fix: SRT 按 cue 解析并按 ID 回填译文，OCR 空帧分段与失败重试
This commit is contained in:
@@ -1,231 +1,120 @@
|
||||
"""翻译批处理行数对齐测试（先红后绿）。
|
||||
"""R05 回归：合法 SRT 多行/空 cue、稳定 ID 翻译和非法模型输出重试。
|
||||
|
||||
背景：真实任务 run_51242078d76e（CJOD-255-长视频）产出的中文字幕存在
|
||||
"内容-时间错位"——例如第 756 条「好像喜欢害羞的样子」被贴到 3805.34s
|
||||
（该时间实际是日文「4つんばんですか(趴着吗)」的位置），而这条译文本应是
|
||||
第 758 条「恥ずかしいのが好きみたいなので(喜欢害羞姿势)」的译文。
|
||||
|
||||
根因：nodes/llm.py 的 translate_lines 按 CHUNK_SIZE=20 分批把日文行发给
|
||||
LLM，返回的译文行用 translated.extend() **无条件顺序拼接**，全批结束后只在
|
||||
invoke 末尾做"多截断、少补空"。只要某批 LLM 返回行数 != 输入行数（实测大量
|
||||
批次出现译文 21 行/原文 20 行），该批之后**所有字幕文本整体错位**，而时间戳
|
||||
（从原文复制）保持不变 —— 造成"文本对错时间，程序从时间戳上看不出问题"。
|
||||
|
||||
修复（见 nodes/llm.py）：
|
||||
1. 系统提示词新增"逐行独立翻译 + 碎片句按语境给含义 + 禁止合并/拆分"，
|
||||
从源头减少 LLM 重组断句导致的行数不一致；
|
||||
2. 程序侧兜底 _repair_batch：返回行数 != 输入行数时，
|
||||
- 多行：末尾多余行合并到前一行（碎片本质同一句，时间轴保留）；
|
||||
- 少行：末尾补空串占位（宁缺勿错位，不挤占相邻字幕时间轴）。
|
||||
|
||||
本测试分两层：
|
||||
1. _repair_batch / translate_lines 确定性单元测试（红 -> 绿）；
|
||||
2. 真实数据 + 真实 LLM 集成测试（非 mock），验证产物与原文逐条对齐。
|
||||
数据/Key 缺失时 skip。
|
||||
历史 run_51242078d76e 出现文本贴错时间；仅检查行数或在末尾合并/补空无法
|
||||
定位中间缺失。本测试在 HTTP 边界注入 JSON 响应，调用真实翻译实现。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
|
||||
from tests.realdata_contract import parse_srt_entries
|
||||
|
||||
WORKSPACE = Path(__file__).resolve().parent.parent
|
||||
TRANSCRIPT = Path(
|
||||
"/home/cat/Downloads/39.105.149.197/202609051737"
|
||||
"/run_51242078d76e/steps/asr/transcript.srt"
|
||||
)
|
||||
|
||||
CHUNK_SIZE = 20
|
||||
from nodes import llm
|
||||
from wov_sdk.models import InvokeRequest
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 层一 helper：可注入的假 HTTP 客户端（与 nodes/llm.py 的 urllib 契约一致）
|
||||
# ---------------------------------------------------------------------------
|
||||
def _http(monkeypatch, answers):
|
||||
"""按次序返回真实 chat.completions 结构，并记录发出的输入。"""
|
||||
calls = []
|
||||
iterator = iter(answers)
|
||||
|
||||
class Response:
|
||||
def __init__(self, content):
|
||||
self.content = content
|
||||
|
||||
def __enter__(self):
|
||||
return self
|
||||
|
||||
def __exit__(self, *args):
|
||||
return False
|
||||
|
||||
def read(self):
|
||||
return json.dumps({"choices": [{"message": {"content": self.content}}],
|
||||
"usage": {"total_tokens": 10}}).encode()
|
||||
|
||||
def open_request(request, **kwargs):
|
||||
calls.append(json.loads(request.data))
|
||||
answer = next(iterator)
|
||||
return Response(answer if isinstance(answer, str) else json.dumps(answer))
|
||||
|
||||
monkeypatch.setattr("urllib.request.urlopen", open_request)
|
||||
return calls
|
||||
|
||||
|
||||
class _FakeUrlOpen:
|
||||
"""模拟 urllib.request.urlopen：按调用次数依次返回预置的 LLM 输出。"""
|
||||
|
||||
def __init__(self, contents: list[str]):
|
||||
self._contents = contents
|
||||
self._calls = 0
|
||||
|
||||
def __enter__(self):
|
||||
return self
|
||||
|
||||
def __exit__(self, exc_type, exc, tb):
|
||||
return False
|
||||
|
||||
def read(self) -> bytes:
|
||||
content = self._contents[self._calls]
|
||||
self._calls += 1
|
||||
payload = {"choices": [{"message": {"content": content}}]}
|
||||
return json.dumps(payload).encode("utf-8")
|
||||
|
||||
|
||||
def _patch_translate_llm(monkeypatch, batch_outputs: list[str]) -> None:
