795 lines
40 KiB
Python
795 lines
40 KiB
Python
# -*- coding: utf-8 -*-
|
||
"""全链路自检(不连 ComfyUI 生成)。"""
|
||
import os
|
||
import re
|
||
import sys
|
||
|
||
from ..config import loader
|
||
from ..core import paths as pmod
|
||
from ..core import workflow as wfmod
|
||
from ..lyrics import meter
|
||
from ..rules import ironlaw
|
||
from .assemble import assemble, do_validate
|
||
from .common import ROOT, now_id
|
||
|
||
# 第三张专辑的低音吟唱 demo 歌词,稀疏档与 chill 集成自测都读它。
|
||
# 素材库迁到 _archive/lyrics_corpus 后,selftest 仍读包内 probe。
|
||
# 旧文件名 02_InstLowChill.txt 已不在树里,以现有 03 为准。
|
||
_CHILL_DEMO = os.path.join(
|
||
ROOT, "lyrics", "probe", "03_InstLowChill6.txt")
|
||
|
||
|
||
def _t_runid():
|
||
"""运行 ID 必须唯一。
|
||
|
||
旧实现只到秒,连跑三首会落进同一个 runs 目录,lyrics.txt / tags.txt /
|
||
manifest.json 互相覆盖,dry-run 记录直接失效。这里既验唯一性,
|
||
也验真实目录里连续三次分配不冲突。
|
||
"""
|
||
layout = pmod.path_layout(loader.load({}))
|
||
scratch = os.path.join(layout["tmp"], "selftest_runid")
|
||
if os.path.isdir(scratch):
|
||
for f in os.listdir(scratch):
|
||
p = os.path.join(scratch, f)
|
||
if os.path.isdir(p):
|
||
os.rmdir(p)
|
||
os.rmdir(scratch)
|
||
os.makedirs(scratch)
|
||
try:
|
||
ids = [now_id("t", scratch) for _ in range(3)]
|
||
# 目录必须已被原子占位,防止后续写入互相覆盖
|
||
assert all(os.path.isdir(os.path.join(scratch, i)) for i in ids), \
|
||
"运行目录未占位: %s" % ids
|
||
finally:
|
||
for f in os.listdir(scratch):
|
||
os.rmdir(os.path.join(scratch, f))
|
||
os.rmdir(scratch)
|
||
assert len(set(ids)) == 3, "运行 ID 冲突: %s" % ids
|
||
|
||
# 模拟目录已存在时必须自增,而不是覆盖旧记录
|
||
base = now_id("t", None)
|
||
assert len(base) > len("t-20261004-041513"), \
|
||
"运行 ID 缺少微秒段: %s" % base
|
||
root2 = os.path.join(layout["tmp"], "selftest_runid2")
|
||
if os.path.isdir(root2):
|
||
for f in os.listdir(root2):
|
||
os.rmdir(os.path.join(root2, f))
|
||
os.rmdir(root2)
|
||
os.makedirs(root2)
|
||
try:
|
||
# now_id 已原子占位,这里再取一次必须自动自增,而不是复用旧目录
|
||
first = now_id("t", root2)
|
||
second = now_id("t", root2)
|
||
finally:
|
||
for d in os.listdir(root2):
|
||
os.rmdir(os.path.join(root2, d))
|
||
os.rmdir(root2)
|
||
assert first != second, "目录已存在时未自增: %s / %s" % (first, second)
|
||
return [("运行 ID 唯一", "连跑三次 %s ~ %s 互不冲突" % (ids[0], ids[-1]))]
|
||
|
||
|
||
def _t_path_layout():
|
||
"""配置里的目录键必须能解析,且与 DEFAULTS 对齐。"""
|
||
cfg = loader.load({})
|
||
p = cfg["paths"]
|
||
for key in ("runs_dir", "tmp_dir", "lyrics_dir", "qc_cache_dir",
|
||
"style_cache_dir", "output_prefix", "baseline_dir",
|
||
"catalog_file"):
|
||
assert key in p and p[key], "paths.%s 缺失" % key
|
||
layout = pmod.path_layout(cfg)
|
||
runs_n = layout["runs"].replace("\\", "/")
|
||
tmp_n = layout["tmp"].replace("\\", "/")
|
||
assert runs_n.replace("\\", "/").endswith("/data/runs") or \
|
||
runs_n.endswith("data/runs"), "runs 路径异常: %s" % runs_n
|
||
assert tmp_n.replace("\\", "/").endswith("/data/tmp") or \
|
||
tmp_n.endswith("data/tmp"), "tmp 路径异常: %s" % tmp_n
|
||
assert os.path.isdir(layout["baseline"]), "baseline 目录缺失: %s" % layout["baseline"]
|
||
assert os.path.isfile(layout["template"]), "模板缺失: %s" % layout["template"]
|
||
for name in ("qc_baseline.json", "qc_baseline_old.json", "style_ref.json"):
|
||
assert os.path.isfile(os.path.join(layout["baseline"], name)), \
|
||
"缺少基线文件 %s" % name
|
||
return [("目录布局", "仓根 data/ + baseline + 包内模板 已解析")]
|
||
|
||
|
||
def _t_paths():
|
||
root = pmod.comfy_root()
|
||
assert os.path.isdir(os.path.join(root, "models", "checkpoints")), \
|
||
"ComfyUI 根目录推导错误: %s" % root
|
||
return [("ComfyUI 根推导", root)]
|
||
|
||
|
||
def _t_config_priority():
|
||
os.environ["SONGFORGE_SHIFT"] = "4.4"
|
||
try:
|
||
cfg = loader.load({})
|
||
assert abs(cfg["sampler"]["shift"] - 4.4) < 1e-6, "环境变量未生效"
|
||
cfg = loader.load({"sampler.shift": 6.0})
|
||
assert abs(cfg["sampler"]["shift"] - 6.0) < 1e-6, "命令行未覆盖环境变量"
|
||
finally:
|
||
os.environ.pop("SONGFORGE_SHIFT", None)
|
||
base = loader.load({})["sampler"]["shift"]
|
||
assert abs(base - 5.5) < 1e-6, "默认 shift 被污染: %s" % base
|
||
