StepCOVNet

Running Audio to SM File Generator

Currently only produces .txt files. Use SMDataTools to convert .txt to .sm

python stepmania_note_generator.py -i --input <string> -o --output <string> --model <string> -v --verbose <int>

• -i --input input directory path to audio files
• -o --output output directory path to .txt files
• -m --model input directory path to StepCOVNet model````
• OPTIONAL: -v --verbose 1 shows full verbose, 0 shows no verbose; default is 0

Creating Training Dataset

Link to training data: Google Drive

To create a training dataset, you need to parse the .sm files and convert sound files into .wav files:

• SMDataTools should be used to parse the .sm files into .txt files.
• wav_converter.py can be used to convert the audio files into .wav files. The default sample rate is 16000hz.

Once the parsed .txt files and .wav files are generated, place the .wav files into separate directories and run training_data_collection.py.

python training_data_collection.py -w --wav <string> -t --timing <string> -o --output <string> --multi <int> --limit <int> --cores <int> --name <string> --distributed <int>

• -w --wav input directory path to .wav files
• -t --timing input directory path to timing files
• -o --output output directory path to output dataset
• OPTIONAL: --multi 1 collects STFTs using frame_size of [2048, 1024, 4096], 0 collects STFTs using frame_size of [2048]; default is 0
• OPTIONAL: --limit > 0 stops data collection at limit, -1 means unlimited; default is -1
• OPTIONAL: --cores > 0 sets the number of cores to use when collecting data; -1 means uses the number of physical cores; default is 1
• OPTIONAL: --name name to give the dataset; default names dataset based on the configuration parameters
• OPTIONAL: --distributed 0 creates a single dataset, 1 creates a distributed dataset; default is 0

Training Model

Once training dataset has been created, run train.py.

python train.py -i --input <string> -o --output <string> -d --difficulty <int> --lookback <int> --limit <int> --name <string> --log <string>

• -i --input input directory path to training dataset
• -o --output output directory path to save model
• OPTIONAL: -d --difficulty [0, 1, 2, 3, 4] sets the song difficulty to use when training to ["challenge", "hard", "medium", "easy", "beginner"], respectively; default is 0 or "challenge"
• OPTIONAL: --lookback > 2 uses timeseries based on lookback when modeling; default is 3
• OPTIONAL: --limit > 0 limits the amount of training samples used during training, -1 uses all the samples; default is -1
• OPTIONAL: --name name to give the finished model; default names model based on dat aset used
• OPTIONAL: --log output directory path to store tensorboard data

Credits

• Inspiration from the paper Dance Dance Convolution
• Most of the source code derived from musical-onset-efficient
• Jhaco for support and collaboration