2026-10-09 16:41:13 +08:00

178 lines
8.0 KiB
Python

#!/usr/bin/env python3
"""校准数据构造: LeRobot v3 数据集 + 契约 key_map -> SDK vla_dataset 布局。
输入契约 (与训练镜像一致, 见 quant_design_v1.md §4/§6):
DATASET_PATH LeRobot v3 数据集根 (meta/info.json) 或共父目录多库聚合 (单库为首个验证形态)
CHECKPOINT_PATH convert 产物 (含 model_structure.json), 契约优先来源; CONTRACT_JSON 兜底
NORM_STATS_PATH checkpoint assets 内 norm_stats.json (state q01/q99 归一化)
OUTPUT_DIR 必填; 产出 OUTPUT_DIR/quant_input/calib/
CALIB_SIZE 采样条数, 默认 32
SEED 采样种子, 默认 0
产物 (SDK vla_dataset, vendored datasets/vla/vla_dataset.py 消费):
action/{sid}/state.npy [valid_state_dim] float32, quantile 归一化到 [-1,1]
action/{sid}/x_t.npy [action_horizon, action_dim] float32, 固定种子噪声
image/{sid}/image_{i}.jpg uint8 原始帧 (不做 letterbox/归一化, SDK input_preprocess 自理)
prompt.json [{"text": 任务文本}] 按 sid 顺序
calib_manifest.json 采样清单 (可复现)
"""
from __future__ import annotations
import json
import logging
import os
import sys
from pathlib import Path
import numpy as np
sys.path.insert(0, str(Path(__file__).resolve().parent))
import contract as _contract # noqa: E402
# 本地数据集契约: 拒绝网络 (与训练镜像 HF_HUB_OFFLINE=1 同语义), 须在 lerobot import 前生效
os.environ.setdefault("HF_HUB_OFFLINE", "1")
os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
logger = logging.getLogger("make_calib_data")
def load_contract() -> _contract.Contract:
ckpt = os.environ.get("CHECKPOINT_PATH", "").strip()
if ckpt:
structure = Path(ckpt) / "model_structure.json"
if structure.is_file():
return _contract.Contract.from_model_structure(str(structure))
raw = os.environ.get("CONTRACT_JSON", "").strip()
if raw:
return _contract.Contract.from_json(raw)
sys.exit("[hyperparam] 需要 CHECKPOINT_PATH(含 model_structure.json) 或 CONTRACT_JSON")
def resolve_dataset_root(raw: str) -> tuple[Path, str]:
"""单库根 (meta/info.json) 或父目录单库聚合; 多库聚合暂不支持并显式报错。"""
root = Path(raw)
if (root / "meta" / "info.json").is_file():
return root, root.name
subs = sorted(p for p in root.iterdir() if (p / "meta" / "info.json").is_file()) if root.is_dir() else []
if len(subs) == 1:
return subs[0], subs[0].name
if not subs:
sys.exit(f"[dataset] 未找到 LeRobot 数据集 (meta/info.json): {root}")
sys.exit(f"[dataset] 多库聚合校准暂不支持 ({len(subs)} 库), 请指定单库: {[s.name for s in subs]}")
def quantile_norm(x: np.ndarray, q01: np.ndarray, q99: np.ndarray) -> np.ndarray:
span = q99 - q01
if np.any(span <= 0):
bad = np.where(span <= 0)[0].tolist()
sys.exit(f"[norm] q99<=q01 的维度: {bad}")
y = (2.0 * x - (q01 + q99)) / span
return np.clip(y, -1.0, 1.0)
def main() -> None:
output_dir = Path(os.environ.get("OUTPUT_DIR", "").strip())
if not output_dir:
sys.exit("[hyperparam] OUTPUT_DIR not set")
ds_raw = os.environ.get("DATASET_PATH", "").strip()
if not ds_raw:
sys.exit("[hyperparam] DATASET_PATH not set")
ns_path = Path(os.environ.get("NORM_STATS_PATH", "").strip())
if not ns_path.is_file():
sys.exit(f"[hyperparam] NORM_STATS_PATH 无效: {ns_path}")
calib_size = int(os.environ.get("CALIB_SIZE", "32"))
seed = int(os.environ.get("SEED", "0"))
ct = load_contract()
ds_root, ds_name = resolve_dataset_root(ds_raw)
# 契约 key_map 逆推: 槽位 -> 外部键
cam_keys = []
for slot in ct.cameras:
src = [k for k, v in ct.key_map.items() if isinstance(v, list) and slot in v]
if len(src) != 1:
sys.exit(f"[key_map] 槽位 {slot} 的图像映射不唯一: {src}")
cam_keys.append(src[0])
state_src = [k for k in ct.key_map if ct.key_map[k] == "state"]
