from transformers import AutoModel
import soundfile
import torch

model_name = "HiDolen/Mini-BS-RoFormer"
model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
)

# 加载音频
file = "./Bruno Mars - Runaway Baby.mp3"
waveform, sr = soundfile.read(file)
assert sr == 44100  # 采样率必须为 44100Hz
waveform = torch.tensor(waveform).T.float()

# 进行推理
result = model.separate(
    waveform,
    chunk_size=44100 * 6,
    overlap_size=44100 * 3,
    gap_size=0,
    batch_size=16,
    verbose=True,
)

# 保存处理结果
for i in range(result.shape[0]):
    soundfile.write(f"separated_stem_{i}.wav", result[i].cpu().numpy().T, 44100)

GPU 推理：

python

from transformers import AutoModel
import soundfile
import torch

model_name = "HiDolen/Mini-BS-RoFormer"
model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
)
model.to("cuda")

# 加载音频
file = "./Bruno Mars - Runaway Baby.mp3"
waveform, sr = soundfile.read(file)
assert sr == 44100  # 采样率必须为 44100Hz
waveform = torch.tensor(waveform).T.float()
waveform = waveform.to("cuda")

# 进行推理
result = model.separate(
    waveform,
    chunk_size=44100 * 6,
    overlap_size=44100 * 3,
    gap_size=0,
    batch_size=16,
    verbose=True,
)

# 保存处理结果
for i in range(result.shape[0]):
    soundfile.write(f"separated_stem_{i}.wav", result[i].cpu().numpy().T, 44100)

Training Details

使用 MUSDB18HQ 数据进行训练。

不使用原论文中提到的 Multi-STFT 损失项以提高训练速度。

学习率 5e-4，以 batch_size=6 训练 60k 步。

Acknowledgments

https://github.com/lucidrains/BS-RoFormer
https://arxiv.org/abs/2309.02612 (Music Source Separation with Band-Split RoPE Transformer)