ProphetNet (#7157)

* add new model prophetnet

prophetnet modified

modify codes as suggested v1

add prophetnet test files

* still bugs, because of changed output formats of encoder and decoder

* move prophetnet into the latest version

* clean integration tests

* clean tokenizers

* add xlm config to init

* correct typo in init

* further refactoring

* continue refactor

* save parallel

* add decoder_attention_mask

* fix use_cache vs. past_key_values

* fix common tests

* change decoder output logits

* fix xlm tests

* make common tests pass

* change model architecture

* add tokenizer tests

* finalize model structure

* no weight mapping

* correct n-gram stream attention mask as discussed with qweizhen

* remove unused import

* fix index.rst

* fix tests

* delete unnecessary code

* add fast integration test

* rename weights

* final weight remapping

* save intermediate

* Descriptions for Prophetnet Config File

* finish all models

* finish new model outputs

* delete unnecessary files

* refactor encoder layer

* add dummy docs

* code quality

* fix tests

* add model pages to doctree

* further refactor

* more refactor, more tests

* finish code refactor and tests

* remove unnecessary files

* further clean up

* add docstring template

* finish tokenizer doc

* finish prophetnet

* fix copies

* fix typos

* fix tf tests

* fix fp16

* fix tf test 2nd try

* fix code quality

* add test for each model

* merge new tests to branch

* Update model_cards/microsoft/prophetnet-large-uncased-cnndm/README.md

Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

* Update model_cards/microsoft/prophetnet-large-uncased-cnndm/README.md

Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

* Update src/transformers/modeling_prophetnet.py

Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

* Update utils/check_repo.py

Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

* apply sams and sylvains comments

* make style

* remove unnecessary code

* Update README.md

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update README.md

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/configuration_prophetnet.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* implement lysandres comments

* correct docs

* fix isort

* fix tokenizers

* fix copies

Co-authored-by: weizhen <weizhen@mail.ustc.edu.cn>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sam Shleifer <sshleifer@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

tests/test_modeling_common.py

@@ -112,6 +112,7 @@ class ModelTesterMixin:
             model.eval()
             with torch.no_grad():
                 outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
             out_2 = outputs[0].cpu().numpy()
             out_2[np.isnan(out_2)] = 0
 
@@ -152,6 +153,7 @@ class ModelTesterMixin:
             with torch.no_grad():
                 first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
                 second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
+
             out_1 = first.cpu().numpy()
             out_2 = second.cpu().numpy()
             out_1 = out_1[~np.isnan(out_1)]
@@ -183,10 +185,12 @@ class ModelTesterMixin:
 
     def test_attention_outputs(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
         seq_len = getattr(self.model_tester, "seq_length", None)
         decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
         encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
-        decoder_key_length = getattr(self.model_tester, "key_length", decoder_seq_length)
+        decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
         encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
         chunk_length = getattr(self.model_tester, "chunk_length", None)
         if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
@@ -211,7 +215,7 @@ class ModelTesterMixin:
             model.eval()
             with torch.no_grad():
                 outputs = model(**self._prepare_for_class(inputs_dict, model_class), return_dict=True)
-            attentions = outputs[-1]
+            attentions = outputs["attentions"] if "attentions" in outputs.keys() else outputs[-1]
             self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
 
             if chunk_length is not None:
@@ -227,8 +231,14 @@ class ModelTesterMixin:
             out_len = len(outputs)
 
             if self.is_encoder_decoder:
-                correct_outlen = 4
-                decoder_attention_idx = 1
+                correct_outlen = (
+                    self.model_tester.base_model_out_len if hasattr(self.model_tester, "base_model_out_len") else 4
+                )
+                decoder_attention_idx = (
+                    self.model_tester.decoder_attention_idx
+                    if hasattr(self.model_tester, "decoder_attention_idx")
+                    else 1
+                )
 
                 # loss is at first position
                 if "labels" in inputs_dict:
@@ -238,6 +248,7 @@ class ModelTesterMixin:
                 if model_class in MODEL_FOR_QUESTION_ANSWERING_MAPPING.values():
                     correct_outlen += 1  # start_logits and end_logits instead of only 1 output
                     decoder_attention_idx += 1
+
                 self.assertEqual(out_len, correct_outlen)
 
                 decoder_attentions = outputs[decoder_attention_idx]
@@ -256,9 +267,16 @@ class ModelTesterMixin:
             model.eval()
             with torch.no_grad():
                 outputs = model(**self._prepare_for_class(inputs_dict, model_class))
-            self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
 
