Automated Rhythm Generation

ARG: Automated Rhythm Generation.

Let's generate rhythm game maps automatically!

Features

Neural Networks

• RhythmGen: Generate rhythm sequence from music piece.
  • RGGRU: RhythmGen with CNN feature extractor and GRU. (460K)
  • RGGRUAT: RhythmGen with CNN feature extractor, GRU, and multi-head attention. (720K)
  • RGTR: RhythmGen with CNN feature extractor and Transformer. (15M)
  • RGRoFormer: RhythmGen with CNN feature extractor, Transformer and RoPE. (420K)
  • Others: RhythmGen with other architectures.
• RhythmRec: Reconstruct music piece from rhythm sequence.
  • No, we didn't implement this yet.

Utils

• RhythmAnnotation (ryan): Rhythm annotation tool and format. See ryan for more details.

Installation

pip install automated-rhythm-generation

Training

from arg import Trainer

trainer = Trainer(
    "RGGRU", # Model architecture
    "JacobLinCool/taiko-2023-1.1", # Dataset
    difficulty="hard",
    num_epochs=300,
    learning_rate=0.001,
    batch_size=32,
    max_length=10.0,
)

trainer.train(
    push="JacobLinCool/RhythmGenGRU-1-hard",
    hf_token=HF_TOKEN,
)

Or you can use the command line interface:

python -m arg.train RGGRU JacobLinCool/taiko-2023-1.1 --difficulty hard --push JacobLinCool/RhythmGenGRU-1-hard

Inference

from arg import RGGRU, generate_tja
import librosa

model = RGGRU.from_pretrained("JacobLinCool/RhythmGenGRU-1-hard")

audio, sr = librosa.load("path/to/music.mp3", sr=16000)
seq = model.predict(audio)
tja = generate_tja(seq)

with open("path/to/output.tja", "w") as f:
    f.write(tja)

Or you can use the command line interface:

python -m arg.infer RGGRU JacobLinCool/RhythmGenGRU-1-hard path/to/music.mp3

Others

To see the model architecture: python -m arg.model.RGGRU, python -m arg.model.RGRoFormer, etc.