Add integration tests for xlm roberta modelling and xlm roberta tokenzier (#3014)

* add first files

* add xlm roberta integration tests

* make style

* flake 8 issues solved

tests/test_modeling_xlm_roberta.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# 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 .utils import slow
+
+
+if is_torch_available():
+    import torch
+    from transformers import XLMRobertaModel
+
+
+class XLMRobertaModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_xlm_roberta_base(self):
+        model = XLMRobertaModel.from_pretrained("xlm-roberta-base")
+        input_ids = torch.tensor([0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]).unsqueeze(
+            0
+        )  # The dog is cute and lives in the garden house
+
+        expected_output_shape = torch.Size((1, 12, 768))  # batch_size, sequence_length, embedding_vector_dim
+        expected_output_values_last_dim = torch.tensor(
+            [-0.0101, 0.1218, -0.0803, 0.0801, 0.1327, 0.0776, -0.1215, 0.2383, 0.3338, 0.3106, 0.0300, 0.0252]
+        ).unsqueeze(0)
+        #  xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.base')
+        #  xlmr.eval()
+        #  expected_output_values_last_dim = xlmr.extract_features(input_ids[0])[:, :, -1]
+
+        output = model(input_ids)[0].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))
+
+    @slow
+    def test_xlm_roberta_large(self):
+        model = XLMRobertaModel.from_pretrained("xlm-roberta-large")
+        input_ids = torch.tensor([0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]).unsqueeze(
+            0
+        )  # The dog is cute and lives in the garden house
+
+        expected_output_shape = torch.Size((1, 12, 1024))  # batch_size, sequence_length, embedding_vector_dim
+        expected_output_values_last_dim = torch.tensor(
+            [-0.0699, -0.0318, 0.0705, -0.1241, 0.0999, -0.0520, 0.1004, -0.1838, -0.4704, 0.1437, 0.0821, 0.0126]
+        ).unsqueeze(0)
+        #  xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.large')
+        #  xlmr.eval()
+        #  expected_output_values_last_dim = xlmr.extract_features(input_ids[0])[:, :, -1]
+
+        output = model(input_ids)[0].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))

tests/test_tokenization_xlm_roberta.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# 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.tokenization_xlm_roberta import XLMRobertaTokenizer
+
+from .utils import slow
+
+
+class XLMRobertaTokenizationIntegrationTest(unittest.TestCase):
+    @slow
+    def test_tokenization_base_easy_symbols(self):
+        tokenizer = XLMRobertaTokenizer.from_pretrained("xlm-roberta-base")
+
+        symbols = "Hello World!"
+        original_tokenizer_encodings = [0, 35378, 6661, 38, 2]
+        # xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.base')  # xlmr.large has same tokenizer
+        # xlmr.eval()
+        # xlmr.encode(symbols)
+
+        self.assertListEqual(original_tokenizer_encodings, tokenizer.encode(symbols))
+
+    @slow
+    def test_tokenization_base_hard_symbols(self):
+        tokenizer = XLMRobertaTokenizer.from_pretrained("xlm-roberta-base")
+
+        symbols = 'This is a very long text with a lot of weird characters, such as: . , ~ ? ( ) " [ ] ! : - . Also we will add words that should not exsist and be tokenized to <unk>, such as saoneuhaoesuth'
+        original_tokenizer_encodings = [
+            0,
+            3293,
+            83,
+            10,
+            4552,
+            4989,
+            7986,
+            678,
+            10,
+            5915,
+            111,
+            179459,
+            124850,
+            4,
+            6044,
+            237,
+            12,
+            6,
+            5,
+            6,
+            4,
+            6780,
+            705,
+            15,
+            1388,
+            44,
+            378,
+            10114,
+            711,
+            152,
+            20,
+            6,
+            5,
+            22376,
+            642,
+            1221,
+            15190,
+            34153,
+            450,
+            5608,
+            959,
+            1119,
+            57702,
+            136,
+            186,
+            47,
+            1098,
+            29367,
+            47,
+            4426,
+            3678,
+            2740,
+            4,
+            6044,
+            237,
+            6284,
+            50901,
+            528,
+            31,
+            90,
+            34,
+            927,
+            2,
+        ]
+        # xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.base')  # xlmr.large has same tokenizer
+        # xlmr.eval()
+        # xlmr.encode(symbols)
+
+        self.assertListEqual(original_tokenizer_encodings, tokenizer.encode(symbols))