Fix functional TF Whisper and modernize tests (huggingface#24301)

* Revert whisper change and modify the test_compile_tf_model test

* make fixup

* Tweak test slightly

* Add functional model saving to test

* Ensure TF can infer shapes for data2vec

* Add override for efficientformer

* Mark test as slow

src/transformers/models/data2vec/modeling_tf_data2vec_vision.py

@@ -1383,11 +1383,11 @@ def call(
         # only keep certain features, and reshape
         # note that we do +1 as the encoder_hidden_states also includes the initial embeddings
         features = [feature for idx, feature in enumerate(encoder_hidden_states) if idx + 1 in self.config.out_indices]
-        batch_size = shape_list(pixel_values)[0]
         patch_resolution = self.config.image_size // self.config.patch_size
 
         def reshape_features(x):
-            x = tf.reshape(x, (batch_size, patch_resolution, patch_resolution, -1))
+            # We do it this way so TF can always infer the non-batch dims at compile time
+            x = tf.reshape(x, (-1, patch_resolution, patch_resolution, self.config.hidden_size))
             return x
 
         features = [reshape_features(x[:, 1:, :]) for x in features]

src/transformers/models/whisper/modeling_tf_whisper.py

@@ -766,12 +766,11 @@ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, past_key_
         # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
         batch_size, seq_len = input_shape[0], input_shape[1]
 
-        if seq_len > 1:
-            combined_attention_mask = _make_causal_mask(input_shape, past_key_values_length=past_key_values_length)
-        else:
-            combined_attention_mask = _expand_mask(
-                tf.ones((batch_size, seq_len + past_key_values_length)), tgt_len=seq_len
-            )
+        combined_attention_mask = tf.cond(
+            tf.math.greater(seq_len, 1),
+            lambda: _make_causal_mask(input_shape, past_key_values_length=past_key_values_length),
+            lambda: _expand_mask(tf.ones((batch_size, seq_len + past_key_values_length)), tgt_len=seq_len),
+        )
 
         if attention_mask is not None:
             # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]

tests/models/data2vec/test_modeling_tf_data2vec_vision.py

@@ -240,10 +240,6 @@ def test_for_image_segmentation(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_for_image_segmentation(*config_and_inputs)
 
-    @unittest.skip("Test was written for TF 1.x and isn't really relevant here")
-    def test_compile_tf_model(self):
-        pass
-
     def test_attention_outputs(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         config.return_dict = True

tests/models/efficientformer/test_modeling_tf_efficientformer.py

@@ -344,6 +344,24 @@ def test_attention_outputs(self):
                     [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                 )
 
+    def test_compile_tf_model(self):
+        # We use a simplified version of this test for EfficientFormer because it requires training=False
+        # and Keras refuses to let us force that during functional construction
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            # Prepare our model
+            model = model_class(config)
+            # These are maximally general inputs for the model, with multiple None dimensions
+            # Hopefully this will catch any conditionals that fail for flexible shapes
+            functional_inputs = {
+                key: tf.keras.Input(shape=val.shape[1:], dtype=val.dtype, name=key)
+                for key, val in model.input_signature.items()
+                if key in model.dummy_inputs
+            }
+            outputs_dict = model(functional_inputs)
+            self.assertTrue(outputs_dict is not None)
+
 
 # We will verify our results on an image of cute cats
 def prepare_img():

tests/models/funnel/test_modeling_tf_funnel.py

@@ -390,10 +390,6 @@ def test_for_question_answering(self):
     def test_saved_model_creation(self):
         pass
 
-    def test_compile_tf_model(self):
-        # This test fails the CI. TODO Lysandre re-enable it
-        pass
-
 
 @require_tf
 class TFFunnelBaseModelTest(TFModelTesterMixin, unittest.TestCase):

tests/models/lxmert/test_modeling_tf_lxmert.py

@@ -532,55 +532,6 @@ def test_save_load(self):
 
                 self.assert_outputs_same(after_outputs, outputs)
 