|
||||
"""统一打桩：把 translate_lines 内 urlopen 换成 _FakeUrlOpen。"""
|
||||
import urllib.request
|
||||
|
||||
# 直接替换 urllib.request.urlopen（nodes/llm.py 也是经它调用）。
|
||||
|
||||
# 直接替换 urllib.request.urlopen（nodes/llm.py 也是经它调用）。
|
||||
fake = _FakeUrlOpen(batch_outputs)
|
||||
monkeypatch.setattr(urllib.request, "urlopen", lambda req, timeout=None: fake)
|
||||
monkeypatch.setenv("LLM_API_KEY", "test-key")
|
||||
monkeypatch.setenv("LLM_API_BASE", "http://fake/v1/chat/completions")
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 层一：_repair_batch 确定性单元测试
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_repair_batch_extra_lines_merged() -> None:
|
||||
"""多行：LLM 返回 21 行但输入 20 行，末尾多余行应合并到前一行。"""
|
||||
from nodes.llm import _repair_batch
|
||||
|
||||
out = _repair_batch([f"译{i}" for i in range(21)], 20)
|
||||
assert len(out) == 20
|
||||
# 最后一行 = 原第 19（索引19）+第 20（索引20）行的合并。
|
||||
assert out[19] == "译19 译20"
|
||||
|
||||
|
||||
def test_repair_batch_fewer_lines_padded() -> None:
|
||||
"""少行：LLM 返回 19 行但输入 20 行，末尾补空串占位不挤占时间轴。"""
|
||||
from nodes.llm import _repair_batch
|
||||
|
||||
out = _repair_batch([f"译{i}" for i in range(19)], 20)
|
||||
assert len(out) == 20
|
||||
assert out[19] == ""
|
||||
|
||||
|
||||
def test_repair_batch_exact_unchanged() -> None:
|
||||
"""正好对齐：原样返回。"""
|
||||
from nodes.llm import _repair_batch
|
||||
|
||||
out = _repair_batch([f"译{i}" for i in range(20)], 20)
|
||||
assert len(out) == 20
|
||||
assert out == [f"译{i}" for i in range(20)]
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 层一：translate_lines 整批校验（多行/少行场景经修复后必须对齐）
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
def test_translate_lines_aligns_extra_line(monkeypatch) -> None:
|
||||
"""输入 40 行（两批 20），首批 LLM 返回 21 行：修复后必须对齐为 40 行。"""
|
||||
from nodes import llm as llm_node
|
||||
|
||||
src_lines = [f"原文{i}" for i in range(40)]
|
||||
batch1_wrong = "\n".join([f"译{i}" for i in range(21)]) # 21 行错位源
|
||||
batch2_ok = "\n".join([f"译{i}" for i in range(20, 40)])
|
||||
_patch_translate_llm(monkeypatch, [batch1_wrong, batch2_ok])
|
||||
|
||||
result = llm_node.translate_lines(src_lines, {})
|
||||
assert len(result) == len(src_lines), (
|
||||
f"translate_lines 未把多行合并对齐：输入 {len(src_lines)} 行，返回 {len(result)} 行"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
def test_translate_lines_aligns_missing_line(monkeypatch) -> None:
|
||||
"""第二批 LLM 少行时触发重试：重试返回正确 20 行后必须仍为 40 行。"""
|
||||
from nodes import llm as llm_node
|
||||
|
||||
src_lines = [f"原文{i}" for i in range(40)]
|
||||
batch1_ok = "\n".join([f"译{i}" for i in range(20)])
|
||||
# 第二批第一次返回 19 行（少行）-> 触发重试；第二次返回正确 20 行。
|
||||
batch2_short = "\n".join([f"译{i}" for i in range(20, 39)]) # 19 行
|
||||
batch2_retry = "\n".join([f"译{i}" for i in range(20, 40)]) # 20 行
|
||||
_patch_translate_llm(monkeypatch, [batch1_ok, batch2_short, batch2_retry])
|
||||
|
||||
result = llm_node.translate_lines(src_lines, {})
|
||||
assert len(result) == len(src_lines), (
|
||||
f"translate_lines 未把少行补齐：输入 {len(src_lines)} 行，返回 {len(result)} 行"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
def test_translate_lines_pads_after_retries_exhausted(monkeypatch) -> None:
|
||||
"""少行且重试耗尽：必须补空串占位，仍保持与输入等长（宁缺勿错位）。"""
|
||||
