284 lines
9.5 KiB
Python
284 lines
9.5 KiB
Python
"""三池：held-out test + 验证 + 诊断，分层采样 + 冻结持久化。
|
||
|
||
三池切分对应训练循环中的 DataLoader 阶段——从题目全集中按
|
||
test -> validation -> diagnosis 的顺序 progressive exclusion，
|
||
保证 question_id 互斥。test 池用自然分布（correct_ratio=None），
|
||
验证池/诊断池按对错比例分层采样。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
from dataclasses import dataclass, field
|
||
from typing import TYPE_CHECKING
|
||
|
||
from app.question_gen import stratified_sample
|
||
from core.types import GeneratedQuestion
|
||
|
||
if TYPE_CHECKING:
|
||
from pathlib import Path
|
||
|
||
from app.harness.config import RunConfig
|
||
|
||
|
||
@dataclass
|
||
class Pools:
|
||
"""冻结的三池及其基线指标。
|
||
|
||
字段:
|
||
diagnosis: 诊断池（用于错误归因，对应 loss.backward）。
|
||
validation: 验证池（按类局部验证，每题型有保底样本）。
|
||
test: held-out 测试池（自然分布，用于最终无偏评估）。
|
||
baseline_run_id: 基线 run 标识。
|
||
baseline_val_accuracy: 基线在验证池上的准确率。
|
||
correctness: 三池所有题的 question_id -> 基线是否答对。
|
||
"""
|
||
|
||
diagnosis: list[GeneratedQuestion]
|
||
validation: list[GeneratedQuestion]
|
||
test: list[GeneratedQuestion]
|
||
baseline_run_id: str
|
||
baseline_val_accuracy: float
|
||
correctness: dict[str, bool] = field(default_factory=dict)
|
||
|
||
|
||
def build_pools(
|
||
questions: list[GeneratedQuestion],
|
||
correctness: dict[str, bool],
|
||
diag_cfg: dict,
|
||
val_cfg: dict,
|
||
test_cfg: dict,
|
||
baseline_run_id: str,
|
||
) -> Pools:
|
||
"""先抽 held-out test，再抽验证集，最后抽诊断池，三池互斥。
|
||
|
||
参数:
|
||
questions: 题目全集。
|
||
correctness: question_id -> 基线是否答对。
|
||
diag_cfg: 诊断池采样配置（size/correct_ratio/task_types[/seed]）。
|
||
val_cfg: 验证池采样配置，可含 min_per_class 做按类保底。
|
||
test_cfg: 测试池配置（size[/seed]）；走自然分布，不强制对错比与题型。
|
||
baseline_run_id: 基线 run 标识。
|
||
|
||
返回:
|
||
冻结的三池 Pools。
|
||
|
||
关键实现细节:
|
||
切分顺序 test -> validation -> diagnosis；后两步从剩余题中采样以保证
|
||
question_id 互斥。test 池用 correct_ratio=None 的自然分布采样。
|
||
"""
|
||
test = _sample_excluding(
|
||
questions,
|
||
set(),
|
||
correctness,
|
||
size=test_cfg["size"],
|
||
correct_ratio=None,
|
||
task_types=None,
|
||
seed=test_cfg.get("seed", 0),
|
||
min_per_class=None,
|
||
)
|
||
selected_ids = {q.question_id for q in test}
|
||
|
||
validation = _sample_excluding(questions, selected_ids, correctness, **val_cfg)
|
||
selected_ids |= {q.question_id for q in validation}
|
||
|
||
diagnosis = _sample_excluding(questions, selected_ids, correctness, **diag_cfg)
|
||
|
||
val_correct = sum(1 for q in validation if correctness.get(q.question_id))
|
||
baseline_val_accuracy = val_correct / len(validation) if validation else 0.0
|
||
return Pools(
|
||
diagnosis=diagnosis,
|
||
validation=validation,
|
||
test=test,
