48 lines
1.9 KiB
Python
48 lines
1.9 KiB
Python
#!/usr/bin/env python3
|
|
"""板端原始 x_t -> 与 torch dump 同口径的后处理 (对齐 unified_dumper._postprocess_action)。
|
|
|
|
变换 (vendor tools/vla_eval/action_postprocess.py 语义):
|
|
1) 有效维 denorm: out[..., :14] = out*std[:14] + mean[:14] (norm_stats actions 的 mean/std)
|
|
2) AbsoluteActions: 关节维 (6+1 掩码, 夹爪不偏移) += state[..., :14] (e2e 的归一化 state, 与 torch dump 同源同空间)
|
|
输入: BOARD_RAW_DIR (板端 board_out/{sid}.npy), E2E_DIR (.e2e_data/{sid}/action_i0.bin), NORM_STATS_PATH
|
|
输出: OUT_DIR/{sid}.npy (写进 eval_dump/hbm/hbm/ 供 sample_compare)
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import os
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
import numpy as np
|
|
|
|
ALOHA_DELTA_ACTION_MASK = (True,) * 6 + (False,) + (True,) * 6 + (False,)
|
|
|
|
|
|
def main() -> None:
|
|
raw_dir = Path(os.environ["BOARD_RAW_DIR"])
|
|
e2e_dir = Path(os.environ["E2E_DIR"])
|
|
out_dir = Path(os.environ["OUT_DIR"])
|
|
n_dims = int(os.environ.get("VALID_ACTION_DIM", "14"))
|
|
ns = json.loads(Path(os.environ["NORM_STATS_PATH"]).read_text())
|
|
stats = ns.get("norm_stats", ns)["actions"]
|
|
mean, std = np.array(stats["mean"]), np.array(stats["std"])
|
|
mask = np.asarray(ALOHA_DELTA_ACTION_MASK)
|
|
|
|
out_dir.mkdir(parents=True, exist_ok=True)
|
|
n = 0
|
|
for p in sorted(raw_dir.glob("*.npy"), key=lambda x: int(x.stem)):
|
|
sid = p.stem
|
|
a = np.load(p).astype(np.float64)
|
|
a[..., :n_dims] = a[..., :n_dims] * std[:n_dims] + mean[:n_dims]
|
|
state = np.fromfile(e2e_dir / sid / "action_i0.bin", dtype=np.float16).reshape(1, -1).astype(np.float64)
|
|
contrib = np.where(mask, state[..., :n_dims], 0.0)
|
|
a[..., :n_dims] += contrib[:, np.newaxis, :]
|
|
np.save(out_dir / f"{sid}.npy", a.astype(np.float32))
|
|
n += 1
|
|
print(f"DROBOTICFLOW_PROGRESS 100 postprocess_board {n} files -> {out_dir}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|