Fix loss calculation in TFXXXForTokenClassification models (#15294)

* Fix loss calculation in TFFunnelForTokenClassification

* revert the change in TFFunnelForTokenClassification

* fix FunnelForTokenClassification loss

* fix other TokenClassification loss

* fix more

* fix more

* add num_labels to ElectraForTokenClassification

* revert the change to research projects
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/models/rembert/modeling_rembert.py

@@ -1413,16 +1413,7 @@ class RemBertForTokenClassification(RemBertPreTrainedModel):
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)
-                active_labels = torch.where(
-                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
-                )
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 
         if not return_dict:
             output = (logits,) + outputs[2:]

src/transformers/models/roberta/modeling_roberta.py

@@ -1414,16 +1414,7 @@ class RobertaForTokenClassification(RobertaPreTrainedModel):
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)
-                active_labels = torch.where(
-                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
-                )
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 
         if not return_dict:
             output = (logits,) + outputs[2:]

src/transformers/models/roformer/modeling_roformer.py

@@ -1472,16 +1472,7 @@ class RoFormerForTokenClassification(RoFormerPreTrainedModel):
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)
-                active_labels = torch.where(
-                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
-                )
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 
         if not return_dict:
             output = (logits,) + outputs[1:]

src/transformers/models/squeezebert/modeling_squeezebert.py

@@ -984,16 +984,7 @@ class SqueezeBertForTokenClassification(SqueezeBertPreTrainedModel):
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)
-                active_labels = torch.where(
-                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
-                )
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 
         if not return_dict:
             output = (logits,) + outputs[2:]

src/transformers/models/xlm/modeling_xlm.py

@@ -1169,16 +1169,7 @@ class XLMForTokenClassification(XLMPreTrainedModel):
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)
-                active_labels = torch.where(
-                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
-                )
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 
         if not return_dict:
             output = (logits,) + outputs[1:]

src/transformers/models/xlnet/modeling_xlnet.py

@@ -1680,16 +1680,7 @@ class XLNetForTokenClassification(XLNetPreTrainedModel):
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)
-                active_labels = torch.where(
-                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
-                )
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 
         if not return_dict:
             output = (logits,) + outputs[1:]

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py

@@ -1455,16 +1455,7 @@ class {{cookiecutter.camelcase_modelname}}ForTokenClassification({{cookiecutter.
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)
-                active_labels = torch.where(
-                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
-                )
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 
         if not return_dict:
             output = (logits,) + outputs[1:]