search

Tsung-Ping / Harmony-Transformer

Deep learning model for chord recognition
42 stars 9 forks source link

Apply on symbolic dataset #2

Open ajmehrlich opened 3 years ago

ajmehrlich commented 3 years ago

Hello! In the paper you also use the symbolic BPS-FH dataset for comparisons. I have some MIDI files I would like to test this model on, but since I am new to pretty much everything related to AI, I have no idea which changes are necessary to apply this code to symbolic data, particularly how to integrate tonal centroids and chord change. Is there a chance you can provide a quick summarization or something like this? I would appreciate it.

Tsung-Ping commented 3 years ago

Hi!

You can feed any type of data into the model as long as the data are represented as sequences of features. For example, you can transcribe your midi files into piano-roll representations (using package like pretty_midi), compute tonal centroids from the piano rolls, and naively concatenate the tonal centroid features and the piano roll representations. As for the chord change, it is processed inside the model, so you don't need to care about it. Note that you need to change the hyperparameters of the model (e.g., n_steps, frequency_size, etc.) according to your data shape.

Another way is to synthesize the midi files to audios, and then feed them into the model. You are recommended to visit omnizart, a project of our lab, whose chord recognition part employ the Harmony Transformer, and can process audio data directly.

ajmehrlich commented 3 years ago

Thank you for your response. I'll try your suggestions.

I have another question: You opened another repository for harmony transformer v2, but it's empty. Do you plan to release it soon or will this take a while?

Tsung-Ping commented 3 years ago

Hi,

I will release the code in a few weeks.