From 37feb091853d900ac415a60b3ec21e9880b70fb9 Mon Sep 17 00:00:00 2001
From: qing <1776177662@qq.com>
Date: Sat, 15 Aug 2026 09:55:22 +0000
Subject: [PATCH] =?UTF-8?q?fix(actpred):=20=E5=9B=BA=E5=AE=9A=200=20?=
 =?UTF-8?q?=E5=88=97=E6=9C=AA=E8=A2=AB=E8=A6=86=E7=9B=96=E2=80=94=E2=80=94?=
 =?UTF-8?q?=E6=8E=A9=E7=A0=81=E6=8C=89=E6=9D=83=E9=87=8D=E8=80=8C=E9=9D=9E?=
 =?UTF-8?q?=E5=B8=B8=E9=87=8F=E5=80=BC?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

fixed_zero_columns 对 spk.F 返回常量 0，调用处掩码 fixed!=0 将其
排除，spk.F 垃圾预测（69-423Hz）继续进闭环。改为按权重<=0 掩码
覆盖（含固定 0 的列）。修正后舵机闭环 MSE 14477 假象 → 293 真实
（hand 类 55K 是固定 0 vs 真值 1000 的指标伪影）。+1 回归测试。
---
 actpred/predict.py      |  6 ++++--
 actpred/test_actpred.py | 16 ++++++++++++++++
 actpred/train.py        |  4 ++--
 3 files changed, 22 insertions(+), 4 deletions(-)

diff --git a/actpred/predict.py b/actpred/predict.py
index 9c3e77e..0fccb7b 100644
--- a/actpred/predict.py
+++ b/actpred/predict.py
@@ -128,8 +128,10 @@ class ActionPredictorInference:
                 from action_norm import default_column_weights
                 col_w = default_column_weights(self.action_dim)
             fixed = fixed_zero_columns(col_w, self.action_dim)
-            m = fixed != 0.0
-            result[m] = fixed[m]
+            # 按权重掩码覆盖（固定 0 的列也须覆盖——掩码不能只看常量值）
+            w = np.asarray(col_w, dtype=np.float32)
+            result[w[:self.action_dim] <= 0.0] = (
+                fixed[w[:self.action_dim] <= 0.0])
             if not np.all(np.isfinite(result)):
                 result = np.nan_to_num(result, nan=0.0, posinf=0.0, neginf=0.0)
             return result
diff --git a/actpred/test_actpred.py b/actpred/test_actpred.py
index 6f212d8..5641495 100644
--- a/actpred/test_actpred.py
+++ b/actpred/test_actpred.py
@@ -2525,6 +2525,22 @@ class TestOneHot(unittest.TestCase):
         cfg2 = apply_overrides(Config(), parser.parse_args([]))
         self.assertEqual(cfg2.train.val_split_mode, "sample")  # 新默认
 
+    def test_fixed_zero_value_still_applied(self):
+        """固定 0 的列（spk.F）也必须被覆盖——掩码按权重而非常量值"""
+        from bins import decode, fixed_zero_columns, START
+        from action_norm import default_column_weights
+        # 构造 spk.F 头输出 1000Hz 附近 logits（垃圾预测）
+        logits = np.full((1, 1864), -10.0, dtype=np.float32)
+        logits[0, START[10] + 25] = 0.0      # 1000Hz 附近箱
+        raw = decode(logits, 18)
+        self.assertGreater(float(raw[0, 10]), 100.0)   # 未覆盖时是垃圾
+        w = default_column_weights(18)
+        fixed = fixed_zero_columns(w, 18)
+        mask = np.asarray(w, dtype=np.float32) <= 0.0
+        raw[0, mask] = fixed[mask]
+        self.assertEqual(float(raw[0, 10]), 0.0)      # 固定后归零
+        self.assertEqual(float(raw[0, 11]), 10.0)
+
     def test_action_jitter(self):
         """角度抖动增强：历史加有界高斯噪声，目标保持干净"""
         with tempfile.TemporaryDirectory() as tmpdir:
diff --git a/actpred/train.py b/actpred/train.py
index 1a5e214..6a0f7fd 100644
--- a/actpred/train.py
+++ b/actpred/train.py
@@ -130,8 +130,8 @@ def train_one_epoch(model, loader, optimizer, device, max_grad_norm: float = Non
                         # 权重 0 列固定为常量，防自预测垃圾进入历史
                         fixed = fixed_zero_columns(
                             weights.cpu().numpy(), action_dim)
-                        m = fixed != 0.0
-                        raw[m] = fixed[m]
+                        w = weights.cpu().numpy()[:action_dim]
+                        raw[w <= 0.0] = fixed[w <= 0.0]
                         abs_action = torch.from_numpy(normalize_action(
                             raw, action_dim, mode=range_mode)).to(device)
                     else: