fixes bugs to handle non-dict output (#18897)

src/transformers/models/owlvit/modeling_owlvit.py

@@ -1277,7 +1277,7 @@ class OwlViTForObjectDetection(OwlViTPreTrainedModel):
         )
 
         # Resize class token
-        image_embeds = outputs.image_embeds
+        image_embeds = outputs[-3]
         new_size = tuple(np.array(image_embeds.shape) - np.array((0, 1, 0)))
         class_token_out = torch.broadcast_to(image_embeds[:, :1, :], new_size)
 
@@ -1293,11 +1293,11 @@ class OwlViTForObjectDetection(OwlViTPreTrainedModel):
             image_embeds.shape[-1],
         )
         image_embeds = image_embeds.reshape(new_size)
-        text_embeds = outputs.text_embeds
+        text_embeds = outputs[-4]
 
         # Last hidden states from text and vision transformers
-        text_model_last_hidden_state = outputs.text_model_output.last_hidden_state
-        vision_model_last_hidden_state = outputs.vision_model_output.last_hidden_state
+        text_model_last_hidden_state = outputs[-2][0]
+        vision_model_last_hidden_state = outputs[-1][0]
 
         return (text_embeds, image_embeds, text_model_last_hidden_state, vision_model_last_hidden_state)