from nodes import llm as llm_node
|
||||
|
||||
src_lines = [f"原文{i}" for i in range(40)]
|
||||
batch1_ok = "\n".join([f"译{i}" for i in range(20)])
|
||||
batch2_short = "\n".join([f"译{i}" for i in range(20, 39)])
|
||||
from nodes.llm import MAX_BATCH_RETRIES
|
||||
|
||||
# 首次调用 + 重试重发，共 MAX_BATCH_RETRIES 次对 batch2 的调用都返回 19 行。
|
||||
responses = [batch1_ok] + [batch2_short] * MAX_BATCH_RETRIES
|
||||
_patch_translate_llm(monkeypatch, responses)
|
||||
|
||||
result = llm_node.translate_lines(src_lines, {})
|
||||
assert len(result) == len(src_lines), (
|
||||
f"重试耗尽后未能补空串：输入 {len(src_lines)} 行，返回 {len(result)} 行"
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 层二：真实数据 + 真实 LLM 集成测试（非 mock）
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _llm_credentials_ok() -> bool:
|
||||
"""是否具备真实 LLM 调用条件（加载 .env 后 Key 非空）。"""
|
||||
try:
|
||||
from dotenv import load_dotenv
|
||||
|
||||
load_dotenv(WORKSPACE / ".env")
|
||||
except Exception:
|
||||
pass
|
||||
return bool(os.getenv("LLM_API_KEY"))
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
def test_pipeline_zh_cn_timetext_alignment(tmp_path) -> None:
|
||||
"""真实数据 + 真实 LLM：完整翻译流水线后，译文必须与原文时间逐条对齐。
|
||||
|
||||
方法：把真实日文 transcript.srt 喂给 llm.invoke（真实 LLM API），产出
|
||||
cn.srt；逐条比较 cn.srt 与原文的 (start, 行序) 严格一致。
|
||||
"""
|
||||
if not _llm_credentials_ok():
|
||||
pytest.skip("未配置 LLM_API_KEY，跳过真实 LLM 集成测试")
|
||||
if not TRANSCRIPT.is_file():
|
||||
pytest.skip("缺少真实 transcript.srt，跳过集成测试")
|
||||
|
||||
from wov_sdk.models import InvokeRequest
|
||||
from nodes import llm as llm_node
|
||||
|
||||
out_dir = tmp_path / "out"
|
||||
response = llm_node.invoke(
|
||||
InvokeRequest(
|
||||
run_id="align_llm_test",
|
||||
node_instance_id="",
|
||||
inputs={"srt_uri": str(TRANSCRIPT)},
|
||||
params={"target_language": "zh-CN"},
|
||||
output_dir=str(out_dir),
|
||||
)
|
||||
)
|
||||
def test_multiline_and_empty_cues_keep_timestamps(monkeypatch, tmp_path):
|
||||
"""多行正文视为一条 cue，空 cue 不发送翻译，时间轴不会进入模型输入。"""
|
||||
source = tmp_path / "input.srt"
|
||||
source.write_text("\ufeff7\n00:00:01,000 --> 00:00:02,000\nこんにちは\n元気ですか\n\n"
|
||||
"8\n00:00:03,000 --> 00:00:04,000\n\n"
|
||||
"9\n00:00:05,000 --> 00:00:06,000\nはい\n", encoding="utf-8")
|
||||
calls = _http(monkeypatch, [[{"id": 3, "text": "是的"}, {"id": 1, "text": "你好\n还好吗"}]])
|
||||
response = llm.invoke(InvokeRequest(run_id="r", node_instance_id="", inputs={"srt_uri": str(source)}, output_dir=str(tmp_path / "out")))
|
||||
assert response.status == "completed", response.error
|
||||
assert json.loads(calls[0]["messages"][1]["content"]) == [
|
||||
{"id": 1, "text": "こんにちは\n元気ですか"}, {"id": 3, "text": "はい"}]
|
||||
text = Path(response.outputs["cn_srt_uri"]).read_text()
|
||||
assert text.count("-->") == 3
|
||||
assert "00:00:01,000 --> 00:00:02,000\n你好\n还好吗" in text
|
||||
assert "00:00:03,000 --> 00:00:04,000\n\n" in text
|
||||
assert "00:00:05,000 --> 00:00:06,000\n是的" in text
|
||||
|
||||
zh_path = Path(response.outputs["cn_srt_uri"])
|
||||
zh_entries = parse_srt_entries(zh_path.read_text(encoding="utf-8"))
|
||||
src_entries = parse_srt_entries(TRANSCRIPT.read_text(encoding="utf-8"))
|
||||
assert len(zh_entries) == len(src_entries), (
|
||||