-            self_attentions = outputs[-1]
+            if hasattr(self.model_tester, "num_hidden_states_types"):
+                added_hidden_states = self.model_tester.num_hidden_states_types
+            elif self.is_encoder_decoder:
+                added_hidden_states = 2
+            else:
+                added_hidden_states = 1
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs["attentions"] if "attentions" in outputs else outputs[-1]
             self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
             if chunk_length is not None:
                 self.assertListEqual(
@@ -571,8 +589,8 @@ class ModelTesterMixin:
             model.eval()
 
             with torch.no_grad():
-                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
-            hidden_states = outputs[-1]
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class), return_dict=True)
+            hidden_states = outputs["hidden_states"] if "hidden_states" in outputs else outputs[-1]
 
             expected_num_layers = getattr(
                 self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
@@ -822,6 +840,7 @@ class ModelTesterMixin:
             model.eval()
 
             inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
+
             if not self.is_encoder_decoder:
                 input_ids = inputs["input_ids"]
                 del inputs["input_ids"]
@@ -839,7 +858,7 @@ class ModelTesterMixin:
                 inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
 
             with torch.no_grad():
-                model(**inputs)
+                model(**inputs)[0]
 
     def test_lm_head_model_random_no_beam_search_generate(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

tests/test_modeling_encoder_decoder.py

@@ -24,6 +24,7 @@ from .test_modeling_bert import BertModelTester
 from .test_modeling_bert_generation import BertGenerationEncoderTester
 from .test_modeling_common import ids_tensor
 from .test_modeling_gpt2 import GPT2ModelTester
+from .test_modeling_prophetnet import ProphetNetStandaloneDecoderModelTester
 from .test_modeling_roberta import RobertaModelTester
 
 
@@ -41,9 +42,11 @@ if is_torch_available():
         EncoderDecoderConfig,
         EncoderDecoderModel,
         GPT2LMHeadModel,
+        ProphetNetForCausalLM,
         RobertaForCausalLM,
         RobertaModel,
     )
+    from transformers.modeling_outputs import BaseModelOutput
 
 
 @require_torch
@@ -82,10 +85,15 @@ class EncoderDecoderMixin:
             decoder_input_ids=decoder_input_ids,
             attention_mask=attention_mask,
             decoder_attention_mask=decoder_attention_mask,
+            return_dict=True,
         )
 
-        self.assertEqual(outputs_encoder_decoder[0].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,)))
-        self.assertEqual(outputs_encoder_decoder[1].shape, (input_ids.shape + (config.hidden_size,)))
+        self.assertEqual(
+            outputs_encoder_decoder["logits"].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,))
+        )
+        self.assertEqual(
+            outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
+        )
 
     def check_encoder_decoder_model(
         self,
@@ -109,20 +117,30 @@ class EncoderDecoderMixin:
             decoder_input_ids=decoder_input_ids,
             attention_mask=attention_mask,
             decoder_attention_mask=decoder_attention_mask,
+            return_dict=True,
+        )
+        self.assertEqual(
+            outputs_encoder_decoder["logits"].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,))
+        )
+        self.assertEqual(
+            outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
         )
 
-        self.assertEqual(outputs_encoder_decoder[0].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,)))
-        self.assertEqual(outputs_encoder_decoder[1].shape, (input_ids.shape + (config.hidden_size,)))
-        encoder_outputs = (encoder_hidden_states,)
+        encoder_outputs = BaseModelOutput(last_hidden_state=encoder_hidden_states)
         outputs_encoder_decoder = enc_dec_model(
             encoder_outputs=encoder_outputs,
             decoder_input_ids=decoder_input_ids,
             attention_mask=attention_mask,
             decoder_attention_mask=decoder_attention_mask,
+            return_dict=True,
         )
 
