refactor: update BehaviorClassifier to use MLP for frame-wise classif…

…ication

- Replaced the Transformer encoder with an MLP architecture for simpler and more efficient classification.
- Integrated the MLP to classify features extracted from each video frame independently.
- Applied global average pooling across frame outputs to produce a single classification per video.

annolid/behavior/models/classifier.py

@@ -50,15 +50,15 @@ class BehaviorClassifier(nn.Module):
 
     Args:
         feature_extractor (nn.Module): The feature extraction module.
-        d_model (int, optional): The embedding dimension. Defaults to 512.
+        d_model (int, optional): The embedding dimension. Defaults to 768.
         nhead (int, optional): The number of attention heads. Defaults to 8.
         num_layers (int, optional): The number of transformer encoder layers. Defaults to 6.
         dim_feedforward (int, optional): The dimension of the feedforward network. Defaults to 2048.
         dropout (float, optional): The dropout probability. Defaults to 0.1.
         num_classes (int, optional): The number of behavior classes. Defaults to 5.
     """
 
-    def __init__(self, feature_extractor: nn.Module, d_model: int = 512, nhead: int = 8,
+    def __init__(self, feature_extractor: nn.Module, d_model: int = 768, nhead: int = 8,
                  num_layers: int = 6, dim_feedforward: int = 2048, dropout: float = 0.1,
                  num_classes: int = 5):
         super().__init__()