Move test model folders (#17034)

* move test model folders (TODO: fix imports and others)

* fix (potentially partially) imports (in model test modules)

* fix (potentially partially) imports (in tokenization test modules)

* fix (potentially partially) imports (in feature extraction test modules)

* fix import utils.test_modeling_tf_core

* fix path ../fixtures/

* fix imports about generation.test_generation_flax_utils

* fix more imports

* fix fixture path

* fix get_test_dir

* update module_to_test_file

* fix get_tests_dir from wrong transformers.utils

* update config.yml (CircleCI)

* fix style

* remove missing imports

* update new model script

* update check_repo

* update SPECIAL_MODULE_TO_TEST_MAP

* fix style

* add __init__

* update self-scheduled

* fix add_new_model scripts

* check one way to get location back

* python setup.py build install

* fix import in test auto

* update self-scheduled.yml

* update slack notification script

* Add comments about artifact names

* fix for yolos

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

tests/models/roberta/test_modeling_flax_roberta.py

@@ -0,0 +1,160 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+
+from transformers import RobertaConfig, is_flax_available
+from transformers.testing_utils import require_flax, slow
+
+from ...test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_flax_available():
+    from transformers.models.roberta.modeling_flax_roberta import (
+        FlaxRobertaForCausalLM,
+        FlaxRobertaForMaskedLM,
+        FlaxRobertaForMultipleChoice,
+        FlaxRobertaForQuestionAnswering,
+        FlaxRobertaForSequenceClassification,
+        FlaxRobertaForTokenClassification,
+        FlaxRobertaModel,
+    )
+
+
+class FlaxRobertaModelTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_attention_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_choices=4,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_choices = num_choices
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        config = RobertaConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+        return config, input_ids, token_type_ids, attention_mask
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, token_type_ids, attention_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": attention_mask}
+        return config, inputs_dict
+
+    def prepare_config_and_inputs_for_decoder(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, token_type_ids, attention_mask = config_and_inputs
+
+        config.is_decoder = True
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            token_type_ids,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+
+@require_flax
+class FlaxRobertaModelTest(FlaxModelTesterMixin, unittest.TestCase):
+
+    test_head_masking = True
+
+    all_model_classes = (
+        (
+            FlaxRobertaModel,
+            FlaxRobertaForCausalLM,
+            FlaxRobertaForMaskedLM,
+            FlaxRobertaForSequenceClassification,
+            FlaxRobertaForTokenClassification,
+            FlaxRobertaForMultipleChoice,
+            FlaxRobertaForQuestionAnswering,
+        )
+        if is_flax_available()
+        else ()
+    )
+
+    def setUp(self):
+        self.model_tester = FlaxRobertaModelTester(self)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_class_name in self.all_model_classes:
+            model = model_class_name.from_pretrained("roberta-base", from_pt=True)
+            outputs = model(np.ones((1, 1)))
+            self.assertIsNotNone(outputs)