-        self.assertEqual(outputs_encoder_decoder[0].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,)))
-        self.assertEqual(outputs_encoder_decoder[1].shape, (input_ids.shape + (config.hidden_size,)))
+        self.assertEqual(
+            outputs_encoder_decoder["logits"].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,))
+        )
+        self.assertEqual(
+            outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
+        )
 
     def check_encoder_decoder_model_from_pretrained(
         self,
@@ -145,10 +163,15 @@ class EncoderDecoderMixin:
             decoder_input_ids=decoder_input_ids,
             attention_mask=attention_mask,
             decoder_attention_mask=decoder_attention_mask,
+            return_dict=True,
         )
 
-        self.assertEqual(outputs_encoder_decoder[0].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,)))
-        self.assertEqual(outputs_encoder_decoder[1].shape, (input_ids.shape + (config.hidden_size,)))
+        self.assertEqual(
+            outputs_encoder_decoder["logits"].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,))
+        )
+        self.assertEqual(
+            outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
+        )
 
     def check_save_and_load(
         self,
@@ -255,14 +278,19 @@ class EncoderDecoderMixin:
             attention_mask=attention_mask,
             decoder_attention_mask=decoder_attention_mask,
             labels=labels,
+            return_dict=True,
         )
 
-        mlm_loss = outputs_encoder_decoder[0]
+        loss = outputs_encoder_decoder["loss"]
         # check that backprop works
-        mlm_loss.backward()
+        loss.backward()
 
-        self.assertEqual(outputs_encoder_decoder[1].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,)))
-        self.assertEqual(outputs_encoder_decoder[2].shape, (input_ids.shape + (config.hidden_size,)))
+        self.assertEqual(
+            outputs_encoder_decoder["logits"].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,))
+        )
+        self.assertEqual(
+            outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
+        )
 
     def check_encoder_decoder_model_generate(self, input_ids, config, decoder_config, **kwargs):
         encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
@@ -425,6 +453,7 @@ class EncoderDecoderMixin:
                 self.assertLessEqual(max_diff, 1e-5)
 
 
+@require_torch
 class BertEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
     def get_pretrained_model(self):
         return EncoderDecoderModel.from_encoder_decoder_pretrained("bert-base-cased", "bert-base-cased")
@@ -493,6 +522,7 @@ class BertEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
         self.assertEqual(summary, EXPECTED_SUMMARY)
 
 
+@require_torch
 class BertGenerationEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
     def get_pretrained_model(self):
         return EncoderDecoderModel.from_encoder_decoder_pretrained(
@@ -554,6 +584,7 @@ class BertGenerationEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCa
         self.assertEqual(summary, EXPECTED_SUMMARY)
 
 
+@require_torch
 class RoBertaEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
     def get_encoder_decoder_model(self, config, decoder_config):
         encoder_model = RobertaModel(config)
@@ -606,6 +637,7 @@ class RoBertaEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
         return EncoderDecoderModel.from_encoder_decoder_pretrained("roberta-base", "roberta-base")
 
 
+@require_torch
 class GPT2EncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
     def get_encoder_decoder_model(self, config, decoder_config):
         encoder_model = BertModel(config)
@@ -663,3 +695,59 @@ class GPT2EncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
 
     def test_encoder_decoder_model_shared_weights(self):
         pass
+
+
+@require_torch
+class ProphetNetEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
+    def get_encoder_decoder_model(self, config, decoder_config):
+        encoder_model = BertModel(config)
+        decoder_model = ProphetNetForCausalLM(decoder_config)
+        return encoder_model, decoder_model
+
+    def prepare_config_and_inputs(self):
+        model_tester_encoder = BertModelTester(self, batch_size=13)
+        model_tester_decoder = ProphetNetStandaloneDecoderModelTester(
+            self, batch_size=13, hidden_size=32, max_position_embeddings=512
+        )
+        encoder_config_and_inputs = model_tester_encoder.prepare_config_and_inputs()
+        decoder_config_and_inputs = model_tester_decoder.prepare_config_and_inputs_for_decoder()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = encoder_config_and_inputs
+        (
+            decoder_config,
+            decoder_input_ids,
+            decoder_attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            lm_labels,
+        ) = decoder_config_and_inputs
+
+        # make sure that cross attention layers are added
+        decoder_config.add_cross_attention = True
+        #  disable cache for now
+        decoder_config.use_cache = False
+        return {
+            "config": config,
+            "input_ids": input_ids,
+            "attention_mask": input_mask,
+            "decoder_config": decoder_config,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+            "encoder_hidden_states": encoder_hidden_states,
+            "labels": lm_labels,
+        }
+
+    def get_pretrained_model(self):
+        return EncoderDecoderModel.from_encoder_decoder_pretrained(
+            "bert-large-uncased", "patrickvonplaten/prophetnet-decoder-clm-large-uncased"
+        )
+
+    def test_encoder_decoder_model_shared_weights(self):
+        pass

