remove resize code from prev. trained model - Did not work with all model types

Author: nreimers

sentence_transformers/cross_encoder/CrossEncoder.py

@@ -43,23 +43,11 @@ def __init__(self, model_name:str, num_labels:int = None, max_length:int = None,
         if num_labels is None and not classifier_trained:
             num_labels = 1
 
-        resize_num_labels = None
         if num_labels is not None:
-            if hasattr(self.config, 'num_labels') and self.config.num_labels is not None and self.config.num_labels != num_labels:
-                #Resize classifier head
-                resize_num_labels = num_labels
-            else:
-                self.config.num_labels = num_labels
+            self.config.num_labels = num_labels
 
         self.model = AutoModelForSequenceClassification.from_pretrained(model_name, config=self.config)
         self.tokenizer = AutoTokenizer.from_pretrained(model_name, **tokenizer_args)
-
-        if resize_num_labels is not None:
-            print("Warning: Loaded model was trained for {} labels. Resize CrossEncoder to {} labels. You must re-train this model to get meaningful predictions".format(self.config.num_labels, resize_num_labels))
-            self.config.num_labels = resize_num_labels
-            self.model.config.num_labels = resize_num_labels
-            self.model.classifier = torch.nn.Linear(self.config.hidden_size, resize_num_labels)
-
         self.max_length = max_length
 
         if device is None:
@@ -214,7 +202,6 @@ def fit(self,
                     logits = activation_fct(model_predictions.logits)
                     if self.config.num_labels == 1:
                         logits = logits.view(-1)
-
                     loss_value = loss_fct(logits, labels)
                     loss_value.backward()
                     torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_grad_norm)