11

app.py

@@ -6,12 +6,13 @@ A singing voice synthesis system powered by YingMusicSinger,
 with built-in vocal/accompaniment separation via MelBandRoformer.
 """
 
+import os
+import tempfile
+
 import gradio as gr
 import torch
 import torchaudio
-import tempfile
-import os
-import numpy as np
+from initialization import download_files
 
 # ---------------------------------------------------------------------------
 # Model loading (lazy, singleton) / 模型懒加载（单例）
@@ -22,6 +23,7 @@ _separator = None
 
 def get_model(device: str = "cuda:0"):
     """加载 YingMusicSinger 模型 / Load YingMusicSinger model."""
+    download_files(task="infer")
     global _model
     if _model is None:
         from src.YingMusicSinger.infer.YingMusicSinger import YingMusicSinger
@@ -35,6 +37,7 @@ def get_separator(device: str = "cuda:0"):
     加载 MelBandRoformer 分离模型 / Load MelBandRoformer separator.
     Returns a Separator instance ready for inference.
     """
+    download_files(task="infer")
     global _separator
     if _separator is None:
         from src.third_party.MusicSourceSeparationTraining.inference_api import (

initialization.py

@@ -0,0 +1,73 @@
+"""
+YingMusic-Singer Initialization Script
+
+Downloads required checkpoints from HuggingFace based on task type.
+
+Usage:
+    python initialization.py --task infer
+    python initialization.py --task train
+"""
+
+import argparse
+import os
+
+from huggingface_hub import hf_hub_download
+
+REPO_ID = "ASLP-lab/YingMusic-Singer"
+CKPT_DIR = "ckpts"
+
+# Files required for each task
+INFER_FILES = [
+    "ckpts/MelBandRoformer.ckpt",
+    "ckpts/config_vocals_mel_band_roformer_kj.yaml",
+]
+
+TRAIN_EXTRA_FILES = [
+    "ckpts/YingMusicSinger_model.pt",
+    "ckpts/model_ckpt_steps_100000_simplified.ckpt",
+    "ckpts/stable_audio_2_0_vae_20hz_official.ckpt",
+]
+
+TASK_FILES = {
+    "infer": INFER_FILES,
+    "train": INFER_FILES + TRAIN_EXTRA_FILES,
+}
+
+
+def download_files(task: str):
+    files = TASK_FILES[task]
+    os.makedirs(CKPT_DIR, exist_ok=True)
+
+    print(f"Task: {task} | Downloading {len(files)} file(s) to {CKPT_DIR}/")
+    for remote_path in files:
+        filename = os.path.basename(remote_path)
+        local_path = os.path.join(CKPT_DIR, filename)
+
+        if os.path.exists(local_path):
+            print(f"  [skip] {filename} already exists")
+            continue
+
+        print(f"  [download] {filename} ...")
+        hf_hub_download(
+            repo_id=REPO_ID,
+            filename=remote_path,
+            local_dir=".",
+        )
+        print(f"  [done] {filename}")
+
+    print("All downloads complete.")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Download YingMusic-Singer checkpoints"
+    )
+    parser.add_argument(
+        "--task",
+        type=str,
+        required=True,
+        choices=list(TASK_FILES.keys()),
+        help="Task type: 'infer' for inference, 'train' for training",
+    )
+    args = parser.parse_args()
+    download_files(args.task)