tests/test_modeling_t5.py

@@ -43,7 +43,6 @@ class T5ModelTester:
         encoder_seq_length=7,
         decoder_seq_length=9,
         # For common tests
-        seq_length=7,
         is_training=True,
         use_attention_mask=True,
         use_labels=True,

tests/test_modeling_xlm_prophetnet.py

@@ -0,0 +1,141 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team, The Microsoft Research team.
+#
+# 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 is_torch_available
+from transformers.testing_utils import slow, torch_device
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import XLMProphetNetForConditionalGeneration, XLMProphetNetTokenizer
+
+
+class XLMProphetNetModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_pretrained_checkpoint_hidden_states(self):
+        model = XLMProphetNetForConditionalGeneration.from_pretrained("microsoft/xprophetnet-large-wiki100-cased")
+        model.to(torch_device)
+
+        # encoder-decoder outputs
+        encoder_ids = torch.tensor([[17, 96208, 103471, 2]]).to(torch_device)
+        decoder_prev_ids = torch.tensor(
+            [[2, 250, 9953, 34, 69489, 1620, 32, 118424, 624, 210, 105, 2913, 1032, 351]]
+        ).to(torch_device)
+        output = model(
+            input_ids=encoder_ids, attention_mask=None, encoder_outputs=None, decoder_input_ids=decoder_prev_ids
+        )
+        output_predited_logis = output[0]
+        expected_shape = torch.Size((1, 14, 250012))
+        self.assertEqual(output_predited_logis.shape, expected_shape)
+        expected_slice = torch.tensor(
+            [[[-6.6042, -8.3838, 12.4717], [-6.4426, -8.1994, 12.4542], [-6.0851, -7.8209, 12.9493]]]
+        ).to(torch_device)
+        self.assertTrue(torch.allclose(output_predited_logis[:, :3, :3], expected_slice, atol=1e-4))
+
+        # encoder outputs
+        encoder_outputs = model.prophetnet.encoder(encoder_ids)[0]
+        expected_encoder_outputs_slice = torch.tensor(
+            [[[-1.4260, -0.7628, 0.8453], [-1.4719, -0.1391, 0.7807], [-1.7678, 0.0114, 0.4646]]]
+        ).to(torch_device)
+        expected_shape_encoder = torch.Size((1, 4, 1024))
+        self.assertEqual(encoder_outputs.shape, expected_shape_encoder)
+        self.assertTrue(torch.allclose(encoder_outputs[:, :3, :3], expected_encoder_outputs_slice, atol=1e-4))
+
+        # decoder outputs
+        decoder_outputs = model.prophetnet.decoder(
+            decoder_prev_ids,
+            encoder_hidden_states=encoder_outputs,
+        )
+        predicting_streams = decoder_outputs[1].view(1, model.config.ngram, 14, -1)
+        predicting_streams_logits = model.lm_head(predicting_streams)
+        next_first_stream_logits = predicting_streams_logits[:, 0]
+        self.assertTrue(torch.allclose(next_first_stream_logits[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_ntg_hidden_states(self):
+        model = XLMProphetNetForConditionalGeneration.from_pretrained(
+            "microsoft/xprophetnet-large-wiki100-cased-xglue-ntg"
+        )
+        model.to(torch_device)
+
+        encoder_ids = torch.tensor([[17, 96208, 103471, 2]]).to(torch_device)
+        decoder_prev_ids = torch.tensor(
+            [[2, 250, 9953, 34, 69489, 1620, 32, 118424, 624, 210, 105, 2913, 1032, 351]]
+        ).to(torch_device)
+        output = model(
+            input_ids=encoder_ids, attention_mask=None, encoder_outputs=None, decoder_input_ids=decoder_prev_ids
+        )