tests/models/roberta/test_modeling_roberta.py

@@ -0,0 +1,557 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+from copy import deepcopy
+
+from transformers import RobertaConfig, is_torch_available
+from transformers.testing_utils import TestCasePlus, require_torch, slow, torch_device
+
+from ...generation.test_generation_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        RobertaForCausalLM,
+        RobertaForMaskedLM,
+        RobertaForMultipleChoice,
+        RobertaForQuestionAnswering,
+        RobertaForSequenceClassification,
+        RobertaForTokenClassification,
+        RobertaModel,
+    )
+    from transformers.models.roberta.modeling_roberta import (
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        RobertaEmbeddings,
+        create_position_ids_from_input_ids,
+    )
+
+ROBERTA_TINY = "sshleifer/tiny-distilroberta-base"
+
+
+class RobertaModelTester:
+    def __init__(
+        self,
+        parent,
+    ):
+        self.parent = parent
+        self.batch_size = 13
+        self.seq_length = 7
+        self.is_training = True
+        self.use_input_mask = True
+        self.use_token_type_ids = True
+        self.use_labels = True
+        self.vocab_size = 99
+        self.hidden_size = 32
+        self.num_hidden_layers = 5
+        self.num_attention_heads = 4
+        self.intermediate_size = 37
+        self.hidden_act = "gelu"
+        self.hidden_dropout_prob = 0.1
+        self.attention_probs_dropout_prob = 0.1
+        self.max_position_embeddings = 512
+        self.type_vocab_size = 16
+        self.type_sequence_label_size = 2
+        self.initializer_range = 0.02
+        self.num_labels = 3
+        self.num_choices = 4
+        self.scope = None
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def get_config(self):
+        return RobertaConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            initializer_range=self.initializer_range,
+        )
+
+    def prepare_config_and_inputs_for_decoder(self):
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = self.prepare_config_and_inputs()
+
+        config.is_decoder = True
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = RobertaModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_model_as_decoder(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.add_cross_attention = True
+        model = RobertaModel(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+        )
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_for_causal_lm(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        model = RobertaForCausalLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = RobertaForCausalLM(config=config).to(torch_device).eval()
+
+        # make sure that ids don't start with pad token
+        mask = input_ids.ne(config.pad_token_id).long()
+        input_ids = input_ids * mask
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+
+        # make sure that ids don't start with pad token
+        mask = next_tokens.ne(config.pad_token_id).long()
+        next_tokens = next_tokens * mask
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
+    def create_and_check_for_masked_lm(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = RobertaForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_token_classification(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = RobertaForTokenClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_multiple_choice(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_choices = self.num_choices
+        model = RobertaForMultipleChoice(config=config)
+        model.to(torch_device)
+        model.eval()
+        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        result = model(
+            multiple_choice_inputs_ids,
+            attention_mask=multiple_choice_input_mask,
+            token_type_ids=multiple_choice_token_type_ids,
+            labels=choice_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
+
+    def create_and_check_for_question_answering(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = RobertaForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            start_positions=sequence_labels,
+            end_positions=sequence_labels,
+        )
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class RobertaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            RobertaForCausalLM,
+            RobertaForMaskedLM,
+            RobertaModel,
+            RobertaForSequenceClassification,
+            RobertaForTokenClassification,
+            RobertaForMultipleChoice,
+            RobertaForQuestionAnswering,
+        )
+        if is_torch_available()
+        else ()
+    )
+    all_generative_model_classes = (RobertaForCausalLM,) if is_torch_available() else ()
+    fx_compatible = True
+
+    def setUp(self):
+        self.model_tester = RobertaModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=RobertaConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_various_embeddings(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for type in ["absolute", "relative_key", "relative_key_query"]:
+            config_and_inputs[0].position_embedding_type = type
+            self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_as_decoder(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)
+
+    def test_model_as_decoder_with_default_input_mask(self):
+        # This regression test was failing with PyTorch < 1.3
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        ) = self.model_tester.prepare_config_and_inputs_for_decoder()
+
+        input_mask = None
+
+        self.model_tester.create_and_check_model_as_decoder(
