qing
35ee614ca9
- robot.py numpy 恢复函数级 lazy import，纯 .mt 脚本执行不再需要 numpy（v3-P1-1） - export_onnx 写入训练配置元数据；onnx_infer 自动识别 normalize_mode/predict_delta 并支持残差加回（v3-P1-2） - pybot --loop 补 .onnx 后端分支，zscore 缺 action_stats.npy 时告警（P2-a 闭环） - 回归测试：actpred TestOnnxMetadata +7、pybot 零依赖导入 +1（88/148 全绿） - 附 v3 审计报告：13 项 v2 修复回归核对 + 7 项新发现（P0-2 遗留 + P2/P3 六项）
231 lines
9.6 KiB
Python
231 lines
9.6 KiB
Python
"""CLI 入口 —— 移植自 bot4cj/src/main.cj
|
||
|
||
用法（与仓颉版管道方式兼容）：
|
||
cat scripts/std.mt | python3 -m pybot # stdin 管道
|
||
python3 -m pybot scripts/std.mt # 位置参数（增强）
|
||
python3 -m pybot --dry-run scripts/std.mt # Mock 总线，动作序列到 stdout
|
||
python3 -m pybot --test roundtrip in.mt out.mt # 解析↔序列化往返
|
||
|
||
与仓颉版的差异（见方案 §4.7）：
|
||
--test nn/cnn/cnn1 移除（NN 由 actpred 承担）；
|
||
--test matrix_o/matrix_2 合并为 --test roundtrip（路径改为参数）；
|
||
新增 --dry-run / --verbose；-i / -o 保留参数位（reserved，对接 actpred）。
|
||
日志规范（S02 根治）：全部诊断走 logging(stderr)，stdout 保持纯净。
|
||
"""
|
||
|
||
import argparse
|
||
import logging
|
||
import sys
|
||
|
||
from .hd import i2c, spi
|
||
from .parser import MatrixParser, MatrixSaver
|
||
from .robot import RobotStruct
|
||
|
||
|
||
def _build_argparser():
|
||
p = argparse.ArgumentParser(
|
||
prog="pybot",
|
||
description="机器人 .mt 动作脚本解析器（bot4cj 的 Python 迁移版）")
|
||
p.add_argument("script", nargs="?", default=None,
|
||
help=".mt 脚本路径（缺省从 stdin 读取）")
|
||
p.add_argument("--test", nargs=3, metavar=("MODE", "IN", "OUT"),
|
||
help="自检：roundtrip <in.mt> <out.mt> 解析↔序列化往返")
|
||
p.add_argument("--dry-run", action="store_true",
|
||
help="Mock 总线运行（不碰硬件），动作序列输出到 stdout")
|
||
p.add_argument("--verbose", action="store_true",
|
||
help="DEBUG 级日志（stderr）")
|
||
p.add_argument("-i", metavar="FIFO", default=None,
|
||
help="(reserved) 传感器数据输出管道，对接 actpred")
|
||
p.add_argument("-o", metavar="FIFO", default=None,
|
||
help="(reserved) 动作指令输入管道，对接 actpred")
|
||
p.add_argument("--record", metavar="DIR", default=None,
|
||
help="录制 actpred 格式 episode 到 DIR（需要 @Camera 块；"
|
||
"--dry-run 下用合成帧）")
|
||
p.add_argument("--action-dim", type=int, default=18,
|
||
help="动作向量维度（默认 18，与 actpred 一致）")
|
||
p.add_argument("--loop", metavar="CKPT", default=None,
|
||
help="实时预测闭环：加载 actpred checkpoint，"
|
||
"采样→推理→执行→回写（不执行脚本 @Action 行，"
|
||
"仅用其表头做动作向量索引映射）")
|
||
p.add_argument("--hz", type=float, default=10.0,
|
||
help="闭环控制频率 Hz（默认 10，须 ≥ 推理+执行耗时）")
|
||
p.add_argument("--loop-steps", type=int, default=None, metavar="N",
|
||
help="闭环步数上限（默认无限，调试用）")
|
||
return p
|
||
|
||
|
||
def _roundtrip(in_path, out_path):
|
||
"""解析 in.mt 再序列化到 out.mt（对应仓颉版 --test matrix_o/matrix_2）"""
|
||
objects = []
|
||