|
||
baseline_run_id=baseline_run_id,
|
||
baseline_val_accuracy=baseline_val_accuracy,
|
||
correctness={
|
||
q.question_id: correctness.get(q.question_id, False)
|
||
for q in test + validation + diagnosis
|
||
},
|
||
)
|
||
|
||
|
||
def _sample_excluding(
|
||
questions: list[GeneratedQuestion],
|
||
exclude_ids: set[str],
|
||
correctness: dict[str, bool],
|
||
**cfg: object,
|
||
) -> list[GeneratedQuestion]:
|
||
"""排除已选 question_id 后，按 cfg 对剩余题做分层采样。
|
||
|
||
参数:
|
||
questions: 题目全集。
|
||
exclude_ids: 已被其他池选走的 question_id，从候选中剔除以保证三池互斥。
|
||
correctness: question_id -> 基线是否答对。
|
||
cfg: 透传给 stratified_sample 的采样配置
|
||
（size/correct_ratio/task_types[/seed/min_per_class]）。
|
||
|
||
返回:
|
||
采样后的题目列表。
|
||
"""
|
||
pool = [q for q in questions if q.question_id not in exclude_ids]
|
||
return stratified_sample(pool, correctness, **cfg)
|
||
|
||
|
||
def _q_to_dict(q: GeneratedQuestion) -> dict:
|
||
"""将 GeneratedQuestion 转为可序列化字典。
|
||
|
||
参数:
|
||
q: 题目对象。
|
||
|
||
返回:
|
||
包含全部字段的字典（options/source_nodes 从 tuple 转为 list）。
|
||
"""
|
||
return {
|
||
"question_id": q.question_id,
|
||
"video_id": q.video_id,
|
||
"task_type": q.task_type,
|
||
"question": q.question,
|
||
"options": list(q.options),
|
||
"answer": q.answer,
|
||
"source_nodes": list(q.source_nodes),
|
||
"difficulty": q.difficulty,
|
||
}
|
||
|
||
|
||
def _dict_to_q(d: dict) -> GeneratedQuestion:
|
||
"""从字典恢复 GeneratedQuestion。
|
||
|
||
参数:
|
||
d: 由 _q_to_dict 产出的字典。
|
||
|
||
返回:
|
||
恢复的 GeneratedQuestion 实例（options/source_nodes 恢复为 tuple）。
|
||
"""
|
||
return GeneratedQuestion(
|
||
question_id=d["question_id"],
|
||
video_id=d["video_id"],
|
||
task_type=d["task_type"],
|
||
question=d["question"],
|
||
options=tuple(d["options"]),
|
||
answer=d["answer"],
|
||
source_nodes=tuple(d.get("source_nodes", ())),
|
||
difficulty=d.get("difficulty", "medium"),
|
||
)
|
||
|
||
|
||
def save_pools(pools: Pools, path: Path) -> None:
|
||
"""将三池及基线指标冻结为 JSON。
|
||
|
||
参数:
|
||
pools: 待冻结的三池。
|
||
path: 目标 JSON 文件路径。
|
||
"""
|
||
path.write_text(
|
||
json.dumps(
|
||
{
|
||
"baseline_run_id": pools.baseline_run_id,
|
||
"baseline_val_accuracy": pools.baseline_val_accuracy,
|
||
"correctness": pools.correctness,
|
||
"diagnosis": [_q_to_dict(q) for q in pools.diagnosis],
|
||
"validation": [_q_to_dict(q) for q in pools.validation],
|
||
"test": [_q_to_dict(q) for q in pools.test],
|
||
},
|
||
ensure_ascii=False,
|
||
indent=2,
|
||
),
|
||
encoding="utf-8",
|
||
)
|
||
|
||
|
||
def load_pools(path: Path) -> Pools:
|
||
"""从 JSON 恢复冻结的三池。
|
||
|
||
参数:
|
||
path: 冻结的 pools.json 路径。
|
||
|
||
返回:
|
||
恢复的三池 Pools。
|
||
|
||
异常:
|
||
