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/bart/test_modeling_tf_bart.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 
 from transformers import BartConfig, BartTokenizer, is_tf_available
-from transformers.testing_utils import require_tf, slow
+from transformers.testing_utils import require_tf, slow, tooslow
 from transformers.utils import cached_property
 
 from ...test_configuration_common import ConfigTester
@@ -293,8 +293,8 @@ class TFBartModelTest(TFModelTesterMixin, TFCoreModelTesterMixin, unittest.TestC
                                 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/blenderbot/test_modeling_tf_blenderbot.py

@@ -17,7 +17,7 @@
 import unittest
 
 from transformers import BlenderbotConfig, BlenderbotTokenizer, is_tf_available
-from transformers.testing_utils import require_tf, require_tokenizers, slow
+from transformers.testing_utils import require_tf, require_tokenizers, slow, tooslow
 from transformers.utils import cached_property
 
 from ...test_configuration_common import ConfigTester
@@ -213,8 +213,8 @@ class TFBlenderbotModelTest(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/blenderbot_small/test_modeling_tf_blenderbot_small.py

@@ -17,7 +17,7 @@
 import unittest
 
 from transformers import BlenderbotSmallConfig, BlenderbotSmallTokenizer, is_tf_available
-from transformers.testing_utils import require_tf, require_tokenizers, slow
+from transformers.testing_utils import require_tf, require_tokenizers, slow, tooslow
 from transformers.utils import cached_property
 
 from ...test_configuration_common import ConfigTester
@@ -278,8 +278,8 @@ class TFBlenderbotSmallModelTest(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/clip/test_modeling_tf_clip.py

@@ -606,11 +606,21 @@ class TFCLIPModelTest(TFModelTesterMixin, unittest.TestCase):
             model = TFCLIPModel.from_pretrained(model_name)
             self.assertIsNotNone(model)
 
+    @unittest.skip(reason="Currently `saved_model` doesn't work with nested outputs.")
+    @slow
+    def test_saved_model_creation(self):
+        pass
+
     @unittest.skip(reason="Currently `saved_model` doesn't work with nested outputs.")
     @slow
     def test_saved_model_creation_extended(self):
         pass
 
+    @unittest.skip(reason="`saved_model` doesn't work with nested outputs so no preparation happens.")
+    @slow
+    def test_prepare_serving_output(self):
+        pass
+
 
 # We will verify our results on an image of cute cats
 def prepare_img():

tests/models/funnel/test_modeling_tf_funnel.py

@@ -17,7 +17,7 @@
 import unittest
 
 from transformers import FunnelConfig, is_tf_available
-from transformers.testing_utils import require_tf
+from transformers.testing_utils import require_tf, tooslow
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor, random_attention_mask
@@ -371,8 +371,8 @@ class TFFunnelModelTest(TFModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     def test_compile_tf_model(self):
@@ -407,6 +407,6 @@ class TFFunnelBaseModelTest(TFModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass

tests/models/led/test_modeling_tf_led.py

@@ -17,7 +17,7 @@
 import unittest
 
 from transformers import LEDConfig, 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
@@ -365,8 +365,8 @@ class TFLEDModelTest(TFModelTesterMixin, unittest.TestCase):
         # TODO JP: Make LED XLA compliant
         pass
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     def test_generate_with_headmasking(self):

tests/models/longformer/test_modeling_tf_longformer.py

@@ -17,7 +17,7 @@
 import unittest
 
 from transformers import 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 ...test_configuration_common import ConfigTester
 from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor, random_attention_mask
@@ -326,8 +326,8 @@ class TFLongformerModelTest(TFModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
 
+    @tooslow
     def test_saved_model_creation(self):
-        # This test is too long (>30sec) and makes fail the CI
         pass
 
     def test_xla_mode(self):

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