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    def test_for_causal_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_for_causal_lm(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = RobertaModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    def test_create_position_ids_respects_padding_index(self):
+        """Ensure that the default position ids only assign a sequential . This is a regression
+        test for https://github.com/huggingface/transformers/issues/1761
+
+        The position ids should be masked with the embedding object's padding index. Therefore, the
+        first available non-padding position index is RobertaEmbeddings.padding_idx + 1
+        """
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        model = RobertaEmbeddings(config=config)
+
+        input_ids = torch.as_tensor([[12, 31, 13, model.padding_idx]])
+        expected_positions = torch.as_tensor(
+            [[0 + model.padding_idx + 1, 1 + model.padding_idx + 1, 2 + model.padding_idx + 1, model.padding_idx]]
+        )
+
+        position_ids = create_position_ids_from_input_ids(input_ids, model.padding_idx)
+        self.assertEqual(position_ids.shape, expected_positions.shape)
+        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
+
+    def test_create_position_ids_from_inputs_embeds(self):
+        """Ensure that the default position ids only assign a sequential . This is a regression
+        test for https://github.com/huggingface/transformers/issues/1761
+
+        The position ids should be masked with the embedding object's padding index. Therefore, the
+        first available non-padding position index is RobertaEmbeddings.padding_idx + 1
+        """
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        embeddings = RobertaEmbeddings(config=config)
+
+        inputs_embeds = torch.empty(2, 4, 30)
+        expected_single_positions = [
+            0 + embeddings.padding_idx + 1,
+            1 + embeddings.padding_idx + 1,
+            2 + embeddings.padding_idx + 1,
+            3 + embeddings.padding_idx + 1,
+        ]
+        expected_positions = torch.as_tensor([expected_single_positions, expected_single_positions])
+        position_ids = embeddings.create_position_ids_from_inputs_embeds(inputs_embeds)
+        self.assertEqual(position_ids.shape, expected_positions.shape)
+        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
+
+
+@require_torch
+class RobertaModelIntegrationTest(TestCasePlus):
+    @slow
+    def test_inference_masked_lm(self):
+        model = RobertaForMaskedLM.from_pretrained("roberta-base")
+
+        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+        with torch.no_grad():
+            output = model(input_ids)[0]
+        expected_shape = torch.Size((1, 11, 50265))
+        self.assertEqual(output.shape, expected_shape)
+        # compare the actual values for a slice.
+        expected_slice = torch.tensor(
+            [[[33.8802, -4.3103, 22.7761], [4.6539, -2.8098, 13.6253], [1.8228, -3.6898, 8.8600]]]
+        )
+
+        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.base')
+        # roberta.eval()
+        # expected_slice = roberta.model.forward(input_ids)[0][:, :3, :3].detach()
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_inference_no_head(self):
+        model = RobertaModel.from_pretrained("roberta-base")
+
+        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+        with torch.no_grad():
+            output = model(input_ids)[0]
+        # compare the actual values for a slice.
+        expected_slice = torch.tensor(
+            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0540, -0.0175], [0.0548, 0.0799, 0.1687]]]
+        )
+
+        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.base')
+        # roberta.eval()
+        # expected_slice = roberta.extract_features(input_ids)[:, :3, :3].detach()
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_inference_classification_head(self):
+        model = RobertaForSequenceClassification.from_pretrained("roberta-large-mnli")
+
+        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+        with torch.no_grad():
+            output = model(input_ids)[0]
+        expected_shape = torch.Size((1, 3))
+        self.assertEqual(output.shape, expected_shape)
+        expected_tensor = torch.tensor([[-0.9469, 0.3913, 0.5118]])
+
+        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.large.mnli')
+        # roberta.eval()
+        # expected_tensor = roberta.predict("mnli", input_ids, return_logits=True).detach()
+
+        self.assertTrue(torch.allclose(output, expected_tensor, atol=1e-4))
+
+    # XXX: this might be a candidate for common tests if we have many of those
+    def test_lm_head_ignore_keys(self):
+        keys_to_ignore_on_save_tied = [r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+        keys_to_ignore_on_save_untied = [r"lm_head.decoder.bias"]
+        config = RobertaConfig.from_pretrained(ROBERTA_TINY)
+        config_tied = deepcopy(config)
+        config_tied.tie_word_embeddings = True
+        config_untied = deepcopy(config)
+        config_untied.tie_word_embeddings = False
+        for cls in [RobertaForMaskedLM, RobertaForCausalLM]:
+            model = cls(config_tied)
+            self.assertEqual(model._keys_to_ignore_on_save, keys_to_ignore_on_save_tied, cls)
+
+            # the keys should be different when embeddings aren't tied
+            model = cls(config_untied)
+            self.assertEqual(model._keys_to_ignore_on_save, keys_to_ignore_on_save_untied, cls)
+
+            # test that saving works with updated ignore keys - just testing that it doesn't fail
+            model.save_pretrained(self.get_auto_remove_tmp_dir())

