Switch `return_dict` to `True` by default. (#8530)

* Use the CI to identify failing tests

* Remove from all examples and tests

* More default switch

* Fixes

* More test fixes

* More fixes

* Last fixes hopefully

* Use the CI to identify failing tests

* Remove from all examples and tests

* More default switch

* Fixes

* More test fixes

* More fixes

* Last fixes hopefully

* Run on the real suite

* Fix slow tests

tests/test_modeling_t5.py

@@ -115,7 +115,6 @@ class T5ModelTester:
             bos_token_id=self.pad_token_id,
             pad_token_id=self.pad_token_id,
             decoder_start_token_id=self.decoder_start_token_id,
-            return_dict=True,
         )
 
         return (

tests/test_modeling_tf_albert.py

@@ -121,7 +121,6 @@ class TFAlbertModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
-            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

tests/test_modeling_tf_bart.py

@@ -182,7 +182,6 @@ class TFBartHeadTests(unittest.TestCase):
             eos_token_id=2,
             pad_token_id=1,
             bos_token_id=0,
-            return_dict=True,
             decoder_start_token_id=2,
         )
         return config, input_ids, batch_size
@@ -206,7 +205,6 @@ class TFBartHeadTests(unittest.TestCase):
             encoder_ffn_dim=32,
             decoder_ffn_dim=32,
             max_position_embeddings=48,
-            return_dict=True,
         )
         lm_model = TFBartForConditionalGeneration(config)
         context = tf.fill((7, 2), 4)
@@ -356,7 +354,7 @@ class FasterTFBartModelIntegrationTests(unittest.TestCase):
             padding="longest",
             truncation=True,
         )
-        features = self.xsum_1_1_model.get_encoder()(**batch, return_dict=True).last_hidden_state
+        features = self.xsum_1_1_model.get_encoder()(**batch).last_hidden_state
         import numpy as np
 
         expected = np.array([[-0.0828, -0.0251, -0.0674], [0.1277, 0.3311, -0.0255], [0.2613, -0.0840, -0.2763]])

tests/test_modeling_tf_bert.py

@@ -120,7 +120,6 @@ class TFBertModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
-            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

tests/test_modeling_tf_camembert.py

@@ -39,7 +39,7 @@ class TFCamembertModelIntegrationTest(unittest.TestCase):
             dtype=tf.int32,
         )  # J'aime le camembert !"
 
-        output = model(input_ids, return_dict=True)["last_hidden_state"]
+        output = model(input_ids)["last_hidden_state"]
         expected_shape = tf.TensorShape((1, 10, 768))
         self.assertEqual(output.shape, expected_shape)
         # compare the actual values for a slice.

tests/test_modeling_tf_common.py

@@ -284,7 +284,7 @@ class TFModelTesterMixin:
         if isinstance(after_outputs, tf.Tensor):
             out_1 = after_outputs.numpy()
         elif isinstance(after_outputs, dict):
-            out_1 = after_outputs[list(after_outputs.keys())[0]]
+            out_1 = after_outputs[list(after_outputs.keys())[0]].numpy()
         else:
             out_1 = after_outputs[0].numpy()
         out_2 = outputs[0].numpy()

tests/test_modeling_tf_ctrl.py

@@ -94,7 +94,6 @@ class TFCTRLModelTester(object):
             n_ctx=self.max_position_embeddings,
             # type_vocab_size=self.type_vocab_size,
             # initializer_range=self.initializer_range,
-            return_dict=True,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)

tests/test_modeling_tf_distilbert.py

@@ -91,7 +91,6 @@ class TFDistilBertModelTester:
             attention_dropout=self.attention_probs_dropout_prob,
             max_position_embeddings=self.max_position_embeddings,
             initializer_range=self.initializer_range,
-            return_dict=True,
         )
 
         return config, input_ids, input_mask, sequence_labels, token_labels, choice_labels

tests/test_modeling_tf_electra.py

@@ -97,7 +97,6 @@ class TFElectraModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
-            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

tests/test_modeling_tf_flaubert.py

@@ -114,7 +114,6 @@ class TFFlaubertModelTester:
             summary_type=self.summary_type,
             use_proj=self.use_proj,
             bos_token_id=self.bos_token_id,
-            return_dict=True,
         )
 
         return (

tests/test_modeling_tf_funnel.py

@@ -137,7 +137,6 @@ class TFFunnelModelTester:
             activation_dropout=self.activation_dropout,
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
-            return_dict=True,
         )
 
         return (

tests/test_modeling_tf_gpt2.py

@@ -104,7 +104,6 @@ class TFGPT2ModelTester:
             # initializer_range=self.initializer_range
             bos_token_id=self.bos_token_id,
             eos_token_id=self.eos_token_id,
-            return_dict=True,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)

tests/test_modeling_tf_longformer.py

@@ -594,7 +594,9 @@ class TFLongformerModelIntegrationTest(unittest.TestCase):
         # 'Hello world! ' repeated 1000 times
         input_ids = tf.convert_to_tensor([[0] + [20920, 232, 328, 1437] * 1000 + [2]], dtype=tf.dtypes.int32)
 
