fix: #14486 do not use BertPooler in DPR (#15068)

* fix: #14486 do not use BertPooler in DPR

* fix tf dpr as well

* finish
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

src/transformers/models/dpr/modeling_dpr.py

@@ -175,7 +175,7 @@ class DPREncoder(DPRPreTrainedModel):
 
     def __init__(self, config: DPRConfig):
         super().__init__(config)
-        self.bert_model = BertModel(config)
+        self.bert_model = BertModel(config, add_pooling_layer=False)
         assert self.bert_model.config.hidden_size > 0, "Encoder hidden_size can't be zero"
         self.projection_dim = config.projection_dim
         if self.projection_dim > 0:
@@ -202,8 +202,9 @@ class DPREncoder(DPRPreTrainedModel):
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
         )
-        sequence_output, pooled_output = outputs[:2]
+        sequence_output = outputs[0]
         pooled_output = sequence_output[:, 0, :]
+
         if self.projection_dim > 0:
             pooled_output = self.encode_proj(pooled_output)
 

src/transformers/models/dpr/modeling_tf_dpr.py

@@ -152,7 +152,7 @@ class TFDPREncoderLayer(tf.keras.layers.Layer):
         super().__init__(**kwargs)
 
         # resolve name conflict with TFBertMainLayer instead of TFBertModel
-        self.bert_model = TFBertMainLayer(config, name="bert_model")
+        self.bert_model = TFBertMainLayer(config, add_pooling_layer=False, name="bert_model")
         self.config = config
 
         assert self.config.hidden_size > 0, "Encoder hidden_size can't be zero"
@@ -198,13 +198,13 @@ class TFDPREncoderLayer(tf.keras.layers.Layer):
             training=inputs["training"],
         )
 
-        sequence_output, pooled_output = outputs[:2]
+        sequence_output = outputs[0]
         pooled_output = sequence_output[:, 0, :]
         if self.projection_dim > 0:
             pooled_output = self.encode_proj(pooled_output)
 
         if not inputs["return_dict"]:
-            return (sequence_output, pooled_output) + outputs[2:]
+            return (sequence_output, pooled_output) + outputs[1:]
 
         return TFBaseModelOutputWithPooling(
             last_hidden_state=sequence_output,