tests/models/roberta/test_modeling_tf_roberta.py

@@ -0,0 +1,687 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+from transformers import RobertaConfig, is_tf_available
+from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_tf_available():
+    import numpy
+    import tensorflow as tf
+
+    from transformers.models.roberta.modeling_tf_roberta import (
+        TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TFRobertaForCausalLM,
+        TFRobertaForMaskedLM,
+        TFRobertaForMultipleChoice,
+        TFRobertaForQuestionAnswering,
+        TFRobertaForSequenceClassification,
+        TFRobertaForTokenClassification,
+        TFRobertaModel,
+    )
+
+
+class TFRobertaModelTester:
+    def __init__(
+        self,
+        parent,
+    ):
+        self.parent = parent
+        self.batch_size = 13
+        self.seq_length = 7
+        self.is_training = True
+        self.use_input_mask = True
+        self.use_token_type_ids = True
+        self.use_labels = True
+        self.vocab_size = 99
+        self.hidden_size = 32
+        self.num_hidden_layers = 5
+        self.num_attention_heads = 4
+        self.intermediate_size = 37
+        self.hidden_act = "gelu"
+        self.hidden_dropout_prob = 0.1
+        self.attention_probs_dropout_prob = 0.1
+        self.max_position_embeddings = 512
+        self.type_vocab_size = 16
+        self.type_sequence_label_size = 2
+        self.initializer_range = 0.02
+        self.num_labels = 3
+        self.num_choices = 4
+        self.scope = None
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = RobertaConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            initializer_range=self.initializer_range,
+        )
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def prepare_config_and_inputs_for_decoder(self):
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = self.prepare_config_and_inputs()
+
+        config.is_decoder = True
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = TFRobertaModel(config=config)
+        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
+        result = model(inputs)
+
+        inputs = [input_ids, input_mask]
+        result = model(inputs)
+
+        result = model(input_ids)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_causal_lm_base_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.is_decoder = True
+
+        model = TFRobertaModel(config=config)
+        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
+        result = model(inputs)
+
+        inputs = [input_ids, input_mask]
+        result = model(inputs)
+
+        result = model(input_ids)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_model_as_decoder(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.add_cross_attention = True
+
+        model = TFRobertaModel(config=config)
+        inputs = {
+            "input_ids": input_ids,
+            "attention_mask": input_mask,
+            "token_type_ids": token_type_ids,
+            "encoder_hidden_states": encoder_hidden_states,
+            "encoder_attention_mask": encoder_attention_mask,
+        }
+        result = model(inputs)
+
+        inputs = [input_ids, input_mask]
+        result = model(inputs, token_type_ids=token_type_ids, encoder_hidden_states=encoder_hidden_states)
+
+        # Also check the case where encoder outputs are not passed
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_causal_lm_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.is_decoder = True
+
+        model = TFRobertaForCausalLM(config=config)
+        inputs = {
+            "input_ids": input_ids,
+            "attention_mask": input_mask,
+            "token_type_ids": token_type_ids,
+        }
+        prediction_scores = model(inputs)["logits"]
+        self.parent.assertListEqual(
+            list(prediction_scores.numpy().shape), [self.batch_size, self.seq_length, self.vocab_size]
+        )
+
+    def create_and_check_causal_lm_model_as_decoder(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.add_cross_attention = True
+
+        model = TFRobertaForCausalLM(config=config)
+        inputs = {
+            "input_ids": input_ids,
+            "attention_mask": input_mask,
+            "token_type_ids": token_type_ids,
+            "encoder_hidden_states": encoder_hidden_states,
+            "encoder_attention_mask": encoder_attention_mask,
+        }
+        result = model(inputs)
+
+        inputs = [input_ids, input_mask]
+        result = model(inputs, token_type_ids=token_type_ids, encoder_hidden_states=encoder_hidden_states)
+
+        prediction_scores = result["logits"]
+        self.parent.assertListEqual(
+            list(prediction_scores.numpy().shape), [self.batch_size, self.seq_length, self.vocab_size]
+        )
+
+    def create_and_check_causal_lm_model_past(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        config.is_decoder = True
+
+        model = TFRobertaForCausalLM(config=config)
+
+        # special to `RobertaEmbeddings` in `Roberta`:
+        #   - its `padding_idx` and its effect on `position_ids`
+        #     (TFRobertaEmbeddings.create_position_ids_from_input_ids)
+        #   - `1` here is `TFRobertaEmbeddings.padding_idx`
+        input_ids = tf.where(input_ids == 1, 2, input_ids)
+
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and attn_mask
+        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
+
+        output_from_no_past = model(next_input_ids, output_hidden_states=True).hidden_states[0]
+        output_from_past = model(
+            next_tokens, past_key_values=past_key_values, output_hidden_states=True
+        ).hidden_states[0]
+
+        # select random slice
+        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
+        output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx]
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx]
+
+        # test that outputs are equal for slice
