My only problem with Magenta's Transformer

[Alejandro Ruiz]

The thing is that I ran the code on my machine and it overfits. It needs a way to check that it isn't stealing from the dataset say no more than 6 or 8 continuous notes. If it can't do that it's useless. I mean your piano dataset is huge but after running the program for 20 times I found it composes note by note music of well known classical melodies. That's not ok. That should be avoided. I have a program that uses a 120 songs dataset, so I can know if it's stealing or not, and it never does. https://soundcloud.com/user-610922241

[Ian Simon]

Can you tell me more about what you tried?  Are you priming the model or generating from scratch?

In the checkpoints we've released, we tried hard to reduce the ability of the model to perform pieces from the training set.  And in the samples we released (such as the ones at https://g.co/magenta/listen), we tried hard to remove any samples that are too similar to an existing piece of music.  But it's difficult to get to 100% on these for a number of reasons, including the lack of a clear definition for "too similar".

-Ian

On Fri, Feb 14, 2020 at 4:17 PM Alejandro Ruiz <arodr...@gmail.com> wrote:

The thing is that I ran the code on my machine and it overfits. It needs a way to check that it isn't stealing from the dataset say no more than 6 or 8 continuous notes. If it can't do that it's useless. I mean your piano dataset is huge but after running the program for 20 times I found it composes note by note music of well known classical melodies. That's not ok. That should be avoided. I have a program that uses a 120 songs dataset, so I can know if it's stealing or not, and it never does. https://soundcloud.com/user-610922241

--
Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discu...@tensorflow.org
---
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discu...@tensorflow.org.

[Alejandro Ruiz]

Hi, I tried the model from scratch with the available source code. And after like 15 or 20 times it reproduced a well known classical music theme exactly as it was composed. I don't know what changes you made to your current website model to avoid this better. Can you tell? I think that checking for a maximum of 6 to 8 notes (you should try and see what works better) that are equal to a sample of the dataset should work. I hope this helps.

[Alejandro Ruiz]

Also it would be cool that in the website, one can decide how many notes in a sequence can be equal to one sequence of the dataset.

[Alejandro Ruiz]

I guess the lowest resolution note matters as well. In my program I have an 8th note duration as the lowest possible note, but if you have a 64th note duration maybe it's different?

[LUCIA TERENZI]

Hey, I am currently looking for the source code to train the model on my own machine, could I ask you where you found it? I could only find ready to go demos and unofficial implementations of the same paper (Huang 2019) but not an official Magenta repository with the model.

Thank you in advance for your help, have a nice day!