The environment prerequisites are as follows:

• python 3.8

The datasets utilized in our paper are as follows:

AIST Dance Video Database (AIST Dance DB) is a shared database containing original street dance videos with copyright-cleared dance music. The database is available here.

The GTZAN dataset is a collection of 1,000 audio files spanning 10 music genres, all having a length of 30 seconds. The audio files are available here.

The Groove MIDI Dataset (GMD) is composed of 13.6 hours of aligned MIDI and (synthesized) audio of human-performed, tempo-aligned expressive drumming. The MIDI data is available in the documentation.

The Lakh Pianoroll Dataset (LPD) is a collection of 174,154 multitrack pianorolls derived from the Lakh MIDI Dataset (LMD). We use its subset lpd-5-cleansed that contains 21,425 five-track pianorolls.

• extract human skeleton keypoints using OpenPose
• extract ground truth music beats
• extract log mel-scaled spectrogram
• convert drum track/multi-track MIDI into token sequence

To train the MBPN.

python ./src/MBPN/train_MBPN.py

To pre-train the Dance style embedding network on AIST.

python ./src/SSM/train_dance_network.py

To pre-train the Music style embedding network on GTZAN.

python ./src/SSM/train_music_network.py

To jointly train the Dance and music style embedding networks.

python ./src/SSM/train_joint.py

To train the Drum Transformer on GrooveMIDI.

python ./src/PCMG/train_drum_Transformer.py

To train the Multi-track Transformer on LPD.

python ./src/PCMG/train_multi_Transformer.py