Return scalar losses instead of per-sample means (#18013)

* Return scalar losses instead of per-sample means

* Make loss shape (1,) instead of scalar

* Allow scalar losses in test_loss_computation

* Allow scalar losses in test_loss_computation

* Allow scalar losses in test_loss_computation

* Remove XLA loss function for RAG

src/transformers/modeling_tf_utils.py

@@ -206,11 +206,9 @@ class TFCausalLanguageModelingLoss:
         unmasked_loss = loss_fn(tf.nn.relu(labels), logits)
         # make sure only labels that are not equal to -100 affect the loss
         loss_mask = tf.cast(labels != -100, dtype=unmasked_loss.dtype)
-        # Avoid division by zero later
-        loss_denominator = tf.math.maximum(tf.cast(1, loss_mask.dtype), tf.reduce_sum(loss_mask, axis=1))
         masked_loss = unmasked_loss * loss_mask
-        reduced_masked_loss = tf.reduce_sum(masked_loss, axis=1) / loss_denominator
-        return reduced_masked_loss
+        reduced_masked_loss = tf.reduce_sum(masked_loss) / tf.reduce_sum(loss_mask)
+        return tf.reshape(reduced_masked_loss, (1,))
 
 
 class TFQuestionAnsweringLoss:
@@ -266,11 +264,10 @@ class TFTokenClassificationLoss:
         # are taken into account as loss
         loss_mask = tf.cast(labels >= 0, dtype=unmasked_loss.dtype)
         # Avoid possible division by zero later
-        loss_denominator = tf.math.maximum(tf.cast(1, loss_mask.dtype), tf.reduce_sum(loss_mask, axis=1))
         # Masked positions will have a loss of NaN because -100 and -1 are not valid labels
         masked_loss = unmasked_loss * loss_mask
-        reduced_masked_loss = tf.reduce_sum(masked_loss, axis=1) / loss_denominator
-        return reduced_masked_loss
+        reduced_masked_loss = tf.reduce_sum(masked_loss) / tf.reduce_sum(loss_mask)
+        return tf.reshape(reduced_masked_loss, (1,))
 
 
 class TFSequenceClassificationLoss:

src/transformers/models/albert/modeling_tf_albert.py

@@ -118,20 +118,18 @@ class TFAlbertPreTrainingLoss:
         # make sure only labels that are not equal to -100
         # are taken into account for the loss computation
         lm_loss_mask = tf.cast(labels["labels"] != -100, dtype=unmasked_lm_losses.dtype)
-        # Avoid division by zero later
-        lm_loss_denominator = tf.math.maximum(tf.cast(1, lm_loss_mask.dtype), tf.reduce_sum(lm_loss_mask, axis=1))
         masked_lm_losses = unmasked_lm_losses * lm_loss_mask
-        reduced_masked_lm_loss = tf.reduce_sum(masked_lm_losses, axis=1) / lm_loss_denominator
+        reduced_masked_lm_loss = tf.reduce_sum(masked_lm_losses) / tf.reduce_sum(lm_loss_mask)
 
         sop_logits = tf.reshape(logits[1], (-1, 2))
         # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway
         unmasked_sop_loss = loss_fn(y_true=tf.nn.relu(labels["sentence_order_label"]), y_pred=sop_logits)
         sop_loss_mask = tf.cast(labels["sentence_order_label"] != -100, dtype=unmasked_sop_loss.dtype)
 
-        # No reduction because this already has shape (num_samples,)
         masked_sop_loss = unmasked_sop_loss * sop_loss_mask
+        reduced_masked_sop_loss = tf.reduce_sum(masked_sop_loss) / tf.reduce_sum(sop_loss_mask)
 
