feat(actpred): HailoPredictor 适配 v4 onehot HEF——1864 logits 解码 + 固定列覆盖

v4 onehot 模型 HEF 输出 (1,1,1864),原回归假设(输出=action_dim)导致
输入分类与维度校验失败。_classify_inputs 改为 pred_token 先判(s[1]==1)、
其余 3 维归 history;_derive_dims 识别 TOTAL_BINS 输出;predict 经 bins.decode
期望值解码并固定权重 0 列(与 predict.py 同语义),新增 column_weights 参数。
_npu_test.py 指向 v4 并增加 onehot 参考分支,NPU 实测 4 角 PASS、
HEF vs torch max_diff 0.5°。
This commit is contained in:
qing committed 2026-08-17 14:29:21 +08:00
1 parent 63c29680ce
commit 0fa32cb5e7
2 files changed
+80 -33

No files matched your search

+42 -10
View File
@@ -39,6 +39,7 @@ from imu_norm import normalize_imu
from action_norm import (normalize_action, denormalize_action,
is_range_mode)
from utils import parse_action_history
from bins import TOTAL_BINS, decode as bin_decode, fixed_zero_columns
IMAGENET_MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
IMAGENET_STD = np.array([0.229, 0.224, 0.225], dtype=np.float32)
@@ -62,7 +63,7 @@ class HailoPredictor:
def __init__(self, hef_path: str, stats_path: str = None, vdevice=None,
predict_delta: bool = False, pred_token=None,
normalize_mode: str = "zscore"):
normalize_mode: str = "zscore", column_weights=None):
"""HEF 推理器。
predict_delta: 模型用 --predict-delta 训练时须置 True,推理输出
@@ -72,12 +73,19 @@ class HailoPredictor:
normalize_mode: 与训练时 DataConfig.normalize_mode 一致。
"zscore"(默认)需传 stats_path;"fixed_range" / "log_range"
免 stats 文件,按固定物理范围归一化(action_norm.py)。
column_weights: 逐列损失权重(onehot HEF 的权重 0 列按
bins.FIXED_OUTPUT 固定,防闭环垃圾回写);None 时用
action_norm.default_column_weights。
"""
from hailo_platform import HEF, VDevice
self._predict_delta = predict_delta
self._range_mode = (normalize_mode
if is_range_mode(normalize_mode) else None)
self._column_weights = None
if column_weights is not None:
self._column_weights = np.asarray(column_weights,
dtype=np.float32)
self._default_pred_token = None
if pred_token is not None:
if isinstance(pred_token, str):
@@ -132,7 +140,6 @@ class HailoPredictor:
imu 为 R2 方案一新增输入(1,1,1,imu_dim),旧 HEF 无此输入。
"""
ins = list(self._in_infos)
out_dim = tuple(self._out_info.shape)[-1]
# 新版:每帧独立 (H,W,3) 输入
frames = [i for i in ins if len(i.shape) == 3
and i.shape[2] == 3 and i.shape[0] > 3]
@@ -145,10 +152,10 @@ class HailoPredictor:
history, pred, imu = None, None, None
for i in rest:
s = tuple(i.shape)
if len(s) == 3 and s[-1] == out_dim:
history = i
elif len(s) == 3 and s[1] == 1:
pred = i
if len(s) == 3 and s[1] == 1:
pred = i # pred_token (1,1,d)
elif len(s) == 3:
history = i # 动作历史 (1,H_len,D)
elif len(s) == 4 and s[0] == 1 and s[1] == 1 and s[2] == 1:
imu = i # R2:imu (1,1,1,imu_dim)
else:
@@ -159,7 +166,8 @@ class HailoPredictor:
raise ValueError("HEF 输入中未找到图像输入")
if history is None:
raise ValueError(
f"HEF 输入中未找到动作历史输入(输出维度 {out_dim}),"
f"HEF 输入中未找到动作历史输入(输出维度 "
f"{tuple(self._out_info.shape)}),"
f"实际: {[(i.name, tuple(i.shape)) for i in ins]}")
return frames, history, pred, imu
@@ -209,9 +217,16 @@ class HailoPredictor:
out_shape = tuple(self._out_info.shape)
out_dim = out_shape[-1]
if out_dim != self.action_dim:
raise ValueError(
f"输出动作维度 {out_dim} 与动作历史维度 {self.action_dim} 不一致")
if out_dim == TOTAL_BINS:
# onehot 模型:输出为多头 logits,predict 时经 bins.decode
# 解码为 action_dim 原始量纲(无需反归一化)
self._onehot = True
else:
self._onehot = False
if out_dim != self.action_dim:
raise ValueError(
f"输出动作维度 {out_dim} 与动作历史维度 "
f"{self.action_dim} 不一致")
# ---------------- 生命周期 ----------------
@@ -310,6 +325,23 @@ class HailoPredictor:
q = self._out_info.quant_info
result = (result - q.qp_zp) * q.qp_scale
result = result.reshape(-1)
if self._onehot:
# 多头 logits → 期望值解码 → 原始量纲(箱生活在原始空间,
# 无需反归一化);权重 0 列固定为常量(与 predict.py 对齐),
# 防闭环垃圾回写
result = bin_decode(result, self.action_dim).reshape(-1)
if self._column_weights is None:
from action_norm import default_column_weights
self._column_weights = default_column_weights(
self.action_dim)
w = np.asarray(self._column_weights, dtype=np.float32)
fixed = fixed_zero_columns(w, self.action_dim)
result[w[:self.action_dim] <= 0.0] = (
fixed[w[:self.action_dim] <= 0.0])
if not np.all(np.isfinite(result)):
print("WARNING: HEF output contains NaN/Inf, replacing with zeros")
result = np.nan_to_num(result, nan=0.0, posinf=0.0, neginf=0.0)
return result
if self._predict_delta:
# 残差模式:Δ̂ + a_{t-1}(归一化空间),与 predict.py 对齐
result = result + action_history[-1]