f"译文条数 {len(zh_entries)} != 原文 {len(src_entries)}：批内行数不一致导致错位。"
|
||||
)
|
||||
|
||||
for i, (ze, se) in enumerate(zip(zh_entries, src_entries)):
|
||||
if abs(ze["start"] - se["start"]) > 0.01:
|
||||
raise AssertionError(
|
||||
f"第 {i} 条译文时间 {ze['start']:.2f} != 原文 {se['start']:.2f}："
|
||||
f"译文文本已整体错位（原文 '{se['text'][:15]}'）"
|
||||
)
|
||||
@pytest.mark.parametrize("bad", [
|
||||
[{"id": 1, "text": "一"}],
|
||||
[{"id": 1, "text": "一"}, {"id": 1, "text": "重复"}],
|
||||
[{"id": 1, "text": "一"}, {"id": 99, "text": "未知"}],
|
||||
[{"id": True, "text": "一"}, {"id": 2, "text": "二"}],
|
||||
[{"id": 1, "text": "一"}, {"id": 2, "text": ""}],
|
||||
"一\n二", "{truncated", {"1": "一", "2": "二"},
|
||||
])
|
||||
def test_invalid_ids_retry_without_positional_repair(monkeypatch, bad):
|
||||
"""缺失/重复/未知 ID 和无结构文本均重试整批，不猜测句子对应关系。"""
|
||||
calls = _http(monkeypatch, [bad, [{"id": 2, "text": "二"}, {"id": 1, "text": "一"}]])
|
||||
assert llm.translate_lines(["first", "second"], {}) == ["一", "二"]
|
||||
assert len(calls) == 2
|
||||
assert calls[0]["messages"][1] == calls[1]["messages"][1]
|
||||
|
||||
|
||||
def test_exhausted_alignment_retries_fail_node(monkeypatch, tmp_path):
|
||||
"""无法对齐时返回 failed，不生成带空占位或错位文本的成功成品。"""
|
||||
source = tmp_path / "input.srt"
|
||||
source.write_text("1\n00:00:01,000 --> 00:00:02,000\nhello\n", encoding="utf-8")
|
||||
calls = _http(monkeypatch, ["无 ID 输出"] * llm.MAX_BATCH_RETRIES)
|
||||
response = llm.invoke(InvokeRequest(run_id="r", node_instance_id="", inputs={"srt_uri": str(source)}, output_dir=str(tmp_path / "out")))
|
||||
assert response.status == "failed"
|
||||
assert len(calls) == llm.MAX_BATCH_RETRIES
|
||||
assert not (tmp_path / "out/cn.srt").exists()
|
||||
|
||||
|
||||
def test_batch_ids_are_global_and_order_independent(monkeypatch):
|
||||
"""跨批 ID 保持全局位置，乱序输出也能按正确 cue 回填。"""
|
||||
answers = [[{"id": i, "text": f"译{i}"} for i in range(20, 0, -1)],
|
||||
[{"id": 22, "text": "译22"}, {"id": 21, "text": "译21"}]]
|
||||
calls = _http(monkeypatch, answers)
|
||||
assert llm.translate_lines([f"原{i}" for i in range(1, 23)], {}) == [f"译{i}" for i in range(1, 23)]
|
||||
assert json.loads(calls[1]["messages"][1]["content"])[0]["id"] == 21
|
||||
|
||||
|
||||
def test_malformed_srt_fails_without_llm(monkeypatch, tmp_path):
|
||||
"""非空坏字幕不能被静默解析为空并成功输出。"""
|
||||
source = tmp_path / "input.srt"
|
||||
source.write_text("1\ninvalid timestamp\nhello\n", encoding="utf-8")
|
||||
calls = _http(monkeypatch, [])
|
||||
response = llm.invoke(InvokeRequest(run_id="r", node_instance_id="", inputs={"srt_uri": str(source)}, output_dir=str(tmp_path / "out")))
|
||||
assert response.status == "failed"
|
||||
assert calls == []
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
def test_real_llm_structured_translation():
|
||||
"""真实接口校准 JSON 协议；无 Key 时跳过，仅使用短日文句子控制调用量。"""
|
||||
import os
|
||||
from dotenv import load_dotenv
|
||||
|
||||
load_dotenv()
|
||||
if not os.getenv("LLM_API_KEY"):
|
||||
pytest.skip("未配置 LLM_API_KEY")
|
||||
result = llm.translate_lines(["こんにちは。", "ありがとうございます。"], {})
|
||||
assert len(result) == 2 and all(result)
|
||||
assert any(word in result[0] for word in ("好", "嗨"))
|
||||
assert "谢" in result[1]
|
||||
|
||||
Reference in New Issue
Block a user