+        output_predited_logis = output[0]
+        expected_shape = torch.Size((1, 14, 250012))
+        self.assertEqual(output_predited_logis.shape, expected_shape)
+        # compare the actual values for a slice.
+        expected_slice = torch.tensor(
+            [[[-8.8815, -9.2996, -4.4506], [-6.7202, -7.8944, -0.9402], [-8.6890, -7.4528, -1.9437]]]
+        ).to(torch_device)
+
+        self.assertTrue(torch.allclose(output_predited_logis[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_xprophetnet_ntg_inference(self):
+        model = XLMProphetNetForConditionalGeneration.from_pretrained(
+            "microsoft/xprophetnet-large-wiki100-cased-xglue-ntg"
+        )
+        model.to(torch_device)
+        model.config.max_length = 512
+
+        tokenizer = XLMProphetNetTokenizer.from_pretrained("microsoft/xprophetnet-large-wiki100-cased-xglue-ntg")
+
+        EN_SENTENCE = "Microsoft Corporation intends to officially end free support for the Windows 7 operating system after January 14, 2020, according to the official portal of the organization. From that day, users of this system will not be able to receive security updates, which could make their computers vulnerable to cyber attacks."
+        RU_SENTENCE = "орпорация Microsoft намерена официально прекратить бесплатную поддержку операционной системы Windows 7 после 14 января 2020 года, сообщается на официальном портале организации . С указанного дня пользователи этой системы не смогут получать обновления безопасности, из-за чего их компьютеры могут стать уязвимыми к кибератакам."
+        ZH_SENTENCE = (
+            "根据该组织的官方门户网站，微软公司打算在2020年1月14日之后正式终止对Windows 7操作系统的免费支持。从那时起，该系统的用户将无法接收安全更新，这可能会使他们的计算机容易受到网络攻击。"
+        )
+
+        input_ids = tokenizer(
+            [EN_SENTENCE, RU_SENTENCE, ZH_SENTENCE], padding=True, max_length=255, return_tensors="pt"
+        ).input_ids
+        input_ids = input_ids.to(torch_device)
+
+        summary_ids = model.generate(
+            input_ids, num_beams=10, length_penalty=1.0, no_repeat_ngram_size=3, early_stopping=True
+        )
+        generated_titles = [tokenizer.decode(g, skip_special_tokens=True) for g in summary_ids]
+        EXPECTED_TITLE_EN = "Microsoft to end Windows 7 free support after January 14, 2020"
+        EXPECTED_TITLE_RU = "Microsoft намерена прекратить бесплатную поддержку Windows 7 после 14 января 2020 года"
+        EXPECTED_TITLE_ZH = "微软打算终止对Windows 7操作系统的免费支持"
+        self.assertListEqual(
+            [EXPECTED_TITLE_EN, EXPECTED_TITLE_RU, EXPECTED_TITLE_ZH],
+            generated_titles,
+        )
+
+        summary_ids_beam1 = model.generate(
+            input_ids, num_beams=1, length_penalty=1.0, no_repeat_ngram_size=3, early_stopping=True
+        )
+        generated_titles_beam1_tok = [
+            tokenizer.convert_ids_to_tokens(g, skip_special_tokens=True) for g in summary_ids_beam1
+        ]
+        EXPECTED_TITLE_EN_BEAM1_TOK = "▁Microsoft ▁to ▁end ▁free ▁support ▁for ▁Windows ▁7".split(" ")
+        EXPECTED_TITLE_RU_BEAM1_TOK = "▁Microsoft ▁намерен а ▁прекрати ть ▁бес плат ную ▁поддержку ▁Windows ▁7 ▁после ▁14 ▁января ▁2020 ▁года".split(
+            " "
+        )
+        EXPECTED_TITLE_ZH_BEAM1_TOK = "微软 公司 打算 终止 对 Windows ▁7 操作 系统的 免费 支持".split(" ")
+        self.assertListEqual(
+            [EXPECTED_TITLE_EN_BEAM1_TOK, EXPECTED_TITLE_RU_BEAM1_TOK, EXPECTED_TITLE_ZH_BEAM1_TOK],
+            generated_titles_beam1_tok,
+        )