-        return reduced_masked_lm_loss + masked_sop_loss
+        return tf.reshape(reduced_masked_lm_loss + reduced_masked_sop_loss, (1,))
 
 
 class TFAlbertEmbeddings(tf.keras.layers.Layer):

src/transformers/models/bert/modeling_tf_bert.py

@@ -130,18 +130,17 @@ class TFBertPreTrainingLoss:
         # make sure only labels that are not equal to -100
         # are taken into account for the loss computation
         lm_loss_mask = tf.cast(labels["labels"] != -100, dtype=unmasked_lm_losses.dtype)
-        # Avoid potential division by zero later
-        lm_loss_denominator = tf.math.maximum(tf.cast(1, lm_loss_mask.dtype), tf.reduce_sum(lm_loss_mask, axis=1))
         masked_lm_losses = unmasked_lm_losses * lm_loss_mask
-        reduced_masked_lm_loss = tf.reduce_sum(masked_lm_losses, axis=1) / lm_loss_denominator
+        reduced_masked_lm_loss = tf.reduce_sum(masked_lm_losses) / tf.reduce_sum(lm_loss_mask)
 
         # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway
         unmasked_ns_loss = loss_fn(y_true=tf.nn.relu(labels["next_sentence_label"]), y_pred=logits[1])
         ns_loss_mask = tf.cast(labels["next_sentence_label"] != -100, dtype=unmasked_ns_loss.dtype)
-        # Just zero out samples where label is -100, no reduction
         masked_ns_loss = unmasked_ns_loss * ns_loss_mask
 
-        return reduced_masked_lm_loss + masked_ns_loss
+        reduced_masked_ns_loss = tf.reduce_sum(masked_ns_loss) / tf.reduce_sum(ns_loss_mask)
+
+        return tf.reshape(reduced_masked_lm_loss + reduced_masked_ns_loss, (1,))
 
 
 class TFBertEmbeddings(tf.keras.layers.Layer):

src/transformers/models/led/modeling_tf_led.py

@@ -2518,7 +2518,6 @@ class TFLEDForConditionalGeneration(TFLEDPreTrainedModel):
         unmasked_loss = loss_fn(tf.nn.relu(labels), logits)
         # make sure only non-padding labels affect the loss
         loss_mask = tf.cast(labels != self.config.pad_token_id, dtype=unmasked_loss.dtype)
-        loss_denominator = tf.math.maximum(tf.cast(1, loss_mask.dtype), tf.reduce_sum(loss_mask, axis=1))
         masked_loss = unmasked_loss * loss_mask
-        reduced_masked_loss = tf.reduce_sum(masked_loss, axis=1) / loss_denominator
-        return reduced_masked_loss
+        reduced_masked_loss = tf.reduce_sum(masked_loss) / tf.reduce_sum(loss_mask)
+        return tf.reshape(reduced_masked_loss, (1,))

src/transformers/models/rag/modeling_tf_rag.py

@@ -1333,46 +1333,29 @@ class TFRagTokenForGeneration(TFRagPreTrainedModel, TFCausalLanguageModelingLoss
     # Adopted modeling_tf_bart + add smooth_loss to match with pytorch version
     def hf_compute_loss(self, labels, y_pred, smooth_epsilon=0.0, from_logits=True, reduce_loss=False):
         """CrossEntropyLoss that ignores pad tokens"""
-        if self.config.tf_legacy_loss:
-            loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
-                from_logits=True,
-                reduction=tf.keras.losses.Reduction.SUM,
-            )
-
-            if from_logits is False:  # convert to logits
-                eps = 1e-9
-                y_pred = tf.clip_by_value(y_pred, clip_value_min=eps, clip_value_max=1 - eps)
-                y_pred = tf.math.log(y_pred)
-
-            logits = y_pred
-            melted_labels = tf.reshape(labels, (-1,))
-            active_loss = tf.not_equal(melted_labels, self.config.generator.pad_token_id)
-
-            reduced_logits = tf.boolean_mask(tf.reshape(logits, (-1, logits.shape[2])), active_loss)
-            labels = tf.boolean_mask(melted_labels, active_loss)
-            nll_loss = loss_fn(labels, reduced_logits)
-
-            smooth_loss = -tf.reduce_sum(reduced_logits, axis=-1)
-            smooth_loss = tf.reduce_sum(smooth_loss)  # sum and squeeze like torch
-            eps_i = smooth_epsilon / reduced_logits.shape[-1]
-
-            loss = (1.0 - smooth_epsilon) * nll_loss + eps_i * smooth_loss
-
-            return loss
-
+        # Matt: As written, this loss is not XLA-compatible, but it's doing some very weird things
+        #       and I don't feel comfortable converting it.
         loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
-            from_logits=from_logits,
-            reduction=tf.keras.losses.Reduction.NONE,
+            from_logits=True,
+            reduction=tf.keras.losses.Reduction.SUM,
         )
 