+        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-6)
+
+    def create_and_check_causal_lm_model_past_with_attn_mask(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        config.is_decoder = True
+
+        model = TFRobertaForCausalLM(config=config)
+
+        # special to `RobertaEmbeddings` in `Roberta`:
+        #   - its `padding_idx` and its effect on `position_ids`
+        #     (TFRobertaEmbeddings.create_position_ids_from_input_ids)
+        #   - `1` here is `TFRobertaEmbeddings.padding_idx`
+        # avoid `padding_idx` in the past
+        input_ids = tf.where(input_ids == 1, 2, input_ids)
+
+        # create attention mask
+        half_seq_length = self.seq_length // 2
+        attn_mask_begin = tf.ones((self.batch_size, half_seq_length), dtype=tf.int32)
+        attn_mask_end = tf.zeros((self.batch_size, self.seq_length - half_seq_length), dtype=tf.int32)
+        attn_mask = tf.concat([attn_mask_begin, attn_mask_end], axis=1)
+
+        # first forward pass
+        outputs = model(input_ids, attention_mask=attn_mask, use_cache=True)
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        past_key_values = outputs.past_key_values
+
+        # change a random masked slice from input_ids
+        random_seq_idx_to_change = ids_tensor((1,), half_seq_length).numpy() + 1
+        random_other_next_tokens = ids_tensor((self.batch_size, self.seq_length), config.vocab_size)
+        vector_condition = tf.range(self.seq_length) == (self.seq_length - random_seq_idx_to_change)
+        condition = tf.transpose(
+            tf.broadcast_to(tf.expand_dims(vector_condition, -1), (self.seq_length, self.batch_size))
+        )
+        input_ids = tf.where(condition, random_other_next_tokens, input_ids)
+        # avoid `padding_idx` in the past
+        input_ids = tf.where(input_ids == 1, 2, input_ids)
+
+        # append to next input_ids and
+        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
+        attn_mask = tf.concat(
+            [attn_mask, tf.ones((attn_mask.shape[0], 1), dtype=tf.int32)],
+            axis=1,
+        )
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=attn_mask,
+            output_hidden_states=True,
+        ).hidden_states[0]
+        output_from_past = model(
+            next_tokens, past_key_values=past_key_values, attention_mask=attn_mask, output_hidden_states=True
+        ).hidden_states[0]
+
+        # select random slice
+        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
+        output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx]
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx]
+
+        # test that outputs are equal for slice
+        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-6)
+
+    def create_and_check_causal_lm_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        config.is_decoder = True
+
+        model = TFRobertaForCausalLM(config=config)
+
+        # special to `RobertaEmbeddings` in `Roberta`:
+        #   - its `padding_idx` and its effect on `position_ids`
+        #     (TFRobertaEmbeddings.create_position_ids_from_input_ids)
+        #   - `1` here is `TFRobertaEmbeddings.padding_idx`
+        # avoid `padding_idx` in the past
+        input_ids = tf.where(input_ids == 1, 2, input_ids)
+
+        input_ids = input_ids[:1, :]
+        input_mask = input_mask[:1, :]
+        self.batch_size = 1
+
+        # first forward pass
+        outputs = model(input_ids, attention_mask=input_mask, use_cache=True)
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_attn_mask = ids_tensor((self.batch_size, 3), 2)
+
+        # append to next input_ids and
+        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
+        next_attention_mask = tf.concat([input_mask, next_attn_mask], axis=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            output_hidden_states=True,
+        ).hidden_states[0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        ).hidden_states[0]
+
+        self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
+
+        # select random slice
+        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
+        output_from_past_slice = output_from_past[:, :, random_slice_idx]
+
+        # test that outputs are equal for slice
+        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
+
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.add_cross_attention = True
+
+        model = TFRobertaForCausalLM(config=config)
+
+        # special to `RobertaEmbeddings` in `Roberta`:
+        #   - its `padding_idx` and its effect on `position_ids`
+        #     (TFRobertaEmbeddings.create_position_ids_from_input_ids)
+        #   - `1` here is `TFRobertaEmbeddings.padding_idx`
+        # avoid `padding_idx` in the past
+        input_ids = tf.where(input_ids == 1, 2, input_ids)
+
+        input_ids = input_ids[:1, :]
+        input_mask = input_mask[:1, :]
+        encoder_hidden_states = encoder_hidden_states[:1, :, :]
+        encoder_attention_mask = encoder_attention_mask[:1, :]
+        self.batch_size = 1
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_attn_mask = ids_tensor((self.batch_size, 3), 2)
+
+        # append to next input_ids and
+        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
+        next_attention_mask = tf.concat([input_mask, next_attn_mask], axis=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        ).hidden_states[0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        ).hidden_states[0]
+
+        self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
+
+        # select random slice
+        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
+        output_from_past_slice = output_from_past[:, :, random_slice_idx]
+
+        # test that outputs are equal for slice
+        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
+