tests/test_tokenization_prophetnet.py

@@ -0,0 +1,191 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team, The Microsoft Research team.
+#
+# 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 os
+import unittest
+
+from transformers.testing_utils import slow
+from transformers.tokenization_bert import (
+    BasicTokenizer,
+    WordpieceTokenizer,
+    _is_control,
+    _is_punctuation,
+    _is_whitespace,
+)
+from transformers.tokenization_prophetnet import VOCAB_FILES_NAMES, ProphetNetTokenizer
+
+from .test_tokenization_common import TokenizerTesterMixin
+
+
+class ProphetNetTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+
+    tokenizer_class = ProphetNetTokenizer
+    test_rust_tokenizer = False
+
+    def setUp(self):
+        super().setUp()
+
+        vocab_tokens = [
+            "[UNK]",
+            "[CLS]",
+            "[SEP]",
+            "[PAD]",
+            "[MASK]",
+            "want",
+            "##want",
+            "##ed",
+            "wa",
+            "un",
+            "runn",
+            "##ing",
+            ",",
+            "low",
+            "lowest",
+        ]
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "UNwant\u00E9d,running"
+        output_text = "unwanted, running"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file)
+
+        tokens = tokenizer.tokenize("UNwant\u00E9d,running")
+        self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [9, 6, 7, 12, 10, 11])
+
+    def test_chinese(self):
+        tokenizer = BasicTokenizer()
+
+        self.assertListEqual(tokenizer.tokenize("ah\u535A\u63A8zz"), ["ah", "\u535A", "\u63A8", "zz"])
+
+    def test_basic_tokenizer_lower(self):
+        tokenizer = BasicTokenizer(do_lower_case=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU?  "), ["hello", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"])
+
+    def test_basic_tokenizer_lower_strip_accents_false(self):
+        tokenizer = BasicTokenizer(do_lower_case=True, strip_accents=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hällo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["h\u00E9llo"])
+
+    def test_basic_tokenizer_lower_strip_accents_true(self):
+        tokenizer = BasicTokenizer(do_lower_case=True, strip_accents=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hallo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"])
+
+    def test_basic_tokenizer_lower_strip_accents_default(self):
+        tokenizer = BasicTokenizer(do_lower_case=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hallo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"])
+
+    def test_basic_tokenizer_no_lower(self):
+        tokenizer = BasicTokenizer(do_lower_case=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU?  "), ["HeLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    def test_basic_tokenizer_no_lower_strip_accents_false(self):
+        tokenizer = BasicTokenizer(do_lower_case=False, strip_accents=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["HäLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    def test_basic_tokenizer_no_lower_strip_accents_true(self):
+        tokenizer = BasicTokenizer(do_lower_case=False, strip_accents=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["HaLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    def test_basic_tokenizer_respects_never_split_tokens(self):
+        tokenizer = BasicTokenizer(do_lower_case=False, never_split=["[UNK]"])
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU? [UNK]"), ["HeLLo", "!", "how", "Are", "yoU", "?", "[UNK]"]
+        )
+
+    def test_wordpiece_tokenizer(self):
+        vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn", "##ing"]
+
+        vocab = {}
+        for (i, token) in enumerate(vocab_tokens):
+            vocab[token] = i
+        tokenizer = WordpieceTokenizer(vocab=vocab, unk_token="[UNK]")
+
+        self.assertListEqual(tokenizer.tokenize(""), [])
+
+        self.assertListEqual(tokenizer.tokenize("unwanted running"), ["un", "##want", "##ed", "runn", "##ing"])
+
+        self.assertListEqual(tokenizer.tokenize("unwantedX running"), ["[UNK]", "runn", "##ing"])
+
+    def test_is_whitespace(self):
+        self.assertTrue(_is_whitespace(" "))
+        self.assertTrue(_is_whitespace("\t"))
+        self.assertTrue(_is_whitespace("\r"))
+        self.assertTrue(_is_whitespace("\n"))
+        self.assertTrue(_is_whitespace("\u00A0"))
+
+        self.assertFalse(_is_whitespace("A"))
+        self.assertFalse(_is_whitespace("-"))
+
+    def test_is_control(self):
+        self.assertTrue(_is_control("\u0005"))
+
+        self.assertFalse(_is_control("A"))
+        self.assertFalse(_is_control(" "))
+        self.assertFalse(_is_control("\t"))
+        self.assertFalse(_is_control("\r"))
+
+    def test_is_punctuation(self):
+        self.assertTrue(_is_punctuation("-"))
+        self.assertTrue(_is_punctuation("$"))
+        self.assertTrue(_is_punctuation("`"))
+        self.assertTrue(_is_punctuation("."))
+
+        self.assertFalse(_is_punctuation("A"))
+        self.assertFalse(_is_punctuation(" "))
+
+    @slow
+    def test_sequence_builders(self):
+        tokenizer = self.tokenizer_class.from_pretrained("microsoft/prophetnet-large-uncased")
+
+        text = tokenizer.encode("sequence builders", add_special_tokens=False)
+        text_2 = tokenizer.encode("multi-sequence build", add_special_tokens=False)
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        assert encoded_sentence == text + [102]
+        assert encoded_pair == text + [102] + text_2 + [102]