with open(in_path, encoding="utf-8") as f:
|
||
parser = MatrixParser(f)
|
||
while True:
|
||
obj = parser.parse()
|
||
if obj is None:
|
||
break
|
||
objects.append(obj)
|
||
with open(out_path, "w", encoding="utf-8") as f:
|
||
MatrixSaver(f).save(objects)
|
||
logging.info("roundtrip: %d blocks, %s -> %s",
|
||
len(objects), in_path, out_path)
|
||
return 0
|
||
|
||
|
||
def _dump_journal(*mocks):
|
||
"""--dry-run 结束后把 Mock 总线动作序列输出到 stdout（等价性验证工件）"""
|
||
for mock in mocks:
|
||
for entry in mock.journal:
|
||
print(" ".join(str(x) for x in entry))
|
||
|
||
|
||
def main(argv=None):
|
||
args = _build_argparser().parse_args(argv)
|
||
logging.basicConfig(
|
||
stream=sys.stderr,
|
||
level=logging.DEBUG if args.verbose else logging.INFO,
|
||
format="%(levelname)s %(name)s: %(message)s")
|
||
|
||
if args.i or args.o:
|
||
logging.warning("-i/-o are reserved for actpred integration, ignored")
|
||
|
||
if args.test:
|
||
mode, in_path, out_path = args.test
|
||
if mode != "roundtrip":
|
||
logging.error("unknown --test mode: %s (supported: roundtrip)", mode)
|
||
return 2
|
||
return _roundtrip(in_path, out_path)
|
||
|
||
mocks = ()
|
||
if args.dry_run:
|
||
mocks = (i2c.use_mock(), spi.use_mock())
|
||
|
||
predictor = None
|
||
if args.loop:
|
||
predictor = _load_predictor(args.loop)
|
||
if predictor is None:
|
||
return 2
|
||
|
||
source = open(args.script, encoding="utf-8") if args.script else sys.stdin
|
||
try:
|
||
robot = RobotStruct(source, dry_run=args.dry_run)
|
||
recorder = None
|
||
if args.record:
|
||
from .recorder import EpisodeRecorder
|
||
recorder = EpisodeRecorder(args.record,
|
||
action_dim=args.action_dim)
|
||
with robot:
|
||
if predictor is not None:
|
||
robot.run_loop(predictor, hz=args.hz,
|
||
max_steps=args.loop_steps, recorder=recorder)
|
||
else:
|
||
robot.run(recorder=recorder)
|
||
finally:
|
||
if args.script:
|
||
source.close()
|
||
|
||
if args.dry_run:
|
||
_dump_journal(*mocks)
|
||
return 0
|
||
|
||
|
||
def _load_predictor(checkpoint):
|
||
"""进程内构建 actpred 推理器；失败返回 None
|
||
|
||
pybot 不硬依赖 actpred：仅 --loop 时才尝试 import，actpred 不在
|
||
sys.path 时给出可操作提示。checkpoint 后缀决定后端（.hef → Hailo NPU，
|
||
.onnx → ONNX Runtime，否则 → PyTorch），推理类由 actpred 提供。
|
||
|
||
HEF 后端会自动查找同目录下的 pred_token.npy 和 companion checkpoint
|
||
（pretrain_best_inference.pt / best_inference.pt），以恢复 pred_token
|
||
默认值和 predict_delta 训练配置——否则 Hailo 推理输出与训练语义不一致。
|
||
ONNX 后端的 normalize_mode/predict_delta 由模型元数据自动识别。
|
||
"""
|
||
import os
|
||
try:
|
||
lower = checkpoint.lower()
|
||
if lower.endswith(".onnx"):
|
||
from actpred.onnx_infer import OnnxPredictor as _P
