Fix for sequence regression fit() in TF (#19316)

Co-authored-by: Your Name <you@example.com>

src/transformers/modeling_tf_utils.py

@@ -274,6 +274,9 @@ class TFSequenceClassificationLoss:
     def hf_compute_loss(self, labels, logits):
         if logits.shape.rank == 1 or logits.shape[1] == 1:
             loss_fn = tf.keras.losses.MeanSquaredError(reduction=tf.keras.losses.Reduction.NONE)
+            if labels.shape.rank == 1:
+                # MeanSquaredError returns a scalar loss if the labels are 1D, so avoid that
+                labels = tf.expand_dims(labels, axis=-1)
         else:
             loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
                 from_logits=True, reduction=tf.keras.losses.Reduction.NONE