Files
songforge/tools/make_style_ref.py
2026-10-09 19:18:33 +08:00

183 lines
6.5 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# -*- coding: utf-8 -*-
"""用历史已认可作品建风格质心语料 -> config/baseline/style_ref.json
跑法(必须在 ComfyUI venv 的 Python 下):
C:\\comfyui-venv\\Scripts\\python.exe tools\\make_style_ref.py
音乐目录 / 输出路径读 config/default.json + local.json(勿写死盘符)。
"""
import json
import os
import re
import sys
HERE = os.path.dirname(os.path.abspath(__file__))
# 坑位:songforge.py(CLI 入口)和 songforge/(包)同名,
# tools/ 内部互相引用一律走文件路径加载。
sys.path.insert(0, HERE)
import qc_style # noqa: E402
_PKG = os.path.dirname(HERE)
def _paths_from_config():
paths = {}
for name in ("default.json", "local.json"):
p = os.path.join(_PKG, "config", name)
if not os.path.isfile(p):
continue
with open(p, encoding="utf-8") as f:
blob = json.load(f)
paths.update(blob.get("paths") or {})
music = paths.get("music_dir") or os.environ.get("SONGFORGE_MUSIC_DIR")
if not music:
cro = paths.get("comfy_root") or os.environ.get("SONGFORGE_COMFY_ROOT")
if cro:
music = os.path.join(
cro, "output", paths.get("output_prefix") or "MyMusic")
if not music:
music = os.path.join(
os.path.dirname(os.path.dirname(_PKG)), "output", "MyMusic")
out = os.path.join(_PKG, "config", "baseline", "style_ref.json")
cache = os.path.join(_PKG, "cache", "style")
return os.path.abspath(music), out, cache
MUSIC, OUT, CACHE = _paths_from_config()
# 前缀白名单:确认过的直接给 family,避免子串规则误伤
EXACT = {
"smile13": "pop",
"hit10": "pop",
"dy10": "pop",
"col2": "pop",
"punk4": "punk",
"niaopunk": "punk",
"urbanhiphop": "hiphop",
"chillhiphop": "hiphop",
"chillhiphopen": "hiphop",
"chillhiphopja": "hiphop",
"upbeathiphop": "hiphop",
"funkyhiphop": "hiphop",
"discofunk": "hiphop",
"partytrap": "hiphop",
"smoothhiphop": "hiphop",
"喜剧hiphop": "hiphop",
"欧式hiphop": "hiphop",
}
# 出现在前缀里就归某家族(前面已排除精确表未命中的)
SUBSTR = (
("punk", "punk"),
("hiphop", "hiphop"),
("trap", "hiphop"),
)
SEED = re.compile(r"_\d{5}\.flac$", re.I)
# 只收每个 seed 序列的第一首。同源的 _00002/_00003 全收进来,语料就被
# 少数几首歌按 seed 数投票,质心被拉偏(踩过:172 个文件其实只有 30 来首歌)。
FIRST_SEED = re.compile(r"_00001\.flac$", re.I)
def family_of(name):
"""从文件名解析风格家族,解析不出来返回 None(直接排除)。"""
stem = SEED.sub("", name)
head = stem.split("_")[0]
if head in EXACT:
return EXACT[head]
low = head.lower()
for key, fam in SUBSTR:
if key in low:
return fam
return None
def main():
if not os.path.isdir(MUSIC):
print("[错误] 找不到音频目录 %s" % MUSIC)
return 2
files = sorted(f for f in os.listdir(MUSIC) if f.lower().endswith(".flac"))
picked = []
for f in files:
fam = family_of(f)
if not fam:
continue
if not FIRST_SEED.search(f): # 只收每首歌的第一个 seed
continue
p = os.path.join(MUSIC, f)
if not os.path.isfile(p):
continue
picked.append((f, fam, p))
if not picked:
print("[错误] 一首都没匹配上,检查 EXACT / SUBSTR 是否还匹配文件名")
return 2
print("语料候选 %d 首,开始算特征(每首几秒)..." % len(picked))
os.makedirs(CACHE, exist_ok=True)
items = []
for i, (f, fam, p) in enumerate(picked, 1):
cp = os.path.join(CACHE, f + ".json")
# 缓存 key 带上音频 mtime + 特征列表,音频换了或特征改了都会自动失效
stamp = "%s|%s" % (os.path.getmtime(p), ",".join(qc_style.FEATURES))
vec = None
if os.path.isfile(cp):
try:
with open(cp, encoding="utf-8") as fh:
blob = json.load(fh)
if blob.get("key") == stamp:
vec = blob["vec"]
except Exception:
vec = None
if vec is None:
try:
# expect_bpm=None:历史文件没有可靠的 BPM 记录,靠自身中位数定折叠锚点,
# 保证语料里的 tempo 尺度自洽。新歌的 BPM 对齐由 qc_worker 单独判。
vec, _raw = qc_style.features(p, expect_bpm=None)
except Exception as exc:
print(" [%d/%d] 跳过 %s -> %s: %s"
% (i, len(picked), f, type(exc).__name__, exc))
continue
with open(cp, "w", encoding="utf-8") as fh:
json.dump(dict(key=stamp, vec=vec), fh)
items.append(dict(file=f, family=fam, vec=vec))
print(" [%d/%d] %-44s %-6s tempo=%5.1f flat=%.4f"
% (i, len(picked), f[:44], fam, vec["tempo"], vec["flatness"]))
counts = {}
for it in items:
counts[it["family"]] = counts.get(it["family"], 0) + 1
print("按家族统计:%s" % counts)
for fam in qc_style.FAMILIES:
if counts.get(fam, 0) < 3:
print("[错误] %s 只有 %d 首,建不出可信质心(至少 3)"
% (fam, counts.get(fam, 0)))
return 2
ref = qc_style.build(items)
with open(OUT, "w", encoding="utf-8") as f:
json.dump(ref, f, ensure_ascii=False, indent=1)
loo = ref["loo"]
print("")
print("已写出 %s" % OUT)
print("样本 %d(%d 首不同的歌),分组留一准确率 %.1f%%,trusted=%s"
% (len(items), loo["n_songs"], loo["acc"] * 100, ref["trusted"]))
print("分家族:%s" % json.dumps(loo["per_family"], ensure_ascii=False))
print("混淆矩阵(行=真实,列=判成):")
for fam, row in sorted(loo["confusion"].items()):
print(" %-7s %s" % (fam, json.dumps(row, ensure_ascii=False)))
if not ref["trusted"]:
print("")
print("[结论] 风格判定不可信,只写进报告当参考,不当放行闸门。")
print(" 风格的第一物理保证是实测 BPM 与设定 BPM 对齐(QC 的 bpm_rel_tol)。")
print(" 注意:语料里 pop 的 chill/disco/funk 变体和 hiphop 制作手法本就连续,")
print(" 分组 LOO 总体约 60~67%、单家族最低约 43%,"
"这是语料重叠,不是分类器没调好。")
return 0
if __name__ == "__main__":
sys.exit(main())