|
||
onnx_dir = os.path.dirname(os.path.abspath(checkpoint))
|
||
# stats_path：zscore 归一化模型的 action_stats.npy
|
||
stats_path = None
|
||
sp = os.path.join(onnx_dir, "action_stats.npy")
|
||
if os.path.exists(sp):
|
||
stats_path = sp
|
||
try:
|
||
pred = _P(checkpoint, stats_path=stats_path)
|
||
except Exception:
|
||
logging.exception("加载 ONNX 推理器失败: %s", checkpoint)
|
||
return None
|
||
if pred.normalize_mode == "zscore" and stats_path is None:
|
||
logging.warning("ONNX 元数据为 zscore 归一化但同目录无 "
|
||
"action_stats.npy，动作历史将不归一化: %s",
|
||
checkpoint)
|
||
return pred
|
||
if lower.endswith(".hef"):
|
||
from actpred.hailo_infer import HailoPredictor as _P
|
||
hef_dir = os.path.dirname(os.path.abspath(checkpoint))
|
||
|
||
# pred_token：export_deploy.py --pred-token-out 的产物
|
||
pred_token = None
|
||
pt_path = os.path.join(hef_dir, "pred_token.npy")
|
||
if os.path.exists(pt_path):
|
||
pred_token = pt_path
|
||
logging.info("found pred_token: %s", pt_path)
|
||
|
||
# stats_path：归一化模型的 action_stats.npy
|
||
stats_path = None
|
||
sp = os.path.join(hef_dir, "action_stats.npy")
|
||
if os.path.exists(sp):
|
||
stats_path = sp
|
||
|
||
# predict_delta：从同目录 checkpoint 读取训练配置
|
||
predict_delta = False
|
||
normalize_mode = "zscore"
|
||
for name in ("pretrain_best_inference.pt",
|
||
"best_inference.pt"):
|
||
ckpt_path = os.path.join(hef_dir, name)
|
||
if os.path.exists(ckpt_path):
|
||
try:
|
||
import torch
|
||
ck = torch.load(ckpt_path, map_location="cpu",
|
||
weights_only=True)
|
||
cfg = ck.get("config", {})
|
||
predict_delta = cfg.get("train", {}).get(
|
||
"predict_delta", False)
|
||
normalize_mode = cfg.get("data", {}).get(
|
||
"normalize_mode", "zscore")
|
||
if predict_delta:
|
||
logging.info(
|
||
"predict_delta=True from %s", ckpt_path)
|
||
if normalize_mode != "zscore":
|
||
logging.info(
|
||
"normalize_mode=%s from %s",
|
||
normalize_mode, ckpt_path)
|
||
except Exception:
|
||
pass
|
||
break
|
||
|
||
try:
|
||
return _P(checkpoint, stats_path=stats_path,
|
||
predict_delta=predict_delta,
|
||
pred_token=pred_token,
|
||
normalize_mode=normalize_mode)
|
||
except Exception:
|
||
logging.exception("加载 HEF 推理器失败: %s", checkpoint)
|
||
return None
|
||
else:
|
||
from actpred.predict import ActionPredictorInference as _P
|
||
except ImportError:
|
||
logging.error("--loop 需要 actpred 模块（把 blackbird 根目录加入 "
|
||
"PYTHONPATH，或在 actpred/ 同级运行）")
|
||
return None
|
||
try:
|
||
return _P(checkpoint)
|
||
except Exception:
|
||
logging.exception("加载 actpred checkpoint 失败: %s", checkpoint)
|
||
return None
|
||
|
||
|
||
if __name__ == "__main__":
|
||
sys.exit(main())
|