Update serving code to enable `saved_model=True` (#18153)

* Add serving_output and serving methods to some vision models

* Add serving outputs for DeiT

* Don't convert hidden states - differing shapes

* Make saveable

* Fix up

* Make swin saveable

* Add in tests

* Fix funnel tests (can't convert to tensor)

* Fix numpy call

* Tidy up a bit

* Add in hidden states - resnet

* Remove numpy

* Fix failing tests - tensor shape and skipping tests

* Remove duplicated function

* PR comments - formatting and var names

* PR comments
Add suggestions made by Joao Gante:
* Use tf.shape instead of shape_list
* Use @tooslow decorator on tests
* Simplify some of the logic

* PR comments
Address Yih-Dar Sheih comments - making tensor names consistent and make types float

* Types consistent with docs; disable test on swin (slow)

* CI trigger

* Change input_features to float32

* Add serving_output for segformer

* Fixup
Co-authored-by: Amy Roberts <amyeroberts@users.noreply.github.com>

tests/models/lxmert/test_modeling_tf_lxmert.py

@@ -20,7 +20,7 @@ import unittest
 import numpy as np
 
 from transformers import LxmertConfig, is_tf_available
-from transformers.testing_utils import require_tf, slow
+from transformers.testing_utils import require_tf, slow, tooslow
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor, random_attention_mask
@@ -600,8 +600,8 @@ class TFLxmertModelTest(TFModelTesterMixin, unittest.TestCase):
                 name = model.get_bias()
                 assert name is None
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     @slow

tests/models/marian/test_modeling_tf_marian.py

@@ -19,7 +19,7 @@ import unittest
 import warnings
 
 from transformers import AutoTokenizer, MarianConfig, MarianTokenizer, TranslationPipeline, is_tf_available
-from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
+from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow, tooslow
 from transformers.utils import cached_property
 
 from ...test_configuration_common import ConfigTester
@@ -246,8 +246,8 @@ class TFMarianModelTest(TFModelTesterMixin, unittest.TestCase):
                 name = model.get_bias()
                 assert name is None
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     def test_resize_token_embeddings(self):

tests/models/mbart/test_modeling_tf_mbart.py

@@ -17,7 +17,7 @@ import tempfile
 import unittest
 
 from transformers import AutoTokenizer, MBartConfig, is_tf_available
-from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
+from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow, tooslow
 from transformers.utils import cached_property
 
 from ...test_configuration_common import ConfigTester
@@ -281,8 +281,8 @@ class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
                                 models_equal = False
                     self.assertTrue(models_equal)
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
 

tests/models/mobilebert/test_modeling_tf_mobilebert.py

@@ -17,7 +17,7 @@
 import unittest
 
 from transformers import MobileBertConfig, is_tf_available
-from transformers.testing_utils import require_tf, slow
+from transformers.testing_utils import require_tf, slow, tooslow
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor, random_attention_mask
@@ -306,8 +306,8 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
                 name = model.get_bias()
                 assert name is None
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     @slow

tests/models/opt/test_modeling_tf_opt.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 
 from transformers import OPTConfig, is_tf_available
-from transformers.testing_utils import require_sentencepiece, require_tf, slow
+from transformers.testing_utils import require_sentencepiece, require_tf, slow, tooslow
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor
@@ -227,8 +227,8 @@ class TFOPTModelTest(TFModelTesterMixin, unittest.TestCase):
                             models_equal = False
                     self.assertTrue(models_equal)
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
 

tests/models/pegasus/test_modeling_tf_pegasus.py

@@ -17,7 +17,7 @@ import tempfile
 import unittest
 
 from transformers import AutoTokenizer, PegasusConfig, is_tf_available
-from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
+from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow, tooslow
 from transformers.utils import cached_property
 
 from ...test_configuration_common import ConfigTester
@@ -244,8 +244,8 @@ class TFPegasusModelTest(TFModelTesterMixin, unittest.TestCase):
                 name = model.get_bias()
                 assert name is None
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     def test_resize_token_embeddings(self):

tests/models/swin/test_modeling_tf_swin.py

@@ -21,7 +21,7 @@ import unittest
 import numpy as np
 
 from transformers import SwinConfig
-from transformers.testing_utils import require_tf, require_vision, slow, to_2tuple
+from transformers.testing_utils import require_tf, require_vision, slow, to_2tuple, tooslow
 from transformers.utils import cached_property, is_tf_available, is_vision_available
 
 from ...test_configuration_common import ConfigTester
@@ -225,6 +225,10 @@ class TFSwinModelTest(TFModelTesterMixin, unittest.TestCase):
     def test_inputs_embeds(self):
         pass
 
+    @tooslow
+    def test_saved_model_creation(self):
+        pass
+
     def test_model_common_attributes(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/models/t5/test_modeling_tf_t5.py

@@ -16,7 +16,7 @@
 import unittest
 
 from transformers import T5Config, is_tf_available
-from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
+from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow, tooslow
 from transformers.utils import cached_property
 
 from ...test_configuration_common import ConfigTester
@@ -305,8 +305,8 @@ class TFT5ModelTest(TFModelTesterMixin, unittest.TestCase):
                 name = model.get_bias()
                 assert name is None
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     @slow

tests/test_modeling_tf_common.py

@@ -205,6 +205,47 @@ class TFModelTesterMixin:
 
             self.assert_outputs_same(after_outputs, outputs)
 
+    @slow
+    def test_saved_model_creation(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.output_hidden_states = False
+        config.output_attentions = False
+
+        if hasattr(config, "use_cache"):
+            config.use_cache = False
+
+        model_class = self.all_model_classes[0]
+
+        class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+        model = model_class(config)
+
+        model(class_inputs_dict)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            model.save_pretrained(tmpdirname, saved_model=True)
+            saved_model_dir = os.path.join(tmpdirname, "saved_model", "1")
+            self.assertTrue(os.path.exists(saved_model_dir))
+
+    def test_prepare_serving_output(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.output_hidden_states = True
+        config.output_attentions = self.has_attentions
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+            outputs = model(inputs)
+            serving_outputs = model.serving_output(outputs)
+
+            for k, v in serving_outputs.items():
+                # Check that we have one of three possible outputs: None, tuple of tensors or a tensor
+                if isinstance(v, tuple):
+                    self.assertTrue(all(isinstance(elem, tf.Tensor) for elem in v))
+                elif v is not None:
+                    self.assertIsInstance(v, tf.Tensor)
+                else:
+                    self.assertIsNone(v)
+
     def test_forward_signature(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/utils/test_modeling_tf_core.py

@@ -201,27 +201,6 @@ class TFCoreModelTesterMixin:
                 val_loss = history.history["val_loss"][0]
                 self.assertTrue(not isnan(val_loss))
 
-    @slow
-    def test_saved_model_creation(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        config.output_hidden_states = False
-        config.output_attentions = False
-
-        if hasattr(config, "use_cache"):
-            config.use_cache = False
-
-        model_class = self.all_model_classes[0]
-
-        class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
-        model = model_class(config)
-
-        model(class_inputs_dict)
-
-        with tempfile.TemporaryDirectory() as tmpdirname:
-            model.save_pretrained(tmpdirname, saved_model=True)
-            saved_model_dir = os.path.join(tmpdirname, "saved_model", "1")
-            self.assertTrue(os.path.exists(saved_model_dir))
-
     @slow
     def test_saved_model_creation_extended(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()