first commit

tests/models/xlm_roberta/test_modeling_xlm_roberta.py

@@ -0,0 +1,88 @@
+# 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 is_torch_available
+from transformers.testing_utils import (
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+    slow,
+)
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import XLMRobertaModel
+
+
+@require_sentencepiece
+@require_tokenizers
+@require_torch
+@slow
+class XLMRobertaModelIntegrationTest(unittest.TestCase):
+    def test_xlm_roberta_base(self):
+        model = XLMRobertaModel.from_pretrained("FacebookAI/xlm-roberta-base", attn_implementation="eager")
+        input_ids = torch.tensor([[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]])
+        # 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]]
+        )
+        #  xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.base')
+        #  xlmr.eval()
+        #  expected_output_values_last_dim = xlmr.extract_features(input_ids[0])[:, :, -1]
+        with torch.no_grad():
+            output = model(input_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        torch.testing.assert_close(output[:, :, -1], expected_output_values_last_dim, rtol=1e-3, atol=1e-3)
+
+    def test_xlm_roberta_base_sdpa(self):
+        input_ids = torch.tensor([[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]])
+        # 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]]
+        )
+
+        model = XLMRobertaModel.from_pretrained("FacebookAI/xlm-roberta-base", attn_implementation="sdpa")
+        with torch.no_grad():
+            output = model(input_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        torch.testing.assert_close(output[:, :, -1], expected_output_values_last_dim, rtol=1e-3, atol=1e-3)
+
+    def test_xlm_roberta_large(self):
+        model = XLMRobertaModel.from_pretrained("FacebookAI/xlm-roberta-large")
+        input_ids = torch.tensor([[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]])
+        # 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]]
+        )
+        #  xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.large')
+        #  xlmr.eval()
+        #  expected_output_values_last_dim = xlmr.extract_features(input_ids[0])[:, :, -1]
+        with torch.no_grad():
+            output = model(input_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        torch.testing.assert_close(output[:, :, -1], expected_output_values_last_dim, rtol=1e-3, atol=1e-3)

tests/models/xlm_roberta/test_tokenization_xlm_roberta.py

@@ -0,0 +1,33 @@
+# 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 XLMRobertaTokenizer
+from transformers.testing_utils import require_sentencepiece, require_tokenizers
+
+# import cached_property
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+@require_sentencepiece
+@require_tokenizers
+class XLMRobertaTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+    from_pretrained_id = "FacebookAI/xlm-roberta-base"
+    tokenizer_class = XLMRobertaTokenizer
+
+    integration_expected_tokens = ['▁This', '▁is', '▁a', '▁test', '▁', '😊', '▁I', '▁was', '▁born', '▁in', '▁9', '2000', ',', '▁and', '▁this', '▁is', '▁fals', 'é', '.', '▁', '生活的', '真', '谛', '是', '▁Hi', '▁Hello', '▁Hi', '▁Hello', '▁Hello', '<s>', '▁hi', '<s>', '▁there', '▁The', '▁following', '▁string', '▁should', '▁be', '▁properly', '▁en', 'code', 'd', ':', '▁Hello', '.', '▁But', '▁ir', 'd', '▁and', '▁ปี', '▁ir', 'd', '▁ด', '▁Hey', '▁how', '▁are', '▁you', '▁doing']  # fmt: skip
+    integration_expected_token_ids = [3293, 83, 10, 3034, 6, 82803, 87, 509, 103122, 23, 483, 13821, 4, 136, 903, 83, 84047, 446, 5, 6, 62668, 5364, 245875, 354, 2673, 35378, 2673, 35378, 35378, 0, 1274, 0, 2685, 581, 25632, 79315, 5608, 186, 155965, 22, 40899, 71, 12, 35378, 5, 4966, 193, 71, 136, 10249, 193, 71, 48229, 28240, 3642, 621, 398, 20594]  # fmt: skip
+    expected_tokens_from_ids = ['▁This', '▁is', '▁a', '▁test', '▁', '😊', '▁I', '▁was', '▁born', '▁in', '▁9', '2000', ',', '▁and', '▁this', '▁is', '▁fals', 'é', '.', '▁', '生活的', '真', '谛', '是', '▁Hi', '▁Hello', '▁Hi', '▁Hello', '▁Hello', '<s>', '▁hi', '<s>', '▁there', '▁The', '▁following', '▁string', '▁should', '▁be', '▁properly', '▁en', 'code', 'd', ':', '▁Hello', '.', '▁But', '▁ir', 'd', '▁and', '▁ปี', '▁ir', 'd', '▁ด', '▁Hey', '▁how', '▁are', '▁you', '▁doing']  # fmt: skip
+    integration_expected_decoded_text = "This is a test 😊 I was born in 92000, and this is falsé. 生活的真谛是 Hi Hello Hi Hello Hello<s> hi<s> there The following string should be properly encoded: Hello. But ird and ปี ird ด Hey how are you doing"