average over multiple prompts

detectron2/modeling/meta_arch/clip_rcnn.py

@@ -751,8 +751,10 @@ class CLIPFastRCNN(nn.Module):
             text_features = self.backbone.encode_text(queries)
         else:
             features = self.backbone(images.tensor)
-            token_embeddings = pre_tokenize([queries])[:, 0].to(images.tensor.device)
+            token_embeddings = pre_tokenize([queries]).to(images.tensor.device)[0]
             text_features = self.lang_encoder.encode_text(token_embeddings)
+            text_features = text_features.mean(0, keepdim=True)
+            text_features = text_features / text_features.norm(dim=-1, keepdim=True)  
 
         if self.backbone_type == "resnet":
             head = self.backbone.layer4