-        unmasked_loss = loss_fn(labels, y_pred)
-        loss_mask = labels != self.config.generator.pad_token_id
-        nll_loss = tf.reduce_sum(unmasked_loss * loss_mask)
+        if from_logits is False:  # convert to logits
+            eps = 1e-9
+            y_pred = tf.clip_by_value(y_pred, clip_value_min=eps, clip_value_max=1 - eps)
+            y_pred = tf.math.log(y_pred)
+
+        logits = y_pred
+        melted_labels = tf.reshape(labels, (-1,))
+        active_loss = tf.not_equal(melted_labels, self.config.generator.pad_token_id)
+
+        reduced_logits = tf.boolean_mask(tf.reshape(logits, (-1, logits.shape[2])), active_loss)
+        labels = tf.boolean_mask(melted_labels, active_loss)
+        nll_loss = loss_fn(labels, reduced_logits)
 
-        # Matt: This makes no sense to me, but I'm just copying the old loss in XLA-compatible form
-        smooth_loss = -tf.reduce_sum(y_pred * tf.expand_dims(labels, -1), axis=-1)
-        smooth_loss = tf.reduce_sum(smooth_loss)
-        eps_i = smooth_epsilon / y_pred.shape[-1]
+        smooth_loss = -tf.reduce_sum(reduced_logits, axis=-1)
+        smooth_loss = tf.reduce_sum(smooth_loss)  # sum and squeeze like torch
+        eps_i = smooth_epsilon / reduced_logits.shape[-1]
 
         loss = (1.0 - smooth_epsilon) * nll_loss + eps_i * smooth_loss
 

tests/models/xlnet/test_modeling_tf_xlnet.py

@@ -417,12 +417,12 @@ class TFXLNetModelTest(TFModelTesterMixin, unittest.TestCase):
                 input_ids = prepared_for_class.pop(input_name)
 
                 loss = model(input_ids, **prepared_for_class)[0]
-                self.assertEqual(loss.shape.as_list(), expected_loss_size)
+                self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
 
                 # Test that model correctly compute the loss with a dict
                 prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
                 loss = model(prepared_for_class)[0]
-                self.assertEqual(loss.shape.as_list(), expected_loss_size)
+                self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
 
                 # Test that model correctly compute the loss with a tuple
                 prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
@@ -453,7 +453,7 @@ class TFXLNetModelTest(TFModelTesterMixin, unittest.TestCase):
                 # Send to model
                 loss = model(tuple_input[:-1])[0]
 
-                self.assertEqual(loss.shape.as_list(), expected_loss_size)
+                self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
 
 
 @require_tf

tests/test_modeling_tf_common.py

@@ -1294,7 +1294,7 @@ class TFModelTesterMixin:
                 model_input = prepared_for_class.pop(input_name)
 
                 loss = model(model_input, **prepared_for_class)[0]
-                self.assertEqual(loss.shape.as_list(), expected_loss_size)
+                self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
 
                 # Test that model correctly compute the loss when we mask some positions
                 prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
@@ -1307,13 +1307,13 @@ class TFModelTesterMixin:
                         labels[0] = -100
                         prepared_for_class["labels"] = tf.convert_to_tensor(labels)
                         loss = model(model_input, **prepared_for_class)[0]
-                        self.assertEqual(loss.shape.as_list(), expected_loss_size)
+                        self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
                         self.assertTrue(not np.any(np.isnan(loss.numpy())))
 
                 # Test that model correctly compute the loss with a dict
                 prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
                 loss = model(prepared_for_class)[0]
-                self.assertEqual(loss.shape.as_list(), expected_loss_size)
+                self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
 
                 # Test that model correctly compute the loss with a tuple
                 prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
@@ -1344,7 +1344,7 @@ class TFModelTesterMixin:
                 # Send to model
                 loss = model(tuple_input[:-1])[0]
 
-                self.assertEqual(loss.shape.as_list(), expected_loss_size)
+                self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
 
     def test_keras_fit(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()