TF: Finalize `unpack_inputs`-related changes (#16499)

* Add unpack_inputs to remaining models

* removed kwargs to `call()` in TF models

* fix TF T5 tests

src/transformers/models/xlm/modeling_tf_xlm.py

@@ -360,7 +360,6 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         # removed: src_enc=None, src_len=None
 
@@ -707,7 +706,6 @@ class TFXLMModel(TFXLMPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         outputs = self.transformer(
             input_ids=input_ids,
@@ -843,7 +841,6 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         transformer_outputs = self.transformer(
             input_ids=input_ids,
@@ -917,7 +914,6 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel, TFSequenceClassificat
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1025,7 +1021,6 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -1150,7 +1145,6 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1237,7 +1231,6 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, TFQuestionAnsweringL
         start_positions=None,
         end_positions=None,
         training=False,
-        **kwargs,
     ):
         r"""
         start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):

src/transformers/models/xlnet/modeling_tf_xlnet.py

@@ -597,7 +597,6 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
 
         if training and use_mems is None:
@@ -1152,7 +1151,6 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         outputs = self.transformer(
             input_ids=input_ids,
@@ -1262,7 +1260,6 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel, TFCausalLanguageModelingLoss):
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFXLNetLMHeadModelOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1394,7 +1391,6 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel, TFSequenceClassif
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFXLNetForSequenceClassificationOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1501,7 +1497,6 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFXLNetForMultipleChoiceOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1623,7 +1618,6 @@ class TFXLNetForTokenClassification(TFXLNetPreTrainedModel, TFTokenClassificatio
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFXLNetForTokenClassificationOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1711,7 +1705,6 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
         start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFXLNetForQuestionAnsweringSimpleOutput, Tuple[tf.Tensor]]:
         r"""
         start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):

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

@@ -653,7 +653,6 @@ class TF{{cookiecutter.camelcase_modelname}}MainLayer(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
 
         if not self.config.is_decoder:
@@ -949,7 +948,6 @@ class TF{{cookiecutter.camelcase_modelname}}Model(TF{{cookiecutter.camelcase_mod
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
         r"""
         encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1049,7 +1047,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForMaskedLM(TF{{cookiecutter.camelca
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1146,7 +1143,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForCausalLM(TF{{cookiecutter.camelca
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFCausalLMOutputWithCrossAttentions, Tuple[tf.Tensor]]:
         r"""
         encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1289,7 +1285,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForSequenceClassification(TF{{cookie
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -1379,7 +1374,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForMultipleChoice(TF{{cookiecutter.c
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -1506,7 +1500,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForTokenClassification(TF{{cookiecut
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1588,7 +1581,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForQuestionAnswering(TF{{cookiecutte
         start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
         r"""
         start_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -2262,7 +2254,6 @@ class TF{{cookiecutter.camelcase_modelname}}Encoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         """
         Args:
@@ -2421,7 +2412,6 @@ class TF{{cookiecutter.camelcase_modelname}}Decoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         r"""
         Args:
@@ -2876,7 +2866,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForConditionalGeneration(TF{{cookiec
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         """
         Returns:

tests/convbert/test_modeling_tf_convbert.py

@@ -355,7 +355,6 @@ class TFConvBertModelTest(TFModelTesterMixin, unittest.TestCase):
 
         for model_class in self.all_model_classes:
             inputs_dict["output_attentions"] = True
-            inputs_dict["use_cache"] = False
             config.output_hidden_states = False
             model = model_class(config)
             outputs = model(self._prepare_for_class(inputs_dict, model_class))

tests/t5/test_modeling_tf_t5.py

@@ -346,6 +346,11 @@ class TFT5ModelTest(TFModelTesterMixin, unittest.TestCase):
         self.assertEqual(model.get_input_embeddings().weight.shape[0], len(tokenizer))
         self.assertNotEqual(model.get_input_embeddings().weight.shape[0], original_vocab_size)
 
+    # This test is run in `TFT5EncoderOnlyModelTest`, where the main layer has the same inputs as the model
+    @unittest.skip(reason="The inputs of the Main Layer are different.")
+    def test_keras_save_load(self):
+        pass
+
 
 class TFT5EncoderOnlyModelTester:
     def __init__(

tests/test_modeling_tf_common.py

@@ -573,7 +573,12 @@ class TFModelTesterMixin:
             pt_model = pt_model_class(config)
 
             tf_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
-            tf_inputs_dict_maybe_with_labels = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            tf_inputs_dict_maybe_with_labels = self._prepare_for_class(
+                inputs_dict,
+                model_class,
+                # Not all models accept "labels" in the forward pass (yet :) )
+                return_labels=True if "labels" in inspect.signature(model_class.call).parameters.keys() else False,
+            )
 
             # Check we can load pt model in tf and vice-versa with model => model functions
             tf_model = transformers.load_pytorch_model_in_tf2_model(tf_model, pt_model, tf_inputs=tf_inputs_dict)
@@ -722,7 +727,6 @@ class TFModelTesterMixin:
 
         for model_class in self.all_model_classes:
             inputs_dict["output_attentions"] = True
-            inputs_dict["use_cache"] = False
             config.output_hidden_states = False
             model = model_class(config)
             outputs = model(self._prepare_for_class(inputs_dict, model_class))
@@ -944,10 +948,6 @@ class TFModelTesterMixin:
             dict_inputs = self._prepare_for_class(inputs_dict, model_class)
             check_equivalence(model, tuple_inputs, dict_inputs)
 
-            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
-            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
-            check_equivalence(model, tuple_inputs, dict_inputs)
-
             tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
             dict_inputs = self._prepare_for_class(inputs_dict, model_class)
             check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
@@ -956,6 +956,12 @@ class TFModelTesterMixin:
             dict_inputs = self._prepare_for_class(inputs_dict, model_class)
             check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})
 
+            # Not all models accept "labels" in the forward pass (yet :) )
+            if "labels" in inspect.signature(model.call).parameters.keys():
+                tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+                dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+                check_equivalence(model, tuple_inputs, dict_inputs)
+
                 tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                 dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                 check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})