Change masking to direct labeling for TPU support. (#2982)

* change masking to direct labelings

* fix black

* switch to ignore index

* .

* fix black

src/transformers/modeling_bert.py

@@ -1382,8 +1382,10 @@ class BertForTokenClassification(BertPreTrainedModel):
             # 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_loss]
-                active_labels = labels.view(-1)[active_loss]
+                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))

src/transformers/modeling_distilbert.py

@@ -818,8 +818,10 @@ class DistilBertForTokenClassification(DistilBertPreTrainedModel):
             # 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_loss]
-                active_labels = labels.view(-1)[active_loss]
+                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))

src/transformers/modeling_roberta.py

@@ -542,13 +542,16 @@ class RobertaForTokenClassification(BertPreTrainedModel):
         logits = self.classifier(sequence_output)
 
         outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
         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_loss]
-                active_labels = labels.view(-1)[active_loss]
+                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))

src/transformers/modeling_xlnet.py

@@ -1264,8 +1264,10 @@ class XLNetForTokenClassification(XLNetPreTrainedModel):
             # 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_loss]
-                active_labels = labels.view(-1)[active_loss]
+                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))