Inference

Contents

Inference#

JAX:

# Generate 4 seconds of audio
mrt jax generate

MLX:

# Generate 4 seconds of audio
mrt mlx generate --bits=8

To print MusicCoCa tokens for a prompt directly without generating audio:

from magenta_rt.musiccoca import MusicCoCa
m = MusicCoCa()
print(m.tokenize(m.embed('a jazz piano trio')).tolist())

# Get tokens from audio
from magenta_rt.audio import Waveform
wav = Waveform.from_file("jazz_piano_trio.wav")
print(m.tokenize(m.embed(wav)).tolist())

Bulk generation#

Bulk-generate 60s audio clips from MusicCoCa prompts for listener evaluation:

python scripts/bulk_generate.py --size=mrt2_base

Outputs are saved to outputs/eval_audio/<size>/.

Contents