points for improvement.

http://coffemidi.com.ar/

Because of how I have modularized the work, implementing a new dataset does not cause me a lot of work, only processing time. we can exchange our contacts in order to talk it better. I am also open to thinking about new projects.

from coffemidi.

Yes, I really liked the website and the interface :) Very nice :)

I would only suggest adding a custom MIDI option (IMHO it's very important for a system like that). And also I would clean up the names/descriptions of the files because some of them have paths in them and also extensions. In my classifier, I only leave the base file name of the MIDI. This way, it is clear what song it is.

Yes, I know that processing can be difficult, especially on a large dataset. I can help with that if you want.

And sure, if you have anything else in mind, let me know. I am primarily interested in Music AI/Algorithmic Music, so anything related would be great. :)

My email is [email protected] if you want to contact me directly.

Alex.

from coffemidi.

I think your AI knowledge and talent deserve at least one gmail hahah

great! I'm going to write you there

from coffemidi.

from coffemidi.