-        loss, prediction_scores = model(input_ids, labels=input_ids)
+        output = model(input_ids, labels=input_ids)
+        loss = output.loss
+        prediction_scores = output.logits
 
         expected_loss = tf.constant(0.0073798)
         expected_prediction_scores_sum = tf.constant(-610476600.0)

tests/test_modeling_tf_lxmert.py

@@ -297,7 +297,6 @@ class TFLxmertModelTester(object):
             matched_label=matched_label,
             ans=ans,
             output_attentions=output_attentions,
-            return_dict=True,
         )
         result = model(
             input_ids,
@@ -352,7 +351,6 @@ class TFLxmertModelTester(object):
             matched_label=matched_label,
             ans=ans,
             output_attentions=not output_attentions,
-            return_dict=True,
         )
 
         self.parent.assertEqual(result.prediction_logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
@@ -695,7 +693,8 @@ class TFLxmertModelTest(TFModelTesterMixin, unittest.TestCase):
                 model = tf.keras.models.load_model(tmpdirname)
                 outputs = model(class_inputs_dict)
 
-                language_hidden_states, vision_hidden_states = outputs[-2], outputs[-1]
+                language_hidden_states = outputs["language_hidden_states"]
+                vision_hidden_states = outputs["vision_hidden_states"]
 
                 self.assertEqual(len(language_hidden_states), self.model_tester.num_hidden_layers["language"] + 1)
                 self.assertEqual(len(vision_hidden_states), self.model_tester.num_hidden_layers["vision"] + 1)
@@ -731,11 +730,9 @@ class TFLxmertModelTest(TFModelTesterMixin, unittest.TestCase):
                 model = tf.keras.models.load_model(tmpdirname)
                 outputs = model(class_inputs_dict)
 
-                language_attentions, vision_attentions, cross_encoder_attentions = (
-                    outputs[-3],
-                    outputs[-2],
-                    outputs[-1],
-                )
+                language_attentions = outputs["language_attentions"]
+                vision_attentions = outputs["vision_attentions"]
+                cross_encoder_attentions = outputs["cross_encoder_attentions"]
 
                 self.assertEqual(len(language_attentions), self.model_tester.num_hidden_layers["language"])
                 self.assertEqual(len(vision_attentions), self.model_tester.num_hidden_layers["vision"])

tests/test_modeling_tf_mobilebert.py

@@ -139,7 +139,6 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
                 type_vocab_size=self.type_vocab_size,
                 initializer_range=self.initializer_range,
                 embedding_size=self.embedding_size,
-                return_dict=True,
             )
 
             return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

tests/test_modeling_tf_openai.py

@@ -99,7 +99,6 @@ class TFOpenAIGPTModelTester:
             n_ctx=self.max_position_embeddings,
             # type_vocab_size=self.type_vocab_size,
             # initializer_range=self.initializer_range,
-            return_dict=True,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)

tests/test_modeling_tf_roberta.py

@@ -97,7 +97,6 @@ class TFRobertaModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
-            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

tests/test_modeling_tf_t5.py

@@ -78,7 +78,6 @@ class TFT5ModelTester:
             bos_token_id=self.pad_token_id,
             pad_token_id=self.pad_token_id,
             decoder_start_token_id=self.pad_token_id,
-            return_dict=True,
         )
 
         return (config, input_ids, input_mask, token_labels)

tests/test_modeling_tf_transfo_xl.py

@@ -77,7 +77,6 @@ class TFTransfoXLModelTester:
             div_val=self.div_val,
             n_layer=self.num_hidden_layers,
             eos_token_id=self.eos_token_id,
-            return_dict=True,
         )
 
         return (config, input_ids_1, input_ids_2, lm_labels)

tests/test_modeling_tf_xlm.py

@@ -114,7 +114,6 @@ class TFXLMModelTester:
             summary_type=self.summary_type,
             use_proj=self.use_proj,
             bos_token_id=self.bos_token_id,
-            return_dict=True,
         )
 
         return (