-    def test_compile_tf_model(self):
-        optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
-        loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
-        metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
-
-        for model_class in self.all_model_classes:
-            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
-                return_obj_labels="PreTraining" in model_class.__name__
-            )
-
-            input_ids = tf.keras.Input(
-                batch_shape=(self.model_tester.batch_size, self.model_tester.seq_length),
-                name="input_ids",
-                dtype="int32",
-            )
-            visual_feats = tf.keras.Input(
-                batch_shape=(
-                    self.model_tester.batch_size,
-                    self.model_tester.num_visual_features,
-                    self.model_tester.visual_feat_dim,
-                ),
-                name="visual_feats",
-                dtype="int32",
-            )
-            visual_pos = tf.keras.Input(
-                batch_shape=(self.model_tester.batch_size, self.model_tester.num_visual_features, 4),
-                name="visual_pos",
-                dtype="int32",
-            )
-
-            # Prepare our model
-            model = model_class(config)
-
-            # Let's load it from the disk to be sure we can use pretrained weights
-            with tempfile.TemporaryDirectory() as tmpdirname:
-                outputs = model(self._prepare_for_class(inputs_dict, model_class))  # build the model
-                model.save_pretrained(tmpdirname)
-                model = model_class.from_pretrained(tmpdirname)
-
-            outputs_dict = model(input_ids, visual_feats, visual_pos)
-            hidden_states = outputs_dict[0]
-
-            # Add a dense layer on top to test integration with other keras modules
-            outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
-
-            # Compile extended model
-            extended_model = tf.keras.Model(inputs=[input_ids, visual_feats, visual_pos], outputs=[outputs])
-            extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
-
     @tooslow
     def test_saved_model_creation(self):
         pass

tests/models/marian/test_modeling_tf_marian.py

@@ -16,7 +16,6 @@
 
 from __future__ import annotations
 
-import tempfile
 import unittest
 import warnings
 
@@ -209,37 +208,6 @@ def test_decoder_model_past_large_inputs(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
         self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
 
-    def test_compile_tf_model(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
-        loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
-        metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
-
-        model_class = self.all_generative_model_classes[0]
-        input_ids = {
-            "decoder_input_ids": tf.keras.Input(batch_shape=(2, 2000), name="decoder_input_ids", dtype="int32"),
-            "input_ids": tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32"),
-        }
-
-        # Prepare our model
-        model = model_class(config)
-        model(self._prepare_for_class(inputs_dict, model_class))  # Model must be called before saving.
-        # Let's load it from the disk to be sure we can use pre-trained weights
-        with tempfile.TemporaryDirectory() as tmpdirname:
-            model.save_pretrained(tmpdirname)
-            model = model_class.from_pretrained(tmpdirname)
-
-        outputs_dict = model(input_ids)
-        hidden_states = outputs_dict[0]
-
-        # Add a dense layer on top to test integration with other keras modules
-        outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
-
-        # Compile extended model
-        extended_model = tf.keras.Model(inputs=[input_ids], outputs=[outputs])
-        extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
-
     @tooslow
     def test_saved_model_creation(self):
         pass

tests/models/mbart/test_modeling_tf_mbart.py

@@ -15,7 +15,6 @@
 
 from __future__ import annotations
 
-import tempfile
 import unittest
 
 from transformers import AutoTokenizer, MBartConfig, is_tf_available
@@ -118,32 +117,6 @@ def check_decoder_model_past_large_inputs(self, config, inputs_dict):
         output, past_key_values = outputs.to_tuple()
         past_key_values = past_key_values[1]
 
-    def test_compile_tf_model(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
-        loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
-        metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
-        model_class = self.all_generative_model_classes[0]
-        input_ids = {
-            "decoder_input_ids": tf.keras.Input(batch_shape=(2, 2000), name="decoder_input_ids", dtype="int32"),
-            "input_ids": tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32"),
-        }
-        # Prepare our model
-        model = model_class(config)
-        model(self._prepare_for_class(inputs_dict, model_class))  # Model must be called before saving.
-        # Let's load it from the disk to be sure we can use pretrained weights
-        with tempfile.TemporaryDirectory() as tmpdirname:
-            model.save_pretrained(tmpdirname)
-            model = model_class.from_pretrained(tmpdirname)
-        outputs_dict = model(input_ids)
-        hidden_states = outputs_dict[0]
-        # Add a dense layer on top to test integration with other keras modules
-        outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
-        # Compile extended model
-        extended_model = tf.keras.Model(inputs=[input_ids], outputs=[outputs])
-        extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
-
 
 def prepare_mbart_inputs_dict(
     config,

tests/models/mobilevit/test_modeling_tf_mobilevit.py

@@ -199,10 +199,6 @@ def test_model_common_attributes(self):
     def test_attention_outputs(self):
         pass
 
