Replace AdaptiveAvgPool2d with fixed-kernel AvgPool2d for ONNX compatibility

AdaptiveAvgPool2d with None dimensions can cause issues with some ONNX
runtimes. Use AvgPool2d with kernel=(img_h//16, 1) to achieve the same
height-to-1 reduction with full ONNX compatibility.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

models/lite_crnn.py

@@ -66,9 +66,11 @@ class LiteCRNN(nn.Module):
             nn.MaxPool2d(2, 2),  # H/2, W/2
         )
 
-        # 经过 4 次高度 pool: img_h / 16 (如 40 → 2, 不够整除时用自适应)
-        # 用 AdaptiveAvgPool 把高度压到 1
-        self.adaptive_pool = nn.AdaptiveAvgPool2d((1, None))  # (B, 128, 1, W')
+        # 经过 4 次高度 pool 后高度 = img_h/16 (向下取整)
+        # 例: 40→20→10→5→2
+        # 用固定 kernel 的 AvgPool 把剩余高度压到 1 (ONNX 兼容)
+        self._pool_h = img_h // 16  # 40→2, 若有余数也向下取整
+        self.height_pool = nn.AvgPool2d(kernel_size=(self._pool_h, 1))
 
         # ---- RNN 序列建模 ----
         self.rnn_input_size = 128
@@ -95,7 +97,7 @@ class LiteCRNN(nn.Module):
         """
         # CNN
         conv = self.cnn(x)                        # (B, 128, H', W')
-        conv = self.adaptive_pool(conv)            # (B, 128, 1, W')
+        conv = self.height_pool(conv)              # (B, 128, 1, W')
         conv = conv.squeeze(2)                     # (B, 128, W')
         conv = conv.permute(0, 2, 1)               # (B, W', 128) — batch_first 序列