prompt_src = [k for k in ct.key_map if ct.key_map[k] == "prompt"]
if len(state_src) != 1 or len(prompt_src) != 1:
sys.exit(f"[key_map] state/prompt 映射不唯一: {state_src} {prompt_src}")
state_key, _ = state_src[0], prompt_src[0] # prompt 文本取 item 的 task 字段 (key_map prompt 源键仅作校验)
ns = json.loads(ns_path.read_text())["norm_stats"]
q01 = np.asarray(ns["state"]["q01"], dtype=np.float64)
q99 = np.asarray(ns["state"]["q99"], dtype=np.float64)
if len(q01) != ct.state_dim:
sys.exit(f"[norm] norm_stats 维度 {len(q01)} != 契约 state.dim {ct.state_dim}")
# ---- LeRobot v3 读取 (与训练 data_loader 同构: HF_LEROBOT_HOME=数据集父目录, 不传 root) ----
try:
import lerobot.datasets.lerobot_dataset as lerobot_dataset
except ImportError:
import lerobot.common.datasets.lerobot_dataset as lerobot_dataset
os.environ.setdefault("HF_LEROBOT_HOME", str(ds_root.parent))
ds = lerobot_dataset.LeRobotDataset(ds_name)
total = len(ds)
if total < calib_size:
sys.exit(f"[dataset] 帧数 {total} < CALIB_SIZE {calib_size}")
rng = np.random.default_rng(seed)
indices = sorted(rng.choice(total, size=calib_size, replace=False).tolist())
logger.info("[calib] dataset=%s frames=%d -> %d 样本 (seed=%d)", ds_root, total, calib_size, seed)
action_dim, horizon = ct.action_dim, ct.action_horizon
noise_rng = np.random.default_rng(seed + 1)
calib_dir = output_dir / "quant_input" / "calib"
img_root, act_root = calib_dir / "image", calib_dir / "action"
img_root.mkdir(parents=True, exist_ok=True)
act_root.mkdir(parents=True, exist_ok=True)
from PIL import Image # noqa: E402
prompts: list[dict] = []
manifest = {"dataset": str(ds_root), "seed": seed, "calib_size": calib_size,
"action_horizon": horizon, "action_dim": action_dim, "samples": []}
for sid, fi in enumerate(indices):
item = ds[fi]
state = np.asarray(item[state_key], dtype=np.float64)[: ct.state_dim]
state_n = quantile_norm(state, q01, q99).astype(np.float32)
sdir = act_root / str(sid)
sdir.mkdir(exist_ok=True)
np.save(sdir / "state.npy", state_n)
np.save(sdir / "x_t.npy", noise_rng.standard_normal((horizon, action_dim)).astype(np.float32))
idir = img_root / str(sid)
idir.mkdir(exist_ok=True)
for i, ck in enumerate(cam_keys):
if ck not in item:
sys.exit(f"[dataset] 帧缺相机键 {ck} (契约槽位 cam{i+1})")
t = np.asarray(item[ck])
if t.ndim != 3:
sys.exit(f"[dataset] 相机 {ck} 帧非 3 维: {t.shape}")
if t.dtype.kind == "f": # lerobot 视频帧: CHW float [0,1]
if t.shape[0] == 3 and t.shape[-1] != 3:
t = t.transpose(1, 2, 0)
t = (np.clip(t, 0.0, 1.0) * 255.0).round().astype(np.uint8)
frame = t.astype(np.uint8)
Image.fromarray(frame[..., :3] if frame.shape[-1] > 3 else frame).save(idir / f"image_{i}.jpg", quality=95)
task_text = str(item.get("task", "")).strip()
if not task_text:
sys.exit(f"[dataset] 帧 {fi} 无任务文本 (task 字段)")
prompts.append({"text": task_text})
manifest["samples"].append({"sid": sid, "frame_index": int(fi),
"episode_index": int(item.get("episode_index", -1)), "task": task_text})
if (sid + 1) % 8 == 0 or sid + 1 == calib_size:
print(f"DROBOTICFLOW_PROGRESS {int((sid + 1) / calib_size * 100)} make_calib_data {sid + 1}/{calib_size}")
(calib_dir / "prompt.json").write_text(json.dumps(prompts, ensure_ascii=False, indent=1), encoding="utf-8")
(calib_dir / "calib_manifest.json").write_text(json.dumps(manifest, ensure_ascii=False, indent=1), encoding="utf-8")
logger.info("[calib] 完成: %s", calib_dir)
if __name__ == "__main__":
main()