-    @unittest.skip("Test was written for TF 1.x and isn't really relevant here")
-    def test_compile_tf_model(self):
-        pass
-
     def test_forward_signature(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/models/pegasus/test_modeling_tf_pegasus.py

@@ -15,7 +15,6 @@
 
 from __future__ import annotations
 
-import tempfile
 import unittest
 
 from transformers import AutoTokenizer, PegasusConfig, is_tf_available
@@ -207,37 +206,6 @@ def test_decoder_model_past_large_inputs(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
         self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
 
-    def test_compile_tf_model(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
-        loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
-        metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
-
-        model_class = self.all_generative_model_classes[0]
-        input_ids = {
-            "decoder_input_ids": tf.keras.Input(batch_shape=(2, 2000), name="decoder_input_ids", dtype="int32"),
-            "input_ids": tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32"),
-        }
-
-        # Prepare our model
-        model = model_class(config)
-        model(self._prepare_for_class(inputs_dict, model_class))  # Model must be called before saving.
-        # Let's load it from the disk to be sure we can use pretrained weights
-        with tempfile.TemporaryDirectory() as tmpdirname:
-            model.save_pretrained(tmpdirname)
-            model = model_class.from_pretrained(tmpdirname)
-
-        outputs_dict = model(input_ids)
-        hidden_states = outputs_dict[0]
-
-        # Add a dense layer on top to test integration with other keras modules
-        outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
-
-        # Compile extended model
-        extended_model = tf.keras.Model(inputs=[input_ids], outputs=[outputs])
-        extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
-
     @tooslow
     def test_saved_model_creation(self):
         pass

tests/models/segformer/test_modeling_tf_segformer.py

@@ -186,10 +186,6 @@ def test_inputs_embeds(self):
     def test_model_common_attributes(self):
         pass
 
-    @unittest.skip("Test was written for TF 1.x and isn't really relevant here")
-    def test_compile_tf_model(self):
-        pass
-
     def test_forward_signature(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/models/vit_mae/test_modeling_tf_vit_mae.py

@@ -283,52 +283,6 @@ def check_pt_tf_models(self, tf_model, pt_model, tf_inputs_dict):
 
         super().check_pt_tf_models(tf_model, pt_model, tf_inputs_dict)
 
-    # overwrite from common since TFViTMAEForPretraining outputs loss along with
-    # logits and mask indices. loss and mask indices are not suitable for integration
-    # with other keras modules.
-    def test_compile_tf_model(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
-        loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
-        metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
-
-        for model_class in self.all_model_classes:
-            # `pixel_values` implies that the input is an image
-            inputs = tf.keras.Input(
-                batch_shape=(
-                    3,
-                    self.model_tester.num_channels,
-                    self.model_tester.image_size,
-                    self.model_tester.image_size,
-                ),
-                name="pixel_values",
-                dtype="float32",
-            )
-
-            # Prepare our model
-            model = model_class(config)
-            model(self._prepare_for_class(inputs_dict, model_class))  # Model must be called before saving.
-            # Let's load it from the disk to be sure we can use pretrained weights
-            with tempfile.TemporaryDirectory() as tmpdirname:
-                model.save_pretrained(tmpdirname, saved_model=False)
-                model = model_class.from_pretrained(tmpdirname)
-
-            outputs_dict = model(inputs)
-            hidden_states = outputs_dict[0]
-
-            # `TFViTMAEForPreTraining` outputs are not recommended to be used for
-            #  downstream application. This is just to check if the outputs of
-            # `TFViTMAEForPreTraining` can be integrated with other keras modules.
-            if model_class.__name__ == "TFViTMAEForPreTraining":
-                hidden_states = outputs_dict["logits"]
-
-            # Add a dense layer on top to test integration with other keras modules
-            outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
-
-            # Compile extended model
-            extended_model = tf.keras.Model(inputs=[inputs], outputs=[outputs])
-            extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
-
     # overwrite from common since TFViTMAEForPretraining has random masking, we need to fix the noise
     # to generate masks during test
     def test_keras_save_load(self):