Add a test for TF mixed precision

tests/test_modeling_tf_albert.py

@@ -294,6 +294,10 @@ def test_model_common_attributes(self):
                 name = model.get_bias()
                 assert name is None
 
+    def test_mixed_precision(self):
+        # TODO JP: Make ALBERT float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

tests/test_modeling_tf_bart.py

@@ -278,6 +278,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make BART float16 compliant
+        pass
+
 
 def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
     """If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""

tests/test_modeling_tf_blenderbot.py

@@ -214,6 +214,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make Blenderbot float16 compliant
+        pass
+
     def test_resize_token_embeddings(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/test_modeling_tf_blenderbot_small.py

@@ -279,6 +279,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make Blenderbot Small float16 compliant
+        pass
+
 
 def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
     """If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""

tests/test_modeling_tf_common.py

@@ -279,6 +279,20 @@ def test_saved_model_with_attentions_output(self):
                     [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                 )
 
+    def test_mixed_precision(self):
+        tf.keras.mixed_precision.experimental.set_policy("mixed_float16")
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+            model = model_class(config)
+            outputs = model(class_inputs_dict)
+
+            self.assertIsNotNone(outputs)
+
+        tf.keras.mixed_precision.experimental.set_policy("float32")
+
     def test_keras_save_load(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/test_modeling_tf_ctrl.py

@@ -221,6 +221,10 @@ def test_model_common_attributes(self):
                 name = model.get_bias()
                 assert name is None
 
+    def test_mixed_precision(self):
+        # TODO JP: Make CTRL float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_CTRL_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

tests/test_modeling_tf_flaubert.py

@@ -330,12 +330,8 @@ def test_model_from_pretrained(self):
             model = TFFlaubertModel.from_pretrained(model_name)
             self.assertIsNotNone(model)
 
-    def test_saved_model_with_hidden_states_output(self):
-        # Should be uncommented during patrick TF refactor
-        pass
-
-    def test_saved_model_with_attentions_output(self):
-        # Should be uncommented during patrick TF refactor
+    def test_mixed_precision(self):
+        # TODO JP: Make Flaubert float16 compliant
         pass
 
 

tests/test_modeling_tf_funnel.py

@@ -371,6 +371,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make Funnel float16 compliant
+        pass
+
 
 @require_tf
 class TFFunnelBaseModelTest(TFModelTesterMixin, unittest.TestCase):
@@ -401,3 +405,7 @@ def test_for_multiple_choice(self):
     def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
+
+    def test_mixed_precision(self):
+        # TODO JP: Make Funnel float16 compliant
+        pass

tests/test_modeling_tf_gpt2.py

@@ -387,6 +387,10 @@ def test_gpt2_sequence_classification_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_gpt2_for_sequence_classification(*config_and_inputs)
 
+    def test_mixed_precision(self):
+        # TODO JP: Make GPT2 float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

tests/test_modeling_tf_led.py

@@ -357,6 +357,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make LED float16 compliant
+        pass
+
     def test_saved_model_with_attentions_output(self):
         # This test don't pass because of the error:
         # condition [13,8,4,5], then [13,8,4,5], and else [13,8,4,6] must be broadcastable

tests/test_modeling_tf_longformer.py

@@ -340,14 +340,25 @@ def test_for_multiple_choice(self):
 
     @slow
     def test_saved_model_with_attentions_output(self):
-        # longformer has special attentions which are not
-        # compatible in graph mode
+        # This test don't pass because of the error:
+        # condition [13,8,4,5], then [13,8,4,5], and else [13,8,4,6] must be broadcastable
+        # This occurs line 323 in modeling_tf_led.py because the condition line 255
+        # returns a tensor of shape
+        # [batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + 2]
+        # if is_global_attn is True and a tensor of shape
+        # [batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + 1]
+        # This is due to the tf.concat call line 703 that adds one dimension
+        # Need to check with PVP how to properly fix this
         pass
 
     def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make Longformer float16 compliant
+        pass
+
 
 @require_tf
 @require_sentencepiece

tests/test_modeling_tf_lxmert.py

@@ -704,6 +704,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make Lxmert float16 compliant
+        pass
+
     @slow
     def test_saved_model_with_hidden_states_output(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

tests/test_modeling_tf_marian.py

@@ -247,6 +247,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make Marian float16 compliant
+        pass
+
     def test_resize_token_embeddings(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/test_modeling_tf_mbart.py

@@ -218,6 +218,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make MBart float16 compliant
+        pass
+
     def test_resize_token_embeddings(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/test_modeling_tf_mobilebert.py

@@ -309,6 +309,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make MobileBert float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         # for model_name in TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

tests/test_modeling_tf_openai.py

@@ -245,6 +245,10 @@ def test_openai_gpt_sequence_classification_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_openai_gpt_for_sequence_classification(*config_and_inputs)
 
+    def test_mixed_precision(self):
+        # TODO JP: Make OpenAIGPT float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

tests/test_modeling_tf_pegasus.py

@@ -245,6 +245,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make Pegasus float16 compliant
+        pass
+
     def test_resize_token_embeddings(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/test_modeling_tf_t5.py

@@ -306,6 +306,10 @@ def test_saved_model_creation(self):
         # This test is too long (>30sec) and makes fail the CI
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make T5 float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         model = TFT5Model.from_pretrained("t5-small")
@@ -435,6 +439,10 @@ def test_model(self):
     def test_train_pipeline_custom_model(self):
         pass
 
+    def test_mixed_precision(self):
+        # TODO JP: Make T5 float16 compliant
+        pass
+
 
 @require_tf
 @require_sentencepiece

tests/test_modeling_tf_transfo_xl.py

@@ -204,6 +204,10 @@ def test_model_common_attributes(self):
                 name = model.get_bias()
                 assert name is None
 
+    def test_mixed_precision(self):
+        # TODO JP: Make TransfoXL float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

tests/test_modeling_tf_xlm.py

@@ -326,6 +326,10 @@ def test_for_multiple_choice(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xlm_for_multiple_choice(*config_and_inputs)
 
+    def test_mixed_precision(self):
+        # TODO JP: Make XLM float16 compliant
+        pass
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_XLM_PRETRAINED_MODEL_ARCHIVE_LIST[:1]: