Transfer Learning with a Transformer based on https://github.com/salu133445/mmt a multitrack music transformer that generates music given a seed. Two different versions of the model are presented and trained in this repository.
- Retraining of top layers on the MCMA - Multitrack Contrapuntal Music Archive. It worked well, but the output is similar to that of the original model.
- Remodelling to setting counterpoint to a melody, instead of generating music given a seed - (continuing a sequence). We follow the approach by Nichols et al. This did not work well, perhaps more low-level retraining is required, or maybe the tasks of sequence prediction and sequence translation are too far apart.
Samples with audio and sheet music are available here
We recommend using Conda. You can create the environment with the following command.
conda env create -f environment.yml- MCMA dataset - This is what we used for the simple transfer learning, contains the scores (music XML) of 470 contrapuntal pieces from the composers Albinoni, Bach, Becker, Buxtehude, and Lully.
- Sequence Translation Dataset - consists of the scores of roughly 700 baroque 2 and 3-voiced pieces and an additional 500 multitrack pieces including orchestral works. The pieces are reconfigured into two track pairs, arbitrarily spliced into 4-measure sections, and filtered to contain more than 10 notes, yielding a dataset of 41,297 four-bar, two-voice segments. 3)The original model is trained on the Symbolic orchestral database (SOD)
Other large data-sets to consider:
We are assuming data is already split into a testing and training set Use the convert_extract_load function, to convert data from midi to json to npy and load them into a dataset.
More on that soon
More pretrained models can be found here. You can use [gdown] to download all the pretrained models via command line as follows.
gdown --id 1HoKfghXOmiqi028oc_Wv0m2IlLdcJglQ --folder