+    def create_and_check_for_masked_lm(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = TFRobertaForMaskedLM(config=config)
+        result = model([input_ids, input_mask, token_type_ids])
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_token_classification(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = TFRobertaForTokenClassification(config=config)
+        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
+        result = model(inputs)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_question_answering(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = TFRobertaForQuestionAnswering(config=config)
+        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
+        result = model(inputs)
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def create_and_check_for_multiple_choice(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_choices = self.num_choices
+        model = TFRobertaForMultipleChoice(config=config)
+        multiple_choice_inputs_ids = tf.tile(tf.expand_dims(input_ids, 1), (1, self.num_choices, 1))
+        multiple_choice_input_mask = tf.tile(tf.expand_dims(input_mask, 1), (1, self.num_choices, 1))
+        multiple_choice_token_type_ids = tf.tile(tf.expand_dims(token_type_ids, 1), (1, self.num_choices, 1))
+        inputs = {
+            "input_ids": multiple_choice_inputs_ids,
+            "attention_mask": multiple_choice_input_mask,
+            "token_type_ids": multiple_choice_token_type_ids,
+        }
+        result = model(inputs)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_tf
+class TFRobertaModelTest(TFModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            TFRobertaModel,
+            TFRobertaForCausalLM,
+            TFRobertaForMaskedLM,
+            TFRobertaForSequenceClassification,
+            TFRobertaForTokenClassification,
+            TFRobertaForQuestionAnswering,
+        )
+        if is_tf_available()
+        else ()
+    )
+    test_head_masking = False
+    test_onnx = False
+
+    def setUp(self):
+        self.model_tester = TFRobertaModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=RobertaConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        """Test the base model"""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_causal_lm_base_model(self):
+        """Test the base model of the causal LM model
+
+        is_deocder=True, no cross_attention, no encoder outputs
+        """
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_causal_lm_base_model(*config_and_inputs)
+
+    def test_model_as_decoder(self):
+        """Test the base model as a decoder (of an encoder-decoder architecture)
+
+        is_deocder=True + cross_attention + pass encoder outputs
+        """
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_causal_lm(self):
+        """Test the causal LM model"""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_causal_lm_model(*config_and_inputs)
+
+    def test_causal_lm_model_as_decoder(self):
+        """Test the causal LM model as a decoder"""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_causal_lm_model_as_decoder(*config_and_inputs)
+
+    def test_causal_lm_model_past(self):
+        """Test causal LM model with `past_key_values`"""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_causal_lm_model_past(*config_and_inputs)
+
+    def test_causal_lm_model_past_with_attn_mask(self):
+        """Test the causal LM model with `past_key_values` and `attention_mask`"""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_causal_lm_model_past_with_attn_mask(*config_and_inputs)
+
+    def test_causal_lm_model_past_with_large_inputs(self):
+        """Test the causal LM model with `past_key_values` and a longer decoder sequence length"""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_causal_lm_model_past_large_inputs(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs(self):
+        """Similar to `test_causal_lm_model_past_with_large_inputs` but with cross-attention"""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFRobertaModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+@require_tf
+@require_sentencepiece
+@require_tokenizers
+class TFRobertaModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_masked_lm(self):
+        model = TFRobertaForMaskedLM.from_pretrained("roberta-base")
+
+        input_ids = tf.constant([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+        output = model(input_ids)[0]
+        expected_shape = [1, 11, 50265]
+        self.assertEqual(list(output.numpy().shape), expected_shape)
+        # compare the actual values for a slice.
+        expected_slice = tf.constant(
+            [[[33.8802, -4.3103, 22.7761], [4.6539, -2.8098, 13.6253], [1.8228, -3.6898, 8.8600]]]
+        )
+        self.assertTrue(numpy.allclose(output[:, :3, :3].numpy(), expected_slice.numpy(), atol=1e-4))
+
+    @slow
+    def test_inference_no_head(self):
+        model = TFRobertaModel.from_pretrained("roberta-base")
+
+        input_ids = tf.constant([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+        output = model(input_ids)[0]
+        # compare the actual values for a slice.
+        expected_slice = tf.constant(
+            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0540, -0.0175], [0.0548, 0.0799, 0.1687]]]
+        )
+        self.assertTrue(numpy.allclose(output[:, :3, :3].numpy(), expected_slice.numpy(), atol=1e-4))
+
+    @slow
+    def test_inference_classification_head(self):
+        model = TFRobertaForSequenceClassification.from_pretrained("roberta-large-mnli")
+
+        input_ids = tf.constant([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+        output = model(input_ids)[0]
+        expected_shape = [1, 3]
+        self.assertEqual(list(output.numpy().shape), expected_shape)
+        expected_tensor = tf.constant([[-0.9469, 0.3913, 0.5118]])
+        self.assertTrue(numpy.allclose(output.numpy(), expected_tensor.numpy(), atol=1e-4))

