Fix regression in regression (#11785)

* Fix regression in regression

* Add test

src/transformers/models/albert/modeling_albert.py

@@ -1037,7 +1037,10 @@ class AlbertForSequenceClassification(AlbertPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/bert/modeling_bert.py

@@ -1528,7 +1528,10 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/big_bird/modeling_big_bird.py

@@ -2671,7 +2671,10 @@ class BigBirdForSequenceClassification(BigBirdPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/convbert/modeling_convbert.py

@@ -1023,7 +1023,10 @@ class ConvBertForSequenceClassification(ConvBertPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/distilbert/modeling_distilbert.py

@@ -642,7 +642,10 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/electra/modeling_electra.py

@@ -964,7 +964,10 @@ class ElectraForSequenceClassification(ElectraPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/funnel/modeling_funnel.py

@@ -1298,7 +1298,10 @@ class FunnelForSequenceClassification(FunnelPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/longformer/modeling_longformer.py

@@ -1872,7 +1872,10 @@ class LongformerForSequenceClassification(LongformerPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/mobilebert/modeling_mobilebert.py

@@ -1279,7 +1279,10 @@ class MobileBertForSequenceClassification(MobileBertPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/reformer/modeling_reformer.py

@@ -2445,7 +2445,10 @@ class ReformerForSequenceClassification(ReformerPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/roberta/modeling_roberta.py

@@ -1178,7 +1178,10 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/squeezebert/modeling_squeezebert.py

@@ -798,7 +798,10 @@ class SqueezeBertForSequenceClassification(SqueezeBertPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/xlm/modeling_xlm.py

@@ -847,7 +847,10 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

src/transformers/models/xlnet/modeling_xlnet.py

@@ -1562,7 +1562,10 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
 
             if self.config.problem_type == "regression":
                 loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels)
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
             elif self.config.problem_type == "single_label_classification":
                 loss_fct = CrossEntropyLoss()
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

tests/test_modeling_common.py

@@ -20,6 +20,7 @@ import os.path
 import random
 import tempfile
 import unittest
+import warnings
 from typing import List, Tuple
 
 from huggingface_hub import HfApi
@@ -1462,7 +1463,14 @@ class ModelTesterMixin:
 
                     inputs["labels"] = inputs["labels"].to(problem_type["dtype"])
 
+                    # This tests that we do not trigger the warning form PyTorch "Using a target size that is different
+                    # to the input size. This will likely lead to incorrect results due to broadcasting. Please ensure
+                    # they have the same size." which is a symptom something in wrong for the regression problem.
+                    # See https://github.com/huggingface/transformers/issues/11780
+                    with warnings.catch_warnings(record=True) as warning_list:
                         loss = model(**inputs).loss
+                    self.assertListEqual(warning_list, [])
+
                     loss.backward()