Remove double bias (#2958)

src/transformers/modeling_albert.py

@@ -600,7 +600,7 @@ class AlbertMLMHead(nn.Module):
         hidden_states = self.LayerNorm(hidden_states)
         hidden_states = self.decoder(hidden_states)
 
-        prediction_scores = hidden_states + self.bias
+        prediction_scores = hidden_states
 
         return prediction_scores
 

src/transformers/modeling_bert.py

@@ -471,7 +471,7 @@ class BertLMPredictionHead(nn.Module):
 
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
-        hidden_states = self.decoder(hidden_states) + self.bias
+        hidden_states = self.decoder(hidden_states)
         return hidden_states
 
 

src/transformers/modeling_roberta.py

@@ -264,7 +264,7 @@ class RobertaLMHead(nn.Module):
         x = self.layer_norm(x)
 
         # project back to size of vocabulary with bias
-        x = self.decoder(x) + self.bias
+        x = self.decoder(x)
 
         return x