scripts/report_rgbd_wide_scale_carrier.py
#!/usr/bin/env python3
"""Report the saved V15 results: pooled EVAL-V3/V4 primary gates, 2x2, DEV (no model/media)."""
# ruff: noqa: E501 -- Chinese scientific report prose
import argparse
import json
from pathlib import Path
V14 = Path("outputs/EXP-3D-RGBD-DATA-SCALE-V14")
METHODS = (
"CONSTANT",
"REPROJ_NN",
"REPROJ_HARMONIC",
"W32_T72",
"W64_T72",
"C0",
"C1",
"HFILL8_W32_T72",
"HFILL8_W64_T72",
"HFILL8_C0",
"HFILL8_C1",
"AHFILL8_C0",
"AHFILL8_C1",
"C0_LAST",
"C1_LAST",
"HFILL8_C0_LAST",
"HFILL8_C1_LAST",
)
NEXT = {
"A": "在 127 个训练场景上，宽度 64 改善了未观测区域的补全，更宽的 carrier 在可部署组合中也超过了经典几何补洞：补全需要同时扩大模型与数据，这支持在 GPU 上继续扩大两者。",
"B": "两个条件只成立一个，需要分开解读：是补全本身随宽度改善、但还不足以胜过经典补洞，还是组合的优势并非来自更宽的 carrier。",
"C": "在 127 个训练场景上，宽度 64 也没有稳定改善补全，组合仍不能胜过经典几何补洞：在 CPU 可及的宽度与数据范围内，这个架构的补全没有扩展。",
}
def read(root, name, default=None):
path = Path(root) / (name + ".json")
if not path.exists() and default is not None:
return default
return json.loads(path.read_text())
def stat(value):
if value is None or value.get("mean") is None:
return "UNDEFINED"
lo, hi = value["ci95"]
return f"{value['mean']:+.4f}，95% CI [{lo:+.4f}, {hi:+.4f}]，场景 {value['positive']}/{value['tied']}/{value['negative']}"
def mean(absolute, method, region):
value = (absolute.get(method, {}).get(region) or {}).get("mean")
return "—" if value is None else f"{value:.4f}"
def c0_reproduces_v14(root, contract):
"""Largest |DEV query AbsRel| difference between V15 C0 and V14 C1, per step."""
diffs = []
for seed in contract["seeds"]:
mine = read(root / "checkpoints" / f"C0_{seed}", "dev_curve", [])
base = read(V14 / "checkpoints" / f"C1_{seed}", "dev_curve", [])
if not mine or len(mine) != len(base):
return None
diffs += [
abs(a["query_absrel"] - b["query_absrel"]) for a, b in zip(mine, base, strict=True)
]
return max(diffs) if diffs else None
def run(root, docs=None):
root, docs = Path(root), Path(docs) if docs else Path(root)
contract = read(root, "training_contract")
result = read(root, "eval_results")
static = read(root, "static_results")
qualification = read(root, "qualification_results")
tests, integrity = (read(root, n, {"status": "PENDING"}) for n in ("tests", "integrity"))
primary = result["primary"]
absolute = primary["absolute_absrel"]
contrasts = primary["contrasts"]
statuses = result["statuses"]
branch = result["INTERPRETATION_BRANCH"]
groups = static["groups"]
reproduction = c0_reproduces_v14(root, contract)
names = (
"WIDTH_AT_SCALE_GAIN",
"HARMONIC_HYBRID_GAIN",
"HFILL8_WIDTH_GAIN",
"WIDTH_AT_T72_GAIN",
"DATA_AT_W32_GAIN",
"DATA_AT_W64_GAIN",
"INTERACTION",
"LAST_WIDTH_GAIN",
"HARMONIC_MINUS_HFILL8_C1_LAST",
"HARMONIC_MINUS_AHFILL8_C1",
)
fields = {
"TRAIN_SCENES": len(contract["train_sets"]["TRAIN_EXT"]),
"EVAL_SCENES": primary["n_scenes"],
"FAR_PIXEL_FRACTION": round(primary["far_pixel_fraction"], 4),
**{
f"EVAL_{method}_{region}": round(absolute[method][region]["mean"], 6)
for method in METHODS
for region in ("ALL", "NEAR", "FAR")
if method in absolute and (absolute[method][region] or {}).get("mean") is not None
},
**{
f"{name}{suffix}": (
(round(gain["mean"], 6) if suffix == "" else gain["ci95"]) if gain else "UNDEFINED"
)
for name, gain in ((name, contrasts[name]["gain"]) for name in names)
for suffix in ("", "_CI")
},
**statuses,
**{
f"{cohort}_{name}": (
round(block["contrasts"][name]["gain"]["mean"], 6)
if block["contrasts"][name]["gain"]
else "UNDEFINED"
)
for cohort, block in result["cohorts"].items()
for name in ("WIDTH_AT_SCALE_GAIN", "HARMONIC_HYBRID_GAIN")
},
"C0_MINUS_V14_C1_EVAL_MAX_ABS": result["consistency"]["c0_minus_v14_c1_max_abs"],
"C0_V14_QUERIES_COMPARED": result["consistency"]["queries_compared"],
"DEV_C0_QUERY_ABSREL": round(
groups["C0|direct"]["query_metrics"]["depth_absrel"]["mean"], 6
),
"DEV_C1_QUERY_ABSREL": round(
groups["C1|direct"]["query_metrics"]["depth_absrel"]["mean"], 6
),
"WIDTH_GAIN_DEV": round(static["surface_gain"]["mean"], 6),
"WIDTH_GAIN_DEV_CI": static["surface_gain"]["ci95"],
"WIDTH_DEV_STATUS": qualification["SURFACE_TRAINING_STATUS"],
"STATIC_DEV_STATUS": qualification["STATIC_DEV_STATUS"],
"C0_REPRODUCES_V14_C1_MAX_DEV_DIFF": reproduction,
"INTERPRETATION_BRANCH": branch,
"FRESH_QUALIFICATION_INCLUDED": False,
"FINAL_HOLDOUT_TOUCHED": False,
"DYNAMIC_TTT_NEXT_STAGE_ALLOWED": False,
"TEST_TIME_INPUT": "RGB+DEPTH+CAMERA",
"TESTS": tests["status"],
"FINAL_INTEGRITY": integrity["status"],
"REPORT_DIR": str(docs),
}
lines = [
"# EXP-3D-RGBD-WIDTH-DATA-V15",
"",
f"宽度 × 数据的检验：在 V14 C1 的配方上（V11 C1 配方，TRAIN72 + TRAIN-EXT 共 {fields['TRAIN_SCENES']} 个训练场景），把 carrier 宽度从 32 加到 64（250,542 个参数）。计划在任何 V15 训练之前写定；C0（宽度 32）必须与 V14 C1 逐位相同。主评价在 EVAL-V3 与 EVAL-V4 的合并队列（{primary['n_scenes']} 个场景）上进行；EVAL-FAR 只作描述。所有预测在读取 query 深度之前封存。全程 CPU。",
"",
f"FAR 为离上下文深度重投影命中点超过 8 像素的像素（占 {primary['far_pixel_fraction']:.1%}）；HFILL8 在其余像素用调和插值补洞，在 FAR 用 carrier；AHFILL8 另把不透明度 < 0.5 的 FAR 像素也交给调和插值（V13）。",
"",
"## 宽度 × 数据（FAR AbsRel，合并队列，描述性）",
"",
"| | 72 个训练场景 | 127 个训练场景 |",
"|---|---:|---:|",
f"| 宽度 32 | {mean(absolute, 'W32_T72', 'FAR')}（V11 C1） | {mean(absolute, 'C0', 'FAR')}（C0 = V14 C1） |",
f"| 宽度 64 | {mean(absolute, 'W64_T72', 'FAR')}（V12 C1） | {mean(absolute, 'C1', 'FAR')}（C1） |",
"",
f"交互作用 [FAR(32,72) − FAR(64,72)] − [FAR(32,127) − FAR(64,127)] = {stat(contrasts['INTERACTION']['gain'])}（< 0 表示更多数据时宽度更有用）。",
"",
"| EVAL-V3 + EVAL-V4 AbsRel | ALL | NEAR | FAR |",
"|---|---:|---:|---:|",
]
for method in METHODS:
if method in absolute:
lines.append(
f"| {method} | "
+ " | ".join(mean(absolute, method, r) for r in ("ALL", "NEAR", "FAR"))
+ " |"
)
lines += [
"",
"## 判定",
"",
f"- WIDTH_AT_SCALE_GAIN = FAR AbsRel(C0 宽度 32) − FAR AbsRel(C1 宽度 64) = {stat(contrasts['WIDTH_AT_SCALE_GAIN']['gain'])}；{statuses['WIDTH_AT_SCALE_STATUS']}",
f"- HARMONIC_HYBRID_GAIN = AbsRel(REPROJ_HARMONIC) − AbsRel(HFILL8 C1) = {stat(contrasts['HARMONIC_HYBRID_GAIN']['gain'])}；{statuses['HARMONIC_HYBRID_LABEL']}",
f"- HFILL8_WIDTH_GAIN = AbsRel(HFILL8 C0) − AbsRel(HFILL8 C1) = {stat(contrasts['HFILL8_WIDTH_GAIN']['gain'])}；{statuses['HFILL8_WIDTH_STATUS']}",
f"- 描述性：72 个场景时宽度的作用 {stat(contrasts['WIDTH_AT_T72_GAIN']['gain'])}；宽度 32 时数据的作用 {stat(contrasts['DATA_AT_W32_GAIN']['gain'])}；宽度 64 时数据的作用 {stat(contrasts['DATA_AT_W64_GAIN']['gain'])}",
f"- 描述性（第 6000 步）：宽度的作用 {stat(contrasts['LAST_WIDTH_GAIN']['gain'])}；调和插值 − HFILL8 C1 = {stat(contrasts['HARMONIC_MINUS_HFILL8_C1_LAST']['gain'])}",
f"- 描述性：调和插值 − AHFILL8 C1 = {stat(contrasts['HARMONIC_MINUS_AHFILL8_C1']['gain'])}",
f"- DEV（次要）：C0 {fields['DEV_C0_QUERY_ABSREL']:.4f}，C1 {fields['DEV_C1_QUERY_ABSREL']:.4f}，WIDTH_GAIN_DEV={stat(static['surface_gain'])}，{fields['WIDTH_DEV_STATUS']}",
f"- 完整性：C0 与 V14 C1 的 DEV 曲线最大差 = {reproduction}；C0 与 V14 C1 封存预测的最大差 = {fields['C0_MINUS_V14_C1_EVAL_MAX_ABS']}（{fields['C0_V14_QUERIES_COMPARED']} 个 query）",
"",
"## 各队列",
"",
"| 队列 | 场景 | WIDTH_AT_SCALE_GAIN | 调和插值 − HFILL8 C1 |",
"|---|---:|---|---|",
]
for cohort, block in result["cohorts"].items():
cells = [
stat(block["contrasts"][name]["gain"])
+ f"，{block['contrasts'][name].get('label', 'UNDEFINED')}"
for name in ("WIDTH_AT_SCALE_GAIN", "HARMONIC_HYBRID_GAIN")
]
lines.append(f"| {cohort} | {block['n_scenes']} | " + " | ".join(cells) + " |")
lines += [
"",
f"分支 {branch}：{NEXT[branch]}",
"",
"说明：DEV 的 8 个场景在 V2–V14 中反复使用，本轮只用它选择 checkpoint 并给出描述性结果；EVAL-V3/V4/FAR 是机制队列，不是资格验证；FRESH-V1/V2 未使用；受保护的 final holdout 未打开。",
]
summary = (
"RGBD WIDTH DATA V15 FINAL\n"
+ "\n".join(
f"{k}={str(v).lower() if isinstance(v, bool) else v}" for k, v in fields.items()
)
+ "\n"
)
(root / "README.md").write_text("\n".join(lines) + "\n")
(root / "STATUS.md").write_text("```text\n" + summary + "```\n")
(root / "terminal_summary.txt").write_text(summary)
print(summary)
return fields
if __name__ == "__main__":
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--root", type=Path, required=True)
parser.add_argument("--docs", type=Path)
args = parser.parse_args()
run(args.root, args.docs)