asigalov61 / pop-melody-transformer Goto Github PK

Any recommendations for how to train in ticks (absolute time) rather than milliseconds? Is it just a matter of changing midi2ms_score to midi2score?

I'm testing the inference notebook with the provided pre-trained models, but I'm not sure what the output of the generator script is supposed to be.

When running the extract melody generator it proceeds to convert the entire POP1k7 dataset (midi files). I canceled this and reduced the number of midi files in Dataset. It writes files to the output folder, but the melody is the exact same as the original midi file. Even after changing the temperature and other settings it writes the same melody.

How is the generator supposed to work? Or did I miss something?