ValueError: 旧格式 pools.json（无 test 池）。
|
||
|
||
关键实现细节:
|
||
旧格式 pools.json（无 test 池）会以清晰的 ValueError 中止——本项目不做
|
||
向后兼容，也不为缺失字段填默认值。删除旧文件后 build_pools 会重新采样切分，
|
||
无需重新推理。
|
||
"""
|
||
d = json.loads(path.read_text(encoding="utf-8"))
|
||
if "test" not in d:
|
||
raise ValueError(
|
||
f"{path} 为旧格式 pools.json（缺 test 池），"
|
||
"请删除后重新切分（build_pools 会重新采样，无需重新推理）。"
|
||
)
|
||
return Pools(
|
||
diagnosis=[_dict_to_q(x) for x in d["diagnosis"]],
|
||
validation=[_dict_to_q(x) for x in d["validation"]],
|
||
test=[_dict_to_q(x) for x in d["test"]],
|
||
baseline_run_id=d["baseline_run_id"],
|
||
baseline_val_accuracy=d["baseline_val_accuracy"],
|
||
correctness=d["correctness"],
|
||
)
|
||
|
||
|
||
def build_or_load_pools(
|
||
config: RunConfig,
|
||
run_id: str,
|
||
task_types: list[str] | None = None,
|
||
) -> Pools:
|
||
"""train 模式的三池获取入口：pools.json 已存在则加载，否则从基线 db 切分并冻结。
|
||
|
||
把 main.py train 分支「pools.json 存在则 load_pools 否则 build_pools 再 save_pools」
|
||
那段抽成纯函数，使 main 与集成测试共用同一切分逻辑、避免重复。pools.json 是
|
||
一次 fresh 训练的冻结切分，resume/重跑同一 workspace 时直接复用以保证三池一致。
|
||
|
||
参数:
|
||
config: 运行配置，提供 workspace_dir 与三池采样旋钮（diag/val/test 各项）。
|
||
run_id: 基线全量记录的 run_id（fresh 时来自 seed.json，决定从哪个 run 读对错）。
|
||
task_types: 可选题型过滤，限定诊断/验证池只采样这些题型；None 表示不过滤。
|
||
|
||
返回:
|
||
冻结的三池 Pools。
|
||
|
||
关键实现:
|
||
切分前从基线 db 的 predictions 表读该 run_id 的逐题对错，作为分层采样依据。
|
||
pools.json 落在 config.workspace_dir 下，存在即视为已冻结，原样加载不重切。
|
||
"""
|
||
from app.harness.log import HarnessLog
|
||
from app.harness.workspace import resolve_paths
|
||
from app.question_gen import load_benchmark
|
||
|
||
pools_path = config.workspace_dir / "pools.json"
|
||
if pools_path.exists():
|
||
return load_pools(pools_path)
|
||
|
||
paths = resolve_paths(config.workspace_dir)
|
||
questions = load_benchmark(paths.questions_dir)
|
||
with HarnessLog(str(paths.db_path), run_id) as log:
|
||
rows = log.query(
|
||
"SELECT question_id, prediction, answer FROM predictions WHERE run_id=?",
|
||
(run_id,),
|
||
)
|
||
correctness = {r["question_id"]: r["prediction"] == r["answer"] for r in rows}
|
||
pools = build_pools(
|
||
questions,
|
||
correctness,
|
||
diag_cfg={
|
||
"size": config.diag_size,
|
||
"correct_ratio": config.diag_correct_ratio,
|
||
"task_types": task_types,
|
||
"seed": 0,
|
||
"min_per_class": None,
|
||
},
|
||
val_cfg={
|
||
"size": config.val_size,
|
||
"correct_ratio": config.val_correct_ratio,
|
||
"task_types": task_types,
|
||
"seed": 0,
|
||
"min_per_class": config.eval_min_per_class,
|
||
},
|
||
test_cfg={"size": config.test_size},
|
||
baseline_run_id=run_id,
|
||
)
|
||
save_pools(pools, pools_path)
|
||
return pools
|