return [("配置优先级 命令行>环境>配置", "shift 5.5 -> 4.4 -> 6.0")]
|
||
|
||
|
||
def _t_model_locked():
|
||
cfg = loader.load({})
|
||
ck = cfg["model"]["checkpoint"]
|
||
assert ck == "ace_step_1.5_turbo_swl_fused.safetensors", "checkpoint 被换: %s" % ck
|
||
assert cfg["model"]["lora"] == "swl_voice_best_ep257.safetensors", "LoRA 被换"
|
||
assert ck not in cfg["model"]["denied_checkpoints"], "当前模型在禁用名单"
|
||
return [("模型锁定", ck + " + " + cfg["model"]["lora"])]
|
||
|
||
|
||
def _t_ironlaw():
|
||
tags = ironlaw.build_tags("Chinese pop")
|
||
for kw in ("swlvoice", "8 bar instrumental intro hook", "no opening vocals",
|
||
"no pure vocal intro", "no humming"):
|
||
assert kw in tags, "tags 缺少 %s" % kw
|
||
lyrics = ironlaw.build_lyrics(
|
||
[("[verse]", "第一句歌词\n第二句歌词")], "kick and snare", bars=8)
|
||
assert lyrics.lstrip().startswith("[intro]"), "未强制器乐前奏"
|
||
assert "[Instrumental: 8 bar" in lyrics, "前奏未写明 8 小节"
|
||
assert not ironlaw.check_rules(lyrics, tags), ironlaw.check_rules(lyrics, tags)
|
||
# 缺前奏必须被拦
|
||
no_intro = "[verse]\n没有前奏直接唱"
|
||
assert ironlaw.check_rules(no_intro, tags), "缺器乐前奏未被拦截"
|
||
# 中文结构标记必须被拦
|
||
marked = lyrics.replace("[verse]", "[主歌]")
|
||
assert ironlaw.check_rules(marked, tags), "中文结构标记未被拦截"
|
||
return [("铁律注入", "8 小节器乐前奏 + 反人声条款齐备,缺前奏/中文标记均被拦")]
|
||
|
||
|
||
def _t_sparse_profile():
|
||
"""稀疏档必须真的放行"只有几句英文语气词"的写法,且仍然拦得住异常值。
|
||
|
||
背景:第三张专辑人声压过音乐,根因之一是 default 密度档按"人声填满全曲"
|
||
反推(175s/124BPM 下限 254 音节),把 84 音节的低音吟唱判成"歌词过少",
|
||
于是每次都被迫写满整首,人声自然盖过器乐。sparse 档就是为了给出
|
||
"可唱小节只占 25%" 的合法区间。
|
||
"""
|
||
cfg = loader.load({})["lyrics"]
|
||
sparse = meter.pick(cfg, "sparse")
|
||
assert sparse["sing_bar_ratio"] < cfg["sing_bar_ratio"], \
|
||
"sparse 档可唱小节占比没有低于 default,音乐主导无从体现"
|
||
assert sparse["ideal_line_syllable"][0] < cfg["ideal_line_syllable"][0], \
|
||
"sparse 档句子下限没有放宽,mmm / ooh 这类短句仍会被拦"
|
||
|
||
dur, bpm = 175, 124
|
||
_, lo, hi = meter.target_words(dur, bpm, sparse)
|
||
# 直接用真实 demo 歌词做样本,而不是抄一份内联的:
|
||
# 抄的内联文本一漂移,测试就测不到真正要生成的那份文件了。
|
||
probe = _CHILL_DEMO
|
||
assert os.path.exists(probe), "找不到低音 demo 歌词 %s" % probe
|
||
with open(probe, encoding="utf-8") as f:
|
||
body = f.read()
|
||
syl = meter.syllable(meter.only_lyrics(body))
|
||
assert meter.check(body, dur, bpm, cfg)[0], \
|
||
"default 档居然放行了稀疏歌词,档间没有区分度"
|
||
errs, st = meter.check(body, dur, bpm, sparse)
|
||
assert not errs, "sparse 档误杀了低音语气词写法: %s" % errs
|
||
assert lo <= syl <= hi, \
|
||
"sparse 放行区间 %d~%d 与实测 %d 音节不一致" % (lo, hi, syl)
|
||
|
||
# 过短仍要拦:sparse 不是"无限制放宽"
|
||
thin = ("[intro]\n[Instrumental: 8 bar]\n[verse]\nmmm\nooh\n")
|
||
assert any("过少" in e for e in meter.check(thin, dur, bpm, sparse)[0]), \
|
||
"sparse 档放过短歌词"
|
||
|
||
# 过多仍要拦:写成满首人声时 sparse 也必须报错,否则会退回旧病
|
||
fat = ("[intro]\n[Instrumental: 8 bar]\n[verse]\n"
|
||
+ "\n".join(["mmm mmm low and slow"] * 90))
|
||
assert any("过多" in e for e in meter.check(fat, dur, bpm, sparse)[0]), \
|
||
"sparse 档放过多歌词,音乐主导会被稀释"
|
||
return [("稀疏密度档",
|
||
"sparse 放行 %d~%d 音节(default 拒绝同一份),过少过多仍被拦"
|
||
% (lo, hi))]
|
||
|
||
|
||
def _t_instrumental():
|
||
"""纯器乐模式的标签必须自洽,且不与风格字典的唱法描述对撞。"""
|
||
tags = ironlaw.build_tags(
|
||
style="English melodic future bass", instruments="huge supersaw lead",
|
||
mood="vast uplifting EDM",
|
||
extra="male low whisper vocal, hushed and intimate",
|
||
instrumental=True)
|
||
# 反人声条款必须在
|
||
for kw in ("no vocals anywhere", "instrumental only", "no voice at all",
|
||
"no humming", "no choir", "no vocal sample", "swlvoice",
|
||
"8 bar instrumental intro hook"):
|
||
assert kw in tags, "纯器乐 tags 缺少 %s" % kw
|
||
# 关键回归点:风格字典的唱法描述绝不能混进来跟 "no vocals" 对撞。
|
||
for bad in ("whisper vocal", "sparse English words", "no belting"):
|
||
assert bad not in tags, "纯器乐 tags 混入了唱法描述: %s" % bad
|
||
# 器乐正向要求必须在(这里只校验铁律常量层,风格字典自带的配器词不算)
|
||
for kw in ("long instrumental passages", "extended instrumental breakdown",
|
||
"huge powerful drop", "instrumental sections are the main focus",
|
||
"drums and bass carry the track"):
|
||