tests/test_tokenization_xlm_prophetnet.py

@@ -0,0 +1,125 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team, The Microsoft Research team.
+#
+# 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 os
+import unittest
+
+from transformers.file_utils import cached_property
+from transformers.testing_utils import slow
+from transformers.tokenization_xlm_prophetnet import SPIECE_UNDERLINE, XLMProphetNetTokenizer
+
+from .test_tokenization_common import TokenizerTesterMixin
+
+
+SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)), "fixtures/test_sentencepiece.model")
+
+
+class XLMProphetNetTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+
+    tokenizer_class = XLMProphetNetTokenizer
+    test_rust_tokenizer = False
+
+    def setUp(self):
+        super().setUp()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = XLMProphetNetTokenizer(SAMPLE_VOCAB, keep_accents=True)
+        tokenizer.save_pretrained(self.tmpdirname)
+
+    def test_full_tokenizer(self):
+        tokenizer = XLMProphetNetTokenizer(SAMPLE_VOCAB, keep_accents=True)
+
+        tokens = tokenizer.tokenize("This is a test")
+        self.assertListEqual(tokens, ["▁This", "▁is", "▁a", "▁t", "est"])
+
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(tokens),
+            [value + tokenizer.fairseq_offset for value in [285, 46, 10, 170, 382]],
+        )
+
+        tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
+        self.assertListEqual(
+            tokens,
+            [
+                SPIECE_UNDERLINE + "I",
+                SPIECE_UNDERLINE + "was",
+                SPIECE_UNDERLINE + "b",
+                "or",
+                "n",
+                SPIECE_UNDERLINE + "in",
+                SPIECE_UNDERLINE + "",
+                "9",
+                "2",
+                "0",
+                "0",
+                "0",
+                ",",
+                SPIECE_UNDERLINE + "and",
+                SPIECE_UNDERLINE + "this",
+                SPIECE_UNDERLINE + "is",
+                SPIECE_UNDERLINE + "f",
+                "al",
+                "s",
+                "é",
+                ".",
+            ],
+        )
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(
+            ids,
+            [
+                value + tokenizer.fairseq_offset
+                for value in [8, 21, 84, 55, 24, 19, 7, -9, 602, 347, 347, 347, 3, 12, 66, 46, 72, 80, 6, -9, 4]
+            ],
+        )
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(
+            back_tokens,
+            [
+                SPIECE_UNDERLINE + "I",
+                SPIECE_UNDERLINE + "was",
+                SPIECE_UNDERLINE + "b",
+                "or",
+                "n",
+                SPIECE_UNDERLINE + "in",
+                SPIECE_UNDERLINE + "",
+                "[UNK]",
+                "2",
+                "0",
+                "0",
+                "0",
+                ",",
+                SPIECE_UNDERLINE + "and",
+                SPIECE_UNDERLINE + "this",
+                SPIECE_UNDERLINE + "is",
+                SPIECE_UNDERLINE + "f",
+                "al",
+                "s",
+                "[UNK]",
+                ".",
+            ],
+        )
+
+    @cached_property
+    def big_tokenizer(self):
+        return XLMProphetNetTokenizer.from_pretrained("microsoft/xprophetnet-large-wiki100-cased")
+
+    @slow
+    def test_tokenization_base_easy_symbols(self):
+        symbols = "Hello World!"
+        original_tokenizer_encodings = [35389, 6672, 49, 2]
+        self.assertListEqual(original_tokenizer_encodings, self.big_tokenizer.encode(symbols))