iomgaa/Video-Tree-TRM5
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor(sampler): replace family_spec param with level+constraint

Browse Source 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
iomgaa
2026-07-14 05:38:14 -04:00
parent b6b6a48503
commit e2325b6535
3 changed files with 100 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files
app/question_gen/pipeline_v2.py
+70 -31
View File
@@ -38,7 +38,7 @@ from app.question_gen.families import QuestionFamilySpec, get_family_for_slot
from app.question_gen.gates import GateReport, GateResult, GateVerdict, run_gates
from app.question_gen.generator_v2 import CandidateQuestion, generate_one_v2
from app.question_gen.postprocess import run_postprocess
from app.question_gen.sampler_v2 import sample_material_v2
from app.question_gen.sampler_v2 import _TASK_TYPE_TO_LEVEL, sample_material_v2
from core.types import GeneratedQuestion
if TYPE_CHECKING:
@@ -101,8 +101,6 @@ class PipelineConfig:
concurrency: 并发 slot 数上限。
seed: 随机种子。
output_dir: 输出目录。
gate_models: 门控模型配置字典。
heavy_agent_model: 重量抽检使用的模型名。
"""
family_ratios: dict[str, float]
@@ -113,8 +111,6 @@ class PipelineConfig:
concurrency: int
seed: int
output_dir: Path
gate_models: dict[str, str]
heavy_agent_model: str
# ---------------------------------------------------------------------------
@@ -163,8 +159,6 @@ def load_pipeline_config(yaml_path: Path) -> PipelineConfig:
concurrency=int(section["concurrency"]),
seed=int(section["seed"]),
output_dir=Path(section["output_dir"]),
gate_models=section["gate"],
heavy_agent_model=str(section["heavy_agent_model"]),
)
@@ -269,6 +263,7 @@ def _is_duplicate(
def _to_generated_question(
candidate: CandidateQuestion,
*,
family: str,
options: tuple[str, ...] | None = None,
answer: str | None = None,
) -> GeneratedQuestion:
@@ -276,6 +271,7 @@ def _to_generated_question(
参数:
candidate: 门控通过的候选题目。
family: 问题家族名称（如 "RETRIEVAL"）。
options: 洗牌后的选项元组（若为 None 则使用 candidate 原始选项）。
answer: 重映射后的答案字母（若为 None 则使用 candidate 原始答案）。
@@ -291,6 +287,7 @@ def _to_generated_question(
answer=answer if answer is not None else candidate.answer,
source_nodes=candidate.source_nodes,
difficulty=candidate.difficulty,
family=family,
skill_target=candidate.skill_target,
difficulty_steps=None,
)
@@ -316,6 +313,7 @@ async def _process_one_slot(
*,
session_id: str,
run_id: str,
all_trees: dict[str, TreeIndex] | None = None,
) -> GeneratedQuestion | None:
"""处理单个 slot 的完整重出循环。
@@ -347,18 +345,36 @@ async def _process_one_slot(
返回:
GeneratedQuestion（通过全部检查）或 None（重出耗尽）。
"""
_RESAMPLE_VIDEO_INTERVAL = 1
async with sem:
prev_reason: str | None = None
current_tree = tree
current_video_id = slot.video_id
for attempt in range(1, config.retry_limit + 1):
# 连续失败 _RESAMPLE_VIDEO_INTERVAL 次后换视频
if attempt > 1 and (attempt - 1) % _RESAMPLE_VIDEO_INTERVAL == 0 and all_trees:
alt_ids = [v for v in all_trees if v != current_video_id]
if alt_ids:
current_video_id = rng.choice(alt_ids)
current_tree = all_trees[current_video_id]
logger.info(
"slot {} 连续 {} 次失败，换视频 {} 重试",
slot.slot_id,
attempt - 1,
current_video_id,
)
# Phase 1: 采样素材
try:
material = sample_material_v2(
tree=tree,
family_spec=slot.family,
tree=current_tree,
task_type=slot.task_type,
used_node_ids=used_node_ids,
rng=rng,
level=_TASK_TYPE_TO_LEVEL[slot.task_type],
constraint=slot.family.sampling,
)
except (RuntimeError, KeyError) as e:
logger.warning(
@@ -374,16 +390,16 @@ async def _process_one_slot(
try:
candidate = await generate_one_v2(
vlm=vlm,
tree=tree,
tree=current_tree,
material=material,
family_spec=slot.family,
task_type=slot.task_type,
seq=slot.seq,
video_id=slot.video_id,
video_id=current_video_id,
reject_reason=prev_reason,
session_id=session_id,
)
except (ValueError, FileNotFoundError) as e:
except (ValueError, FileNotFoundError, OSError, Exception) as e:
logger.warning(
"slot {} 生成失败 (attempt {}/{}): {}",
slot.slot_id,
@@ -399,7 +415,7 @@ async def _process_one_slot(
item_id=item_id,
run_id=run_id,
slot_id=slot.slot_id,
video_id=slot.video_id,
video_id=current_video_id,
family=slot.family.name,
task_type=slot.task_type,
skill_target=slot.family.skill_target,
@@ -446,15 +462,27 @@ async def _process_one_slot(
store.update_gates(item_id, verbatim_report)
continue
# Phase 6: 四门质量检查
report = await run_gates(
candidate=candidate,
tree=tree,
llm=llm,
family_spec=slot.family,
postprocess=pp,
session_id=session_id,
)
# Phase 6: 四门质量检查（key_verify 使用 VLM 看帧+文本）
try:
report = await run_gates(
candidate=candidate,
tree=tree,
llm=llm,
family_spec=slot.family,
postprocess=pp,
vlm=vlm,
session_id=session_id,
)
except Exception as e:
logger.warning(
"slot {} 门控调用异常 (attempt {}/{}): {}",
slot.slot_id,
attempt,
config.retry_limit,
e,
)
prev_reason = f"gate_error: {e}"
continue
store.update_gates(item_id, report)
if not report.passed:
@@ -481,7 +509,12 @@ async def _process_one_slot(
continue
# Phase 8: 通过全部检查 → 接受（使用洗牌后的选项和答案）
result = _to_generated_question(candidate, options=pp.options, answer=pp.answer)
result = _to_generated_question(
candidate,
family=slot.family.name,
options=pp.options,
answer=pp.answer,
)
# 将题目 embedding 加入池（flatten 确保 1D）
embed_pool.append(embed_fn(candidate.question).flatten())
# 标记使用的节点
@@ -571,16 +604,16 @@ async def _heavy_check_one(
_DEFAULT_TASK_TYPES: list[str] = [
"Action Recognition",
"Action Reasoning",
"Action Prediction",
"Action Sequence",
"Attribute Perception",
"Counting Problem",
"Information Synopsis",
"Object Recognition",
"Object Reasoning",
"Object Interaction",
"Scene Understanding",
"Event Reasoning",
"Causal Reasoning",
"Temporal Reasoning",
"OCR Problems",
"Spatial Perception",
"Spatial Reasoning",
"Temporal Perception",
"Temporal Reasoning",
]
@@ -697,6 +730,7 @@ async def run_pipeline_v2(
*,
task_types: list[str] | None = None,
progress: dict[str, str] | None = None,
on_accept: Callable[[GeneratedQuestion], None] | None = None,
) -> PipelineResult:
"""v2 出题管线主入口 — 编排全部 slot 的生成、检查与抽检。
@@ -718,6 +752,7 @@ async def run_pipeline_v2(
config: 管线配置。
task_types: 任务类型列表（默认使用 12 类标准集）。
progress: 已完成 slot 映射 {slot_id → "accepted"|"rejected"}。
on_accept: 每接受一题时的回调（用于实时持久化，防崩溃丢数据）。
返回:
PipelineResult 实例。
@@ -749,7 +784,7 @@ async def run_pipeline_v2(
if tree is None:
logger.warning("slot {} 对应视频 {} 的树不存在，跳过", slot.slot_id, slot.video_id)
return None
return await _process_one_slot(
result = await _process_one_slot(
slot=slot,
tree=tree,
vlm=vlm,
@@ -763,7 +798,11 @@ async def run_pipeline_v2(
sem=sem,
session_id=session_id,
run_id=run_id,
all_trees=trees,
)
if result is not None and on_accept is not None:
on_accept(result)
return result
results = await asyncio.gather(*[_process_wrapper(s) for s in pending_slots])
app/question_gen/sampler_v2.py
+19 -19
View File
@@ -1,16 +1,17 @@
"""v2 素材采样器 — 基于家族约束的树节点采样与上下文收集。
"""v2 素材采样器 — 基于采样约束的树节点采样与上下文收集。
在 v1 synthesizer 的基础上引入 QuestionFamilySpec 约束验证，
在 v1 synthesizer 的基础上引入 SamplingConstraint 约束验证，
为每次出题提供更丰富的素材上下文（字幕、跨 L2 上下文、帧路径）。
典型调用路径::
material = sample_material_v2(
tree=tree_index,
family_spec=REASONING_FAMILY,
task_type="Causal Reasoning",
task_type="Action Reasoning",
used_node_ids=already_used,
rng=rng,
level=2,
constraint=my_constraint,
)
"""
@@ -24,7 +25,7 @@ from loguru import logger
if TYPE_CHECKING:
import random
from app.question_gen.families import QuestionFamilySpec, SamplingConstraint
from app.question_gen.families import SamplingConstraint
from app.tree.index import L1Node, L2Node, TreeIndex
# ---------------------------------------------------------------------------
@@ -35,18 +36,18 @@ _TASK_TYPE_TO_LEVEL: dict[str, int] = {
# Level 3（细粒度帧级）
"Action Recognition": 3,
"Object Recognition": 3,
"Attribute Perception": 3,
"OCR Problems": 3,
# Level 2（片段/事件级）
"Action Reasoning": 2,
"Action Prediction": 2,
"Action Sequence": 2,
"Object Reasoning": 2,
"Object Interaction": 2,
"Scene Understanding": 2,
"Event Reasoning": 2,
"Causal Reasoning": 2,
"Information Synopsis": 2,
"Counting Problem": 2,
# Level 1（段落/场景级）
"Temporal Reasoning": 1,
"Temporal Perception": 1,
"Spatial Reasoning": 1,
"Spatial Perception": 1,
}
@@ -494,17 +495,18 @@ def _collect_source_nodes(tree: TreeIndex, node_id: str) -> tuple[str, ...]:
def sample_material_v2(
tree: TreeIndex,
family_spec: QuestionFamilySpec,
task_type: str,
used_node_ids: set[str],
rng: random.Random,
*,
level: int,
constraint: SamplingConstraint,
max_attempts: int = 10,
) -> MaterialContext:
"""基于家族约束从视频树中采样素材上下文。
"""基于采样约束从视频树中采样素材上下文。
采样流程：
1. 根据 task_type 确定采样层级
1. 按指定 level 确定采样层级
2. 随机选取候选节点（排除 used_node_ids）
3. 验证 SamplingConstraint 约束
4. 约束不满足则重试（最多 max_attempts 次）
@@ -512,10 +514,11 @@ def sample_material_v2(
参数:
tree: 三层树索引。
family_spec: 问题家族规格（含采样约束）。
task_type: 任务类型字符串。
used_node_ids: 本轮已用节点 ID 集合。
rng: 可控随机数生成器。
level: 采样层级（1/2/3）。
constraint: 采样约束条件。
max_attempts: 最大尝试次数。
返回:
@@ -523,10 +526,7 @@ def sample_material_v2(
异常:
RuntimeError: 耗尽 max_attempts 次尝试仍无法满足约束。
KeyError: task_type 不在 _TASK_TYPE_TO_LEVEL 映射中。
"""
level = _TASK_TYPE_TO_LEVEL[task_type]
constraint = family_spec.sampling
for attempt in range(max_attempts):
# Phase 1: 按层级采样候选节点
@@ -589,5 +589,5 @@ def sample_material_v2(
raise RuntimeError(
f"sample_material_v2: 耗尽 max_attempts={max_attempts} 次尝试，"
f"无法为 task_type='{task_type}' 满足家族 '{family_spec.name}' 的采样约束"
f"无法为 task_type='{task_type}' (level={level}) 满足采样约束"
)
tests/unit/test_sampler_v2.py
+11 -6
View File
@@ -181,10 +181,11 @@ class TestSampleMaterialV2:
rng = random.Random(42)
result = sample_material_v2(
tree=real_tree,
family_spec=RETRIEVAL_FAMILY,
task_type="Action Reasoning",
used_node_ids=set(),
rng=rng,
level=2,
constraint=RETRIEVAL_FAMILY.sampling,
)
assert isinstance(result, MaterialContext)
@@ -213,10 +214,11 @@ class TestSampleMaterialV2:
result = sample_material_v2(
tree=real_tree,
family_spec=RETRIEVAL_FAMILY,
task_type="Action Reasoning",
used_node_ids=used,
rng=rng,
level=2,
constraint=RETRIEVAL_FAMILY.sampling,
)
# 锚节点应该是那个未被排除的 L2
@@ -233,10 +235,11 @@ class TestSampleMaterialV2:
with pytest.raises(RuntimeError, match="max_attempts"):
sample_material_v2(
tree=sparse_tree,
family_spec=VISUAL_FAMILY,
task_type="Object Recognition",
used_node_ids=set(),
rng=rng,
level=3,
constraint=VISUAL_FAMILY.sampling,
max_attempts=3,
)
@@ -248,10 +251,11 @@ class TestSampleMaterialV2:
result = sample_material_v2(
tree=real_tree,
family_spec=REASONING_FAMILY,
task_type="Causal Reasoning",
task_type="Action Reasoning",
used_node_ids=set(),
rng=rng,
level=2,
constraint=REASONING_FAMILY.sampling,
)
# cross_l2_span=True 时必须有跨 L2 文本
@@ -265,10 +269,11 @@ class TestSampleMaterialV2:
result = sample_material_v2(
tree=real_tree,
family_spec=RETRIEVAL_FAMILY,
task_type="Action Reasoning",
used_node_ids=set(),
rng=rng,
level=2,
constraint=RETRIEVAL_FAMILY.sampling,
)
# real_tree 所有节点都有字幕，所以 subtitle_sentences 非空