tests/models/roberta/test_tokenization_roberta.py

@@ -0,0 +1,303 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import itertools
+import json
+import os
+import unittest
+
+from transformers import AddedToken, RobertaTokenizer, RobertaTokenizerFast
+from transformers.models.roberta.tokenization_roberta import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_tokenizers, slow
+
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+@require_tokenizers
+class RobertaTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+    tokenizer_class = RobertaTokenizer
+    rust_tokenizer_class = RobertaTokenizerFast
+    test_rust_tokenizer = True
+    from_pretrained_kwargs = {"cls_token": "<s>"}
+
+    def setUp(self):
+        super().setUp()
+
+        # Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt
+        vocab = [
+            "l",
+            "o",
+            "w",
+            "e",
+            "r",
+            "s",
+            "t",
+            "i",
+            "d",
+            "n",
+            "\u0120",
+            "\u0120l",
+            "\u0120n",
+            "\u0120lo",
+            "\u0120low",
+            "er",
+            "\u0120lowest",
+            "\u0120newer",
+            "\u0120wider",
+            "<unk>",
+        ]
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+        merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""]
+        self.special_tokens_map = {"unk_token": "<unk>"}
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+        with open(self.merges_file, "w", encoding="utf-8") as fp:
+            fp.write("\n".join(merges))
+
+    def get_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_rust_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return RobertaTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "lower newer"
+        output_text = "lower newer"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file, self.merges_file, **self.special_tokens_map)
+        text = "lower newer"
+        bpe_tokens = ["l", "o", "w", "er", "\u0120", "n", "e", "w", "er"]
+        tokens = tokenizer.tokenize(text)  # , add_prefix_space=True)
+        self.assertListEqual(tokens, bpe_tokens)
+
+        input_tokens = tokens + [tokenizer.unk_token]
+        input_bpe_tokens = [0, 1, 2, 15, 10, 9, 3, 2, 15, 19]
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+    def roberta_dict_integration_testing(self):
+        tokenizer = self.get_tokenizer()
+
+        self.assertListEqual(tokenizer.encode("Hello world!", add_special_tokens=False), [0, 31414, 232, 328, 2])
+        self.assertListEqual(
+            tokenizer.encode("Hello world! cécé herlolip 418", add_special_tokens=False),
+            [0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2],
+        )
+
+    @slow
+    def test_sequence_builders(self):
+        tokenizer = self.tokenizer_class.from_pretrained("roberta-base")
+
+        text = tokenizer.encode("sequence builders", add_special_tokens=False)
+        text_2 = tokenizer.encode("multi-sequence build", add_special_tokens=False)
+
+        encoded_text_from_decode = tokenizer.encode(
+            "sequence builders", add_special_tokens=True, add_prefix_space=False
+        )
+        encoded_pair_from_decode = tokenizer.encode(
+            "sequence builders", "multi-sequence build", add_special_tokens=True, add_prefix_space=False
+        )
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        assert encoded_sentence == encoded_text_from_decode
+        assert encoded_pair == encoded_pair_from_decode
+
+    def test_space_encoding(self):
+        tokenizer = self.get_tokenizer()
+
+        sequence = "Encode this sequence."
+        space_encoding = tokenizer.byte_encoder[" ".encode("utf-8")[0]]
+
+        # Testing encoder arguments
+        encoded = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=False)
+        first_char = tokenizer.convert_ids_to_tokens(encoded[0])[0]
+        self.assertNotEqual(first_char, space_encoding)
+
+        encoded = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=True)
+        first_char = tokenizer.convert_ids_to_tokens(encoded[0])[0]
+        self.assertEqual(first_char, space_encoding)
+
+        tokenizer.add_special_tokens({"bos_token": "<s>"})
+        encoded = tokenizer.encode(sequence, add_special_tokens=True)
+        first_char = tokenizer.convert_ids_to_tokens(encoded[1])[0]
+        self.assertNotEqual(first_char, space_encoding)
+
+        # Testing spaces after special tokens
+        mask = "<mask>"
+        tokenizer.add_special_tokens(
+            {"mask_token": AddedToken(mask, lstrip=True, rstrip=False)}
+        )  # mask token has a left space
+        mask_ind = tokenizer.convert_tokens_to_ids(mask)
+
+        sequence = "Encode <mask> sequence"
+        sequence_nospace = "Encode <mask>sequence"
+
+        encoded = tokenizer.encode(sequence)
+        mask_loc = encoded.index(mask_ind)
+        first_char = tokenizer.convert_ids_to_tokens(encoded[mask_loc + 1])[0]
+        self.assertEqual(first_char, space_encoding)
+
+        encoded = tokenizer.encode(sequence_nospace)