assert kw in tags, "纯器乐 tags 缺少器乐主导描述 %s" % kw
|
||
# 绝不能出现把器乐段填满的条款(这会让模型加人声)
|
||
assert "no long empty instrumental" not in tags, \
|
||
"纯器乐 tags 含 no long empty instrumental,会逼模型填人声"
|
||
assert "no long empty instrumental" not in ironlaw.RULE_CHILL_MEMO, \
|
||
"吟唱 tags 含 no long empty instrumental,会逼模型填人声"
|
||
# 前奏不能是"先器乐后人声"那句:纯器乐里它等于命令模型把人声加回来
|
||
assert "vocals enter only after" not in tags, \
|
||
"纯器乐 tags 仍含 'vocals enter only after the instrumental hook'"
|
||
assert "no breakdown" not in tags, \
|
||
"纯器乐 tags 含 no breakdown,与 extended instrumental breakdown 冲突"
|
||
# 歌词体必须是纯结构标记,没有可唱正文
|
||
lyrics = ironlaw.build_inst_lyrics("huge supersaw", bars=8)
|
||
assert lyrics.lstrip().startswith("[intro]"), "纯器乐歌词未以 [intro] 开头"
|
||
body = meter.only_lyrics(lyrics)
|
||
assert body.strip() == "", "纯器乐歌词体出现可唱正文: %r" % body[:80]
|
||
assert meter.syllable(body) == 0, "纯器乐歌词体音节数非 0"
|
||
# 结构段名不能是 [verse] / [chorus]:ACE-Step 见到这两个标记就会按
|
||
# "有人唱的段落"来铺,即使下面只挂 [inst] 也容易把人声请回来。
|
||
for bad_tag in ("[verse]", "[chorus]"):
|
||
assert bad_tag not in lyrics.lower(), \
|
||
"纯器乐歌词体含 %s,会重新诱导人声" % bad_tag
|
||
for good_tag in ("[drop]", "[breakdown]", "[build]"):
|
||
assert good_tag in lyrics.lower(), "纯器乐歌词体缺少 %s 段" % good_tag
|
||
# 结构铁律照旧生效
|
||
assert not ironlaw.check_rules(lyrics, tags), ironlaw.check_rules(lyrics, tags)
|
||
# 密度闸门必须让行纯器乐
|
||
cfg = loader.load({})
|
||
errs, _ = do_validate(lyrics, tags, cfg,
|
||
{"duration": 175, "bpm": 124}, {}, strict=False,
|
||
instrumental=True)
|
||
assert not errs, "纯器乐被密度闸门拦下: %s" % errs
|
||
# chill 模式的稳定条款也不能再要求 no breakdown
|
||
chill_tags = ironlaw.build_tags(style="future bass",
|
||
extra="male low whisper vocal",
|
||
chill_vocal=True)
|
||
assert "no breakdown" not in chill_tags, \
|
||
"chill tags 含 no breakdown,与 quiet breakdown 冲突"
|
||
# 声线稳定性约束已被实测证明是反效果:标签里每多一句谈人声,
|
||
# 人声越从 0 秒就进。chill 现在只保留乐器/调性的稳定性要求。
|
||
assert "consistent vocal timbre" not in chill_tags, \
|
||
"chill tags 又开始强调人声稳定性,实测会退回 0 秒开唱"
|
||
assert "consistent instrumentation in every section" in chill_tags, \
|
||
"chill tags 丢了配器稳定性约束"
|
||
return [("纯器乐模式", "反人声标签自洽、无唱法对撞、无 no breakdown、"
|
||
"结构体不诱发人声、密度闸门让行")]
|
||
|
||
|
||
def _t_chill_delay():
|
||
"""chill 模式:验证标签与结构两侧自洽(不退化成纯器乐,也不退回 0 秒开唱)。
|
||
|
||
五轮实测踩出来的结论:
|
||
chill 1 8 小节前奏 + "vocals enter only after…"
|
||
-> vocals stem 从 0.0 秒就有能量
|
||
chill 2 标签加长到 16 小节 + "first 20 seconds are entirely instrumental"
|
||
+ "the human voice does not enter until the second drop"
|
||
-> 第一句仍然落 0.0 秒,标签侧时间约束全部无效
|
||
chill 3 结构重排 + 强人声描述
|
||
-> 稀疏低音语气词,能量 +45~50 dB,歌词重合 0.98
|
||
(当前唯一确认真唱的一版)
|
||
chill 4 标签彻底不讨论人声(复用纯器乐反人声条款)
|
||
-> 人声归零,等同纯器乐
|
||
chill 5 只保留音色质感,删掉密度/时机描述
|
||
-> 人声同样归零。此前把 ASR 的 4 段当成"少量真唱"是误判:
|
||
文本 "oh new healing / lömpe Minor" 歌词里根本没有,
|
||
能量只 +11 dB、ASR 覆盖率 100%,是持续 pad 幻听。
|
||
|
||
关键结论:标签侧写时长/时机对人声起点零作用;唯一旋钮是
|
||
"密度与时机类人声描述的多少",而它降下去不是"延后"而是"消失"。
|
||
所以自测只能守住"结构占位存在 + 标签不退回纯器乐",
|
||
人声起点必须靠音频 QC(ab_vocal_curve.py / ab_is_real_voice.py)判定。
|
||
"""
|
||
tags = ironlaw.build_tags(style="English melodic future bass",
|
||
instruments="huge supersaw lead",
|
||
mood="vast uplifting emotional EDM",
|
||
extra="male low whisper vocal, "
|
||
"hushed and intimate, sparse English words",
|
||
chill_vocal=True)
|
||
# 人声仍然允许存在,不能把 chill 误改成纯器乐
|
||
for kw in ("male low whisper vocal", "hushed and intimate",
|
||
"low register vocal texture", "soft breathy phrases",
|
||
"instrumental sections are the main focus",
|
||
"long instrumental passages",
|
||
"instrumental breakdown and instrumental drop sections"):
|
||
assert kw in tags, "chill tags 缺少器乐主导描述 %s" % kw
|
||
# 人声禁落器乐段
|
||
for kw in ("no singing in any breakdown", "no singing over the first drop",
|
||
"no long vocal phrases", "no choir"):
|
||
assert kw in tags, "chill tags 缺少段落归属约束 %s" % kw
|
||
# chill 不能退回普通前奏规则:那句只说"稍后进人声",实测拦不住
|
||
assert "vocals enter only after the instrumental hook has played" not in tags, \
|
||
"chill tags 退回 RULE_INTRO,实测 0 秒开唱"
|
||
|
||
# ---- 五轮实测收敛出来的核心断言 ----
|
||
# InstLowChill / 2 / 3 / 4 / 5 五版量出一条干净的梯度,证明 prompt 对
|
||
# 人声的控制力是真实存在的,旋钮 = "标签里有多少句在描述人声":
|
||
# chill 2 强人声描述 + 16 小节前奏
|
||
# -> ASR 23 段,voc/acc 0.483,人声压过音乐
|
||
# chill 3 结构重排 + 强人声描述
|
||
# -> 稀疏低音语气词,voc/acc 0.376,唯一确认真唱的一版
|
||
# chill 4 标签彻底不讨论人声(复用纯器乐反人声条款)
|
||
# -> 人声归零,等同纯器乐
|
||
# chill 5 只留音色质感,删掉密度/时机描述
|
||
# -> 人声同样归零(ASR 全是持续 pad 幻听)
|
||
#
|
||
# 时机类措辞必须继续禁掉:chill 2/3 都写了
|
||
# "second half / enters late / first 20 seconds",第一句全部落 0.0 秒。
|
||
for bad in ("voice enters", "the human voice", "no voice in the first",
|
||
"until the second drop", "in the second half of the track",
|
||
"first 20 seconds"):
|
||
assert bad not in tags.lower(), \
|
||
"chill tags 又开始讨论人声进场时机(%s),实测会退回 0 秒开唱" % bad
|
||
|
||
# 密度类措辞必须保留:这是 chill 3 唯一产生真唱的东西,
|
||
# 删掉(chill 5)人声直接归零,不是延后而是消失。
|
||
for kw in ("voice is sparse and low volume", "used only as sparse texture",
|
||
"sparse english words", "mostly instrumental between phrases"):
|
||
assert kw in tags.lower(), \
|
||
"chill tags 丢了密度触发词 %s,实测会退回纯器乐" % kw
|
||
|
||
# 风格字典的 extra 绝不能混进来(六句里五句在谈人声)。
|
||
# 这里不再靠关键词黑名单(因为 chill 自己现在也用同几个密度词,
|
||
# 黑名单会误伤),改用哨兵值:故意传一个只在 extra 里的短语,
|
||
# 结果里出现就说明 extra 泄漏了。
|
||
probe = "zzz-extra-leak-probe"
|
||
tags_probe = ironlaw.build_tags(style="s",
|
||
instruments="i", mood="m",
|
||
extra=probe,
|
||
chill_vocal=True)
|
||
assert probe not in tags_probe.lower(), "chill 分支把风格字典 extra 混进来了"
|
||
|
||
# 但也不能把人声彻底禁掉,否则就是纯器乐、不是吟唱。
|
||
# 注意只匹配纯器乐专属的反人声条款:
|
||
# - 不能用裸 "no singing":"no singing in any breakdown" 含这个子串,
|
||
# 那是吟唱要的人声约束。
|
||
# - 不能用 "instrumental only":它来自 "intro is instrumental only",
|
||
# 那是前奏必须保留的约束。
|
||
for kw in ("no vocals anywhere", "no voice at all", "no lead vocal",
|
||
"no backing vocal", "this is a backing track",
|
||
"drums bass synths chords and lead melodies only"):
|
||
assert kw not in tags.lower(), \
|
||
"chill tags 含纯器乐条款 %s,吟唱被人声禁令吞掉" % kw
|
||
|
||
return [("吟唱标签梯度",
|
||
"时机类措辞全部禁掉,密度触发词保留,"
|
||
"风格字典 extra 未泄漏")]
|
||
|
||
|
||
def _t_groove_first_beat():
|
||
"""器乐前奏必须从第 0 秒就有底鼓和贝斯,不能是空场。
|
||
|
||
这里有一个被实测推翻过的旧结论,必须留在注释里免得又走回去:
|
||
上一轮认定"前 15 秒是空场",依据是 ab_lowfreq.py 报的
|
||
-68~-75 dBFS。那份数据是错的——它把滤波器的瞬态响应和单帧跳变
|
||
当成了整段电平。ab_bands.py / ab_head_rhythm.py 复测(dBFS,
|
||
head = 前 16 秒,mid = 16~32 秒):
|
||
|
||
GroovePure head sub -66.0 / mid sub -55.8
|
||
head full -32.7 / mid full -25.0
|
||
|
||
也就是说开头有音乐(full band 只低 7.7 dB,hat/snare 瞬态齐备),
|
||
缺的只是 sub/kick。所以真正要修的是"前奏要有节奏组",
|
||
不是"开头别是空场",而"空场"这个说法本身就是误判。
|
||
|
||
修法(见 ironlaw.intro_rhythm_only):前奏括号只保留节奏组,
|
||
剔除 pad / riser / reverb / lead。这一条比堆禁令有效——
|
||
GroovePure 已经证明四条禁令(no ambient intro / no pad only intro /
|
||
no beatless intro / no white noise riser in the opening)全部无效。
|
||
"""
|
||
for flag, kw in ((False, "instrumental"), (True, "chill")):
|
||
tags = ironlaw.build_tags(style="English melodic future bass",
|
||
instruments="huge supersaw lead",
|
||
mood="vast uplifting emotional EDM",
|
||
chill_vocal=flag,
|
||
instrumental=not flag)
|
||
for need in ("full four on the floor groove from the very first beat",
|
||
"kick drum, sub bass and walking bass on bar one",
|
||
"no pad only intro", "no beatless intro",
|
||
"no riser in the opening"):
|
||
assert need in tags.lower(), \
|
||
"%s tags 缺第一拍 groove 约束: %s" % (kw, need)
|
||
|
||
# 普通有声歌曲也一并受益:RULE_INTRO 同样要保证鼓从第一拍进。
|
||
assert "rhythmic drums from the very first beat" in \
|
||
ironlaw.build_tags("test").lower(), "普通歌曲前奏丢了第一拍底鼓"
|
||
|
||
return [("第一拍 groove",
|
||
"chill/纯器乐都要求底鼓+贝斯从第 0 拍进,"
|
||
"并点名禁掉 pad only / beatless / riser 空场信号")]
|
||
|
||
|
||
def _t_intro_instrument_filter():
|
||
"""前奏括号必须过滤掉 pad/riser/reverb/lead,只留节奏组。
|
||
|
||
这是本轮真正起作用的改动。上一轮只在标签里堆禁令,GroovePure 实测
|
||
无效(开头 sub 仍 -66.0 dBFS,中段 -55.8 dBFS),因为真正的病灶是
|
||
歌词体那一行:
|
||
[Instrumental: 8 bar instrumental intro hook, huge supersaw lead,
|
||
sidechain compressed pad, punchy electro kick, crisp clap,
|
||
trap snare, white noise riser, big reverb, melodic drop, ...]
|
||
空场信号(lead/pad/riser/reverb)4 个,节奏组(kick/clap/snare)3 个,
|
||
3:4 输了,模型就把前奏铺成 ambient。
|
||
|
||
所以这里守三件事:
|
||
1) 过滤后必须还留着 kick(不能把节奏组也误删成空括号);
|
||
2) pad / riser / reverb / lead 一个都不能留在前奏括号里;
|
||
3) 标签侧的 instruments 段保持原样——drop 段本来就要 pad 和 lead,
|
||
只过滤前奏括号,不能把整首歌的配器改掉。
|
||
"""
|
||
src = ("huge supersaw lead, sidechain compressed pad, "
|
||
"punchy electro kick, crisp clap, trap snare, "
|
||
"white noise riser, big reverb, melodic drop")
|
||
got = ironlaw.intro_rhythm_only(src)
|
||
low = got.lower()
|
||
|
||
assert "kick" in low, "前奏括号把底鼓也过滤掉了: %r" % got
|
||
for bad in ("pad", "riser", "reverb", "supersaw", "lead", "noise"):
|
||
assert bad not in low, "前奏括号残留空场信号 %s: %r" % (bad, got)
|
||
|
||
# 加法兜底:kick 和 bass 两条必须显式出现在前奏括号里。
|
||
# 上一轮漏了这一条,风格字典里一条贝斯都没写,过滤后只剩三条中高频,
|
||
# 结果两版新音频开头 sub 比全曲峰值低 53.5 / 31.2 dB。
|
||
assert "bass" in low, "前奏括号没有 bass:%r" % got
|
||
|
||
# 空 / 中文 / 全被过滤掉 -> 必须回落到默认节奏组,不能留空括号
|
||
for probe in ("", None, "ambient pad, huge supersaw lead, big reverb"):
|
||
fb = ironlaw.intro_rhythm_only(probe)
|
||
assert "kick" in fb.lower(), "前奏括号回落失败 %r -> %r" % (probe, fb)
|
||
assert "bass" in fb.lower(), "回落值缺 bass %r -> %r" % (probe, fb)
|
||
for bad in ("pad", "riser", "reverb", "supersaw", "lead"):
|
||
assert bad not in fb.lower(), "回落值带空场信号 %s: %r" % (bad, fb)
|
||
|
||
# 回落值不能是 DEFAULT_INSTRUMENT:它含 "chord stabs",而 "chord stab"
|
||
# 就在 _RHYTHM_DROP 里,回落会把刚剔除掉的空场信号塞回前奏。
|
||
assert "stab" not in ironlaw.DEFAULT_INTRO_RHYTHM.lower(), \
|
||
"前奏回落值含 chord stab,与过滤规则自相矛盾"
|
||
|
||
# 标签侧不受影响:instruments 段必须还是原样
|
||
tags = ironlaw.build_tags(style="s", instruments=src, mood="m",
|
||
instrumental=True)
|
||
assert "white noise riser" in tags, \
|
||
"过滤逻辑泄漏到了标签侧,drop 段会丢掉 riser"
|
||
assert "sidechain compressed pad" in tags, \
|
||
"过滤逻辑泄漏到了标签侧,drop 段会丢掉 pad"
|
||
|
||
# 歌词体里前奏那一行必须既没有 riser/pad,又必须有 bass
|
||
lyr = ironlaw.build_inst_lyrics(src, bars=8)
|
||
head = lyr.split("\n")[1].lower()
|
||
assert "riser" not in head and "pad" not in head, \
|
||
"纯器乐歌词体前奏行仍带空场信号: %s" % head
|
||
assert "bass" in head, "纯器乐歌词体前奏行缺 bass: %s" % head
|
||
lyr2 = ironlaw.build_chill_lyrics([("[verse]", "ooh")], src, bars=8,
|
||
total_bars=90)
|
||
head2 = lyr2.split("\n")[1].lower()
|
||
assert "riser" not in head2 and "pad" not in head2, \
|
||
"吟唱歌词体前奏行仍带空场信号: %s" % head2
|
||
assert "bass" in head2, "吟唱歌词体前奏行缺 bass: %s" % head2
|
||
|
||
return [("前奏配器过滤",
|
||
"前奏括号剔除 pad/riser/reverb/lead,并强制补齐 kick+bass,"
|
||
"标签侧配器保持原样")]
|
||
|
||
|
||
def _t_chill_no_conflict():
|
||
"""chill 标签里绝不能出现和"低音英文语气词"直接冲突的禁令。
|
||
|
||
这三条(no humming / no mm mm syllables / no oh oh syllables)第一次改
|
||
chill 铁律就是为了解开这个冲突,加回来等于自己把要求唱的内容禁掉。
|
||
"""
|
||
tags = ironlaw.build_tags(style="s", instruments="i", mood="m",
|
||
chill_vocal=True)
|
||
for bad in ("no humming", "no mm mm syllables", "no oh oh syllables",
|
||
"no droning sustained vowels"):
|
||
assert bad not in tags.lower(), \
|
||
"chill tags 含 %s,与要求的低音语气词直接冲突" % bad
|
||
# 反过来,"稀疏低音语气词"必须真的在标签里,否则就退化成纯器乐了。
|
||
for need in ("male low whisper vocal", "low register vocal texture",
|
||
"soft breathy phrases"):
|
||
assert need in tags.lower(), "chill tags 丢了低音唱腔描述: %s" % need
|
||
|
||
return [("吟唱无冲突",
|
||
"no humming / no mm mm / no oh oh 三条禁令未回归")]
|
||
|
||
|
||
def _t_chill_integration():
|
||
"""chill 集成路径:assemble + check_rules 必须同时通过。
|
||
|
||
上一轮 14/14 全绿却照样生成失败,原因是这个自测只调 build_tags(),
|
||
没走 assemble() + check_rules(),于是标签与歌词头小节数不一致这种
|
||
集成冲突根本测不出来。集成层的坑必须在这里测。
|
||
"""
|
||
cfg = loader.load({})
|
||
st = {"bpm": 124, "duration": 175, "key": "A minor",
|
||
"style": "English melodic future bass",
|
||
"instruments": "huge supersaw lead",
|
||
"mood": "vast uplifting emotional EDM",
|
||
"extra": "male low whisper vocal, sparse English words"}
|
||
|
||
class _A(object):
|
||
chill_vocal = True
|
||
instrumental = False
|
||
|
||
raw = open(_CHILL_DEMO, encoding="utf-8").read()
|
||
tags2, lyrics2 = assemble(_A, cfg, st, raw, "selftest")
|
||
errs = ironlaw.check_rules(lyrics2, tags2)
|
||
assert not errs, "chill 集成校验不通过: %s" % errs
|
||
|
||
# 结构核心断言:第一个有人唱的段之前,纯器乐段必须占满前半段。
|
||
# 注意这只保证"器乐占位存在",不保证人声起点被推后——实测两者无关。
|
||
# 音频层是否真的晚进,必须用 ab_vocal_curve.py / ab_is_real_voice.py 验。
|
||
lines = [l.strip() for l in lyrics2.split("\n") if l.strip()]
|
||
vocal_tags = ("[verse]", "[chorus]", "[hook]", "[bridge]", "[pre-chorus]")
|
||
first_v = next((i for i, l in enumerate(lines)
|
||
if l.lower() in vocal_tags), None)
|
||
assert first_v is not None, "chill 结构体里找不到任何人声段"
|
||
head = [l.lower() for l in lines[:first_v]]
|
||
# 只数"纯器乐段",不能把 [inst] 占位和段名一起算进去。
|
||
# 旧写法把两者相加,一个 breakdown 组就被算成 2,报出"8 个器乐段",
|
||
# 实际上只有 4 组——测试描述在骗人,验收语义必须按组算。
|
||
inst_secs = [h for h in head
|
||
if h in ("[breakdown]", "[build]", "[drop]", "[post-drop]")]
|
||
n_inst = len(inst_secs)
|
||
assert n_inst >= 4, \
|
||
"chill 第一个器乐段之前只有 %d 组纯器乐段,人声会紧跟 intro 进来" % n_inst
|
||
# 每个纯器乐段名后面必须真的挂 [inst],否则模型会当成有人唱的段
|
||
for i, h in enumerate(head):
|
||
if h in ("[breakdown]", "[build]", "[drop]", "[post-drop]"):
|
||
assert i + 1 < len(head) and head[i + 1] == "[inst]", \
|
||
"chill 器乐段 %s 后面没挂 [inst],会被当成可唱段" % h
|
||
|
||
# 器乐占位不能重复:原始歌词里的 [inst] 段已经重排过,
|
||
# 再留一遍会让人声提前回来
|
||
assert head.count("[inst]") <= n_inst, "chill 结构体器乐占位异常重复"
|
||
|
||
# 前奏小节数:标签与歌词头必须一致
|
||
mh = re.search(r"\[instrumental:[^\]]*?(\d+)\s*bar", lyrics2, re.I)
|
||
mt = re.search(r"(\d+)\s*bar instrumental intro", tags2.lower())
|
||
assert mh and mt and mh.group(1) == mt.group(1), \
|
||
"chill 标签 %s 小节与歌词头 %s 小节不一致" % (
|
||
mt.group(1) if mt else "?", mh.group(1) if mh else "?")
|
||
|
||
# 反向闸门:两者不一致必须被拦下
|
||
short_tags = tags2.replace("%s bar instrumental intro" % mt.group(1),
|
||
"16 bar instrumental intro", 1)
|
||
errs = ironlaw.check_rules(lyrics2, short_tags)
|
||
assert any("必须一致" in e for e in errs), \
|
||
"tags 与歌词头前奏小节数不一致未被拦截"
|
||
|
||
# 前奏小节数缺失也必须被拦
|
||
errs = ironlaw.check_rules(lyrics2, re.sub(r"\d+\s*bar instrumental",
|
||
"instrumental", tags2))
|
||
assert any("小节数" in e for e in errs), "tags 未写前奏小节数未被拦截"
|
||
|
||
return [("吟唱结构占位",
|
||
"人声段之前留了 %d 组纯器乐段(只管密度,不管起点),"
|
||
"标签侧完全不讨论人声进场时机" % n_inst)]
|
||
|
||
|
||
def _t_lexicon():
|
||
lex = loader.lexicon()
|
||
assert "长板" in (lex["banned"].get("longboard") or []), "禁用词表缺 长板"
|
||
good = "[intro]\n[Instrumental: 8 bar]\n[verse]\n我在街头双翘上起跳"
|
||
assert not ironlaw.check_lexicon(good, lex), ironlaw.check_lexicon(good, lex)
|
||
bad = "[intro]\n[verse]\n我戴着头盔冲下坡"
|
||
errs = ironlaw.check_lexicon(bad, lex)
|
||
assert errs, "禁用词未被拦截"
|
||
return [("身份词典", "街头双翘通过 / 长板类禁用词 %d 条被拦" % len(errs))]
|
||
|
||
|
||
def _t_density():
|
||
cfg = loader.load({})["lyrics"]
|
||
st = loader.styles()["pop"]
|
||
cap = meter.capacity(st["duration"], st["bpm"], cfg)
|
||
assert cap["hi"] > cap["lo"] > 0, "容量区间异常"
|
||
|
||
# 过短必须被拦
|
||
too_thin = "[intro]\n[Instrumental: 8 bar]\n[verse]\n太短了"
|
||
assert meter.check(too_thin, st["duration"], st["bpm"], cfg)[0], "过短未拦"
|
||
|
||
# 过多必须被拦
|
||
too_fat = "[intro]\n[Instrumental: 8 bar]\n[verse]\n" + \
|
||
"\n".join(["这一句有十二个字刚刚好"] * 300)
|
||
assert meter.check(too_fat, st["duration"], st["bpm"], cfg)[0], "过多未拦"
|
||
|
||
# 按 target_words 反推合规歌词,必须放行(防止阈值过严误杀)
|
||
mid, lo, hi = meter.target_words(st["duration"], st["bpm"], cfg)
|
||
line = "这一句刚好十二个字的样子" # 12 音节
|
||
nlines = max(1, int(mid / 12))
|
||
body = "\n".join([line] * nlines)
|
||
ok = "[intro]\n[Instrumental: 8 bar]\n[verse]\n" + body
|
||
errs, stt = meter.check(ok, st["duration"], st["bpm"], cfg)
|
||
assert not errs, "按 target 写的合规歌词被误杀: %s" % errs
|
||
return [("时长对应硬校验",
|
||
"%ds/%dBPM 容量 %d~%d,目标 %d 字,过短/过多均被拦,合规放行"
|
||
% (st["duration"], st["bpm"], cap["lo"], cap["hi"], mid))]
|
||
|
||
|
||
def _t_regression():
|
||
"""实歌回归:8 首用户已生成并认可的歌必须全部通过密度校验。
|
||
|
||
这是阈值校准的锚点。原来按理论值(每拍2.6~4.4音节、每句下限7)设阈值,
|
||
结果这 8 首全部被判"歌词过少"(填充率只有 26%~43%),说明阈值从未
|
||
对齐过真实输出。任何阈值改动都必须先过这一关。
|
||
"""
|
||
old_dir = r"c:\Users\孙文龙\.trae-cn\work\6ab523de391489c168050edc"
|
||
if not os.path.isdir(old_dir):
|
||
return [("实歌回归", "FAIL 找不到 remake_v5.py 目录")]
|
||
sys.path.insert(0, old_dir)
|
||
try:
|
||
import remake_v5 as old
|
||
except Exception as e:
|
||
return [("实歌回归", "FAIL 无法导入 remake_v5 (%s)" % e)]
|
||
|
||
cfg = loader.load({})["lyrics"]
|
||
bad = []
|
||
worst_pl = 99.0
|
||
for title, seed, bpm, key, dur, inst in old.META:
|
||
lyrics = old.build_song(title)[2]
|
||
errs, st = meter.check(lyrics, dur, bpm, cfg)
|
||
worst_pl = min(worst_pl, st["per_line"])
|
||
if errs:
|
||
bad.append("%s: %s" % (title, errs[0].split(":")[0]))
|
||
if bad:
|
||
return [("实歌回归", "FAIL %d/%d 首被误杀 -> %s"
|
||
% (len(bad), len(old.META), " | ".join(bad)))]
|
||
assert worst_pl >= cfg["ideal_line_syllable"][0], \
|
||
"实测最短平均句长 %.1f 已低于配置下限 %.1f,阈值与实歌脱节" % (
|
||
worst_pl, cfg["ideal_line_syllable"][0])
|
||
return [("实歌回归", "8 首已验证实歌全部通过(最短平均句长 %.1f 音节)"
|
||
% worst_pl)]
|
||
|
||
|
||
def _t_parity():
|
||
"""与旧脚本 remake_v5.py 的 API prompt 逐字段对比。"""
|
||
old_dir = r"c:\Users\孙文龙\.trae-cn\work\6ab523de391489c168050edc"
|
||
sys.path.insert(0, old_dir)
|
||
try:
|
||
import remake_v5 as old
|
||
except Exception as e:
|
||
return [("旧脚本 parity", "跳过:无法导入 remake_v5 (%s)" % e)]
|
||
|
||
title = "身体先听懂"
|
||
owf, otags, olyrics = old.build_song(title)
|
||
oapi = old.convert(owf)
|
||
|
||
cfg = loader.load({})
|
||
template = wfmod.load_template(cfg["paths"]["workflow_template"], cfg=cfg)
|
||
m = [x for x in old.META if x[0] == title][0]
|
||
_, oseed, obpm, okey, odur, oinst = m
|
||
nwf = wfmod.build_workflow(
|
||
template, title=title, tags=otags, lyrics=olyrics, seed=oseed,
|
||
bpm=obpm, duration=odur, keyscale=okey, sampler=cfg["sampler"],
|
||
encode=cfg["encode"], model=cfg["model"],
|
||
output_prefix="MyMusic/parity_%s" % title, output_format="flac",
|
||
language="zh", timesignature="4")
|
||
napi = wfmod.render(nwf)
|
||
|
||
if set(oapi) != set(napi):
|
||
return [("旧脚本 parity", "FAIL 节点集不同 %s vs %s"
|
||
% (sorted(oapi), sorted(napi)))]
|
||
|
||
diffs = []
|
||
for nid in sorted(oapi, key=int):
|
||
for k in sorted(set(oapi[nid]["inputs"]) | set(napi[nid]["inputs"])):
|
||
a = oapi[nid]["inputs"].get(k)
|
||
b = napi[nid]["inputs"].get(k)
|
||
if k in ("tags", "lyrics"):
|
||
# 新版 build_tags 措辞不同,只校验关键铁律都在
|
||
for kw in ("swlvoice", "8 bar instrumental intro hook"):
|
||
if (kw in str(a)) != (kw in str(b)):
|
||
diffs.append("节点%s.%s 铁律不一致" % (nid, k))
|
||
continue
|
||
if k == "filename_prefix":
|
||
continue # 输出前缀本就故意不同
|
||
if a != b:
|
||
diffs.append("节点%s.%s: 旧=%r 新=%r" % (nid, k, a, b))
|
||
if diffs:
|
||
return [("旧脚本 parity", "FAIL %d 处不同:%s"
|
||
% (len(diffs), " | ".join(diffs[:6])))]
|
||
return [("旧脚本 parity", "核心 API payload 与 remake_v5.py 等价(%d 节点;"
|
||
"filename_prefix 与歌词措辞为有意差异)" % len(oapi))]
|
||
|
||
|
||
def _t_fallback():
|
||
"""包内模板 + Comfy 输出目录必须可用;旧 workflows 旁路资产可选。"""
|
||
root = pmod.comfy_root()
|
||
bundled = os.path.join(pmod.PKG_ROOT, "workflows", "template.json")
|
||
music = pmod.output_dir(root, "MyMusic")
|
||
if not os.path.isfile(bundled):
|
||
return [("fallback 资产", "FAIL 缺包内模板 %s" % bundled)]
|
||
if not os.path.isdir(music):
|
||
os.makedirs(music, exist_ok=True)
|
||
host = pmod.comfy_workflows_dir()
|
||
legacy = [
|
||
os.path.join(host, "song_rules.py"),
|
||
os.path.join(host, "urbanhiphop-身体先听懂.json"),
|
||
]
|
||
legacy_ok = sum(1 for p in legacy if os.path.isfile(p))
|
||
bak_root = os.path.join(host, "bak")
|
||
stamps = sorted(os.listdir(bak_root)) if os.path.isdir(bak_root) else []
|
||
return [("fallback 资产",
|
||
"包内模板 + MyMusic 就位;Comfy workflows 旧资产 %d/2,bak %d 份"
|
||
% (legacy_ok, len(stamps)))]
|
||
|
||
|
||
def _t_window():
|
||
"""target_words 报的区间必须与 check 实际放行窗口一致。
|
||
|
||
这是最容易出"文档说能过、实际过不了"的地方:容量 [lo,hi] 是理论值,
|
||
但 check 还叠加了填充率红线,真正生效的窗口必须两者取交。
|
||
"""
|
||
cfg = loader.load({})["lyrics"]
|
||
for name in ("pop", "punk", "urbanhiphop"):
|
||
st = loader.styles()[name]
|
||
dur, bpm = st["duration"], st["bpm"]
|
||
mid, lo, hi = meter.target_words(dur, bpm, cfg)
|
||
assert lo <= mid <= hi, "%s 目标字数 %d 不在放行区间 %d~%d" % (
|
||
name, mid, lo, hi)
|
||
|
||
cap = meter.capacity(dur, bpm, cfg)
|
||
assert cap["lo"] <= hi and lo <= cap["hi"], \
|
||
"%s 放行区间 %d~%d 超出理论容量 %d~%d" % (
|
||
name, lo, hi, int(cap["lo"]), int(cap["hi"]))
|
||
|
||
# 窗口两端必须真的被 check 拦/放,边界不能只存在于文档里
|
||
line = "这一句刚好十二个字的样子"
|
||
|
||
def build(n):
|
||
return ("[intro]\n[Instrumental: 8 bar]\n[verse]\n"
|
||
+ "\n".join([line] * n))
|
||
|
||
n_mid = max(1, int(mid / 12))
|
||
ok_errs, _ = meter.check(build(n_mid), dur, bpm, cfg)
|
||
assert not ok_errs, "%s 按目标字数写的歌词被误杀: %s" % (name, ok_errs)
|
||
|
||
# 少一个字就该报"过少",多一个字就该报"过多",容差必须小
|
||
few = meter.check(build(max(1, int((lo - 30) / 12))), dur, bpm, cfg)[0]
|
||
assert any("过少" in e for e in few), \
|
||
"%s 低于放行下限却未报过少: %s" % (name, few)
|
||
many = meter.check(build(int((hi + 30) / 12) + 1), dur, bpm, cfg)[0]
|
||
assert any("过多" in e for e in many), \
|
||
"%s 高于放行上限却未报过多: %s" % (name, many)
|
||
return [("字数窗口自洽",
|
||
"pop/punk/urbanhiphop 的 target 区间 == check 放行窗口,边界确实生效")]
|
||
|
||
|
||
def cmd_selftest(args):
|
||
from .common import session_logger
|
||
cfg = loader.load({})
|
||
run, log = session_logger(args, cfg, "selftest")
|
||
cases = [_t_runid, _t_paths, _t_path_layout, _t_config_priority,
|
||
_t_model_locked, _t_ironlaw,
|
||
_t_instrumental, _t_chill_delay, _t_groove_first_beat,
|
||
_t_intro_instrument_filter, _t_chill_no_conflict,
|
||
_t_chill_integration,
|
||
_t_sparse_profile, _t_lexicon,
|
||
_t_density, _t_window, _t_regression, _t_parity, _t_fallback]
|
||
npass = 0
|
||
print("== songforge selftest ==")
|
||
log.info("selftest 开始", cases=len(cases))
|
||
for fn in cases:
|
||
try:
|
||
notes = fn()
|
||
for name, detail in notes:
|
||
# 测试函数用返回文本表达失败时,必须显式识别,不能算作通过
|
||
bad = bool(re.search(r"\bFAIL\b|\bERR\b|失败", str(detail)))
|
||
print(" [%s] %-22s %s"
|
||
% ("FAIL" if bad else "ok", name, detail))
|
||
if bad:
|
||
raise AssertionError(detail)
|
||
npass += 1
|
||
except AssertionError as e:
|
||
print(" [--] %-22s %s" % (fn.__name__[3:], e or "断言失败"))
|
||
log.warn("case 失败", case=fn.__name__, err=str(e))
|
||
except Exception as e:
|
||
print(" [ERR] %-22s %s: %s"
|
||
% (fn.__name__[3:], type(e).__name__, e))
|
||
log.error("case 异常", case=fn.__name__, err=str(e))
|
||
print("== 结果 ==")
|
||
print(" %d/%d 通过" % (npass, len(cases)))
|
||
log.info("selftest 结束", passed=npass, total=len(cases), log_dir=run.dir)
|
||
log.close()
|
||
return 0 if npass == len(cases) else 1
|
||
|
||
|
||
# ---------------------------------------------------------------- parser
|
||
|