+        mask_loc = encoded.index(mask_ind)
+        first_char = tokenizer.convert_ids_to_tokens(encoded[mask_loc + 1])[0]
+        self.assertNotEqual(first_char, space_encoding)
+
+    def test_pretokenized_inputs(self):
+        pass
+
+    def test_embeded_special_tokens(self):
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+                sentence = "A, <mask> AllenNLP sentence."
+                tokens_r = tokenizer_r.encode_plus(sentence, add_special_tokens=True, return_token_type_ids=True)
+                tokens_p = tokenizer_p.encode_plus(sentence, add_special_tokens=True, return_token_type_ids=True)
+
+                # token_type_ids should put 0 everywhere
+                self.assertEqual(sum(tokens_r["token_type_ids"]), sum(tokens_p["token_type_ids"]))
+
+                # attention_mask should put 1 everywhere, so sum over length should be 1
+                self.assertEqual(
+                    sum(tokens_r["attention_mask"]) / len(tokens_r["attention_mask"]),
+                    sum(tokens_p["attention_mask"]) / len(tokens_p["attention_mask"]),
+                )
+
+                tokens_r_str = tokenizer_r.convert_ids_to_tokens(tokens_r["input_ids"])
+                tokens_p_str = tokenizer_p.convert_ids_to_tokens(tokens_p["input_ids"])
+
+                # Rust correctly handles the space before the mask while python doesnt
+                self.assertSequenceEqual(tokens_p["input_ids"], [0, 250, 6, 50264, 3823, 487, 21992, 3645, 4, 2])
+                self.assertSequenceEqual(tokens_r["input_ids"], [0, 250, 6, 50264, 3823, 487, 21992, 3645, 4, 2])
+
+                self.assertSequenceEqual(
+                    tokens_p_str, ["<s>", "A", ",", "<mask>", "ĠAllen", "N", "LP", "Ġsentence", ".", "</s>"]
+                )
+                self.assertSequenceEqual(
+                    tokens_r_str, ["<s>", "A", ",", "<mask>", "ĠAllen", "N", "LP", "Ġsentence", ".", "</s>"]
+                )
+
+    def test_change_add_prefix_space_and_trim_offsets_args(self):
+        for trim_offsets, add_prefix_space in itertools.product([True, False], repeat=2):
+            tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                self.tmpdirname, use_fast=True, add_prefix_space=add_prefix_space, trim_offsets=trim_offsets
+            )
+
+            pre_tokenizer_state = json.loads(tokenizer_r.backend_tokenizer.pre_tokenizer.__getstate__())
+            post_processor_state = json.loads(tokenizer_r.backend_tokenizer.post_processor.__getstate__())
+
+            self.assertEqual(pre_tokenizer_state["add_prefix_space"], add_prefix_space)
+
+            self.assertEqual(post_processor_state["add_prefix_space"], add_prefix_space)
+            self.assertEqual(post_processor_state["trim_offsets"], trim_offsets)
+
+    def test_offsets_mapping_with_different_add_prefix_space_and_trim_space_arguments(self):
+        # Test which aims to verify that the offsets are well adapted to the argument `add_prefix_space` and
+        # `trim_offsets`
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                text_of_1_token = "hello"  # `hello` is a token in the vocabulary of `pretrained_name`
+                text = f"{text_of_1_token} {text_of_1_token}"
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=True
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (len(text_of_1_token) + 1, len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=True
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (len(text_of_1_token) + 1, len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=False
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (len(text_of_1_token), len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=False
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (len(text_of_1_token), len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+                text = f" {text}"
+
+                # tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                #     pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=True
+                # )
+                # encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                # self.assertEqual(encoding.offset_mapping[0], (1, 1 + len(text_of_1_token)))
+                # self.assertEqual(
+                #     encoding.offset_mapping[1],
+                #     (1 + len(text_of_1_token) + 1, 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
+                # )
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=True
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (1, 1 + len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (1 + len(text_of_1_token) + 1, 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=False
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (0, 1 + len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (1 + len(text_of_1_token), 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=False
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (0, 1 + len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (1 + len(text_of_1_token), 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )