test_modeling_marian.py 10.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
# coding=utf-8
# Copyright 2020 HuggingFace Inc. 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.file_utils import cached_property
21
from transformers.hf_api import HfApi
22
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
23
24
25
26


if is_torch_available():
    import torch
27

28
29
30
    from transformers import (
        AutoConfig,
        AutoModelWithLMHead,
31
32
        AutoTokenizer,
        MarianConfig,
33
        MarianMTModel,
34
        MarianTokenizer,
35
    )
36
    from transformers.convert_marian_to_pytorch import (
37
        ORG_NAME,
38
39
40
        convert_hf_name_to_opus_name,
        convert_opus_name_to_hf_name,
    )
Sam Shleifer's avatar
Sam Shleifer committed
41
    from transformers.modeling_bart import shift_tokens_right
42
    from transformers.pipelines import TranslationPipeline
43
44
45
46


class ModelManagementTests(unittest.TestCase):
    @slow
47
    def test_model_names(self):
48
        model_list = HfApi().model_list()
49
50
51
52
        model_ids = [x.modelId for x in model_list if x.modelId.startswith(ORG_NAME)]
        bad_model_ids = [mid for mid in model_ids if "+" in model_ids]
        self.assertListEqual([], bad_model_ids)
        self.assertGreater(len(model_ids), 500)
53
54
55


@require_torch
56
57
@require_sentencepiece
@require_tokenizers
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
class MarianIntegrationTest(unittest.TestCase):
    src = "en"
    tgt = "de"
    src_text = [
        "I am a small frog.",
        "Now I can forget the 100 words of german that I know.",
        "Tom asked his teacher for advice.",
        "That's how I would do it.",
        "Tom really admired Mary's courage.",
        "Turn around and close your eyes.",
    ]
    expected_text = [
        "Ich bin ein kleiner Frosch.",
        "Jetzt kann ich die 100 Wörter des Deutschen vergessen, die ich kenne.",
        "Tom bat seinen Lehrer um Rat.",
        "So würde ich das machen.",
        "Tom bewunderte Marias Mut wirklich.",
        "Drehen Sie sich um und schließen Sie die Augen.",
    ]
    # ^^ actual C++ output differs slightly: (1) des Deutschen removed, (2) ""-> "O", (3) tun -> machen

79
80
    @classmethod
    def setUpClass(cls) -> None:
81
82
        cls.model_name = f"Helsinki-NLP/opus-mt-{cls.src}-{cls.tgt}"
        cls.tokenizer: MarianTokenizer = AutoTokenizer.from_pretrained(cls.model_name)
83
84
85
86
87
        cls.eos_token_id = cls.tokenizer.eos_token_id
        return cls

    @cached_property
    def model(self):
88
89
90
91
92
93
        model: MarianMTModel = AutoModelWithLMHead.from_pretrained(self.model_name).to(torch_device)
        c = model.config
        self.assertListEqual(c.bad_words_ids, [[c.pad_token_id]])
        self.assertEqual(c.max_length, 512)
        self.assertEqual(c.decoder_start_token_id, c.pad_token_id)

94
95
96
97
98
        if torch_device == "cuda":
            return model.half()
        else:
            return model

99
100
101
102
103
    def _assert_generated_batch_equal_expected(self, **tokenizer_kwargs):
        generated_words = self.translate_src_text(**tokenizer_kwargs)
        self.assertListEqual(self.expected_text, generated_words)

    def translate_src_text(self, **tokenizer_kwargs):
104
        model_inputs = self.tokenizer.prepare_seq2seq_batch(src_texts=self.src_text, **tokenizer_kwargs).to(
105
106
            torch_device
        )
107
        self.assertEqual(self.model.device, model_inputs.input_ids.device)
108
        generated_ids = self.model.generate(
109
            model_inputs.input_ids, attention_mask=model_inputs.attention_mask, num_beams=2
110
111
112
113
114
        )
        generated_words = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
        return generated_words


115
116
@require_sentencepiece
@require_tokenizers
117
class TestMarian_EN_DE_More(MarianIntegrationTest):
118
119
    @slow
    def test_forward(self):
120
        src, tgt = ["I am a small frog"], ["Ich bin ein kleiner Frosch."]
121
        expected_ids = [38, 121, 14, 697, 38848, 0]
122

123
        model_inputs: dict = self.tokenizer.prepare_seq2seq_batch(src, tgt_texts=tgt).to(torch_device)
Sam Shleifer's avatar
Sam Shleifer committed
124

125
        self.assertListEqual(expected_ids, model_inputs.input_ids[0].tolist())
126
127
128
129

        desired_keys = {
            "input_ids",
            "attention_mask",
Sam Shleifer's avatar
Sam Shleifer committed
130
            "labels",
131
132
        }
        self.assertSetEqual(desired_keys, set(model_inputs.keys()))
Sam Shleifer's avatar
Sam Shleifer committed
133
134
135
        model_inputs["decoder_input_ids"] = shift_tokens_right(model_inputs.labels, self.tokenizer.pad_token_id)
        model_inputs["return_dict"] = True
        model_inputs["use_cache"] = False
136
        with torch.no_grad():
Sam Shleifer's avatar
Sam Shleifer committed
137
138
            outputs = self.model(**model_inputs)
        max_indices = outputs.logits.argmax(-1)
139
        self.tokenizer.batch_decode(max_indices)
140

141
142
    def test_unk_support(self):
        t = self.tokenizer
143
        ids = t.prepare_seq2seq_batch(["||"]).to(torch_device).input_ids[0].tolist()
144
145
146
        expected = [t.unk_token_id, t.unk_token_id, t.eos_token_id]
        self.assertEqual(expected, ids)

147
    def test_pad_not_split(self):
148
        input_ids_w_pad = self.tokenizer.prepare_seq2seq_batch(["I am a small frog <pad>"]).input_ids[0].tolist()
149
        expected_w_pad = [38, 121, 14, 697, 38848, self.tokenizer.pad_token_id, 0]  # pad
150
        self.assertListEqual(expected_w_pad, input_ids_w_pad)
151
152
153
154
155
156
157
158
159
160

    @slow
    def test_batch_generation_en_de(self):
        self._assert_generated_batch_equal_expected()

    def test_auto_config(self):
        config = AutoConfig.from_pretrained(self.model_name)
        self.assertIsInstance(config, MarianConfig)


161
162
@require_sentencepiece
@require_tokenizers
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
class TestMarian_EN_FR(MarianIntegrationTest):
    src = "en"
    tgt = "fr"
    src_text = [
        "I am a small frog.",
        "Now I can forget the 100 words of german that I know.",
    ]
    expected_text = [
        "Je suis une petite grenouille.",
        "Maintenant, je peux oublier les 100 mots d'allemand que je connais.",
    ]

    @slow
    def test_batch_generation_en_fr(self):
        self._assert_generated_batch_equal_expected()


180
181
@require_sentencepiece
@require_tokenizers
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
class TestMarian_FR_EN(MarianIntegrationTest):
    src = "fr"
    tgt = "en"
    src_text = [
        "Donnez moi le micro.",
        "Tom et Mary étaient assis à une table.",  # Accents
    ]
    expected_text = [
        "Give me the microphone.",
        "Tom and Mary were sitting at a table.",
    ]

    @slow
    def test_batch_generation_fr_en(self):
        self._assert_generated_batch_equal_expected()


199
200
@require_sentencepiece
@require_tokenizers
201
202
203
204
class TestMarian_RU_FR(MarianIntegrationTest):
    src = "ru"
    tgt = "fr"
    src_text = ["Он показал мне рукопись своей новой пьесы."]
205
    expected_text = ["Il m'a montré le manuscrit de sa nouvelle pièce."]
206

207
    @slow
208
209
210
211
    def test_batch_generation_ru_fr(self):
        self._assert_generated_batch_equal_expected()


212
213
@require_sentencepiece
@require_tokenizers
214
215
216
class TestMarian_MT_EN(MarianIntegrationTest):
    src = "mt"
    tgt = "en"
217
218
    src_text = ["Billi messu b'mod ġentili, Ġesù fejjaq raġel li kien milqut bil - marda kerha tal - ġdiem."]
    expected_text = ["Touching gently, Jesus healed a man who was affected by the sad disease of leprosy."]
219

220
    @slow
221
222
223
224
    def test_batch_generation_mt_en(self):
        self._assert_generated_batch_equal_expected()


225
226
@require_sentencepiece
@require_tokenizers
Sam Shleifer's avatar
Sam Shleifer committed
227
228
229
class TestMarian_en_zh(MarianIntegrationTest):
    src = "en"
    tgt = "zh"
230
231
232
233
234
235
236
237
    src_text = ["My name is Wolfgang and I live in Berlin"]
    expected_text = ["我叫沃尔夫冈 我住在柏林"]

    @slow
    def test_batch_generation_eng_zho(self):
        self._assert_generated_batch_equal_expected()


238
239
@require_sentencepiece
@require_tokenizers
240
241
class TestMarian_en_ROMANCE(MarianIntegrationTest):
    """Multilingual on target side."""
242

243
244
245
246
247
248
249
250
251
252
253
254
    src = "en"
    tgt = "ROMANCE"
    src_text = [
        ">>fr<< Don't spend so much time watching TV.",
        ">>pt<< Your message has been sent.",
        ">>es<< He's two years older than me.",
    ]
    expected_text = [
        "Ne passez pas autant de temps à regarder la télé.",
        "A sua mensagem foi enviada.",
        "Es dos años más viejo que yo.",
    ]
255

256
257
    @slow
    def test_batch_generation_en_ROMANCE_multi(self):
258
259
        self._assert_generated_batch_equal_expected()

260
261
262
263
    def test_tokenizer_handles_empty(self):
        normalized = self.tokenizer.normalize("")
        self.assertIsInstance(normalized, str)
        with self.assertRaises(ValueError):
264
            self.tokenizer.prepare_seq2seq_batch([""])
265

266
    def test_pipeline(self):
267
268
        device = 0 if torch_device == "cuda" else -1
        pipeline = TranslationPipeline(self.model, self.tokenizer, framework="pt", device=device)
269
270
271
        output = pipeline(self.src_text)
        self.assertEqual(self.expected_text, [x["translation_text"] for x in output])

272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294

@require_torch
class TestConversionUtils(unittest.TestCase):
    def test_renaming_multilingual(self):
        old_names = [
            "opus-mt-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-fi",
            "opus-mt-cmn+cn-fi",  # no group
            "opus-mt-en-de",  # standard name
            "opus-mt-en-de",  # standard name
        ]
        expected = ["opus-mt-ZH-fi", "opus-mt-cmn_cn-fi", "opus-mt-en-de", "opus-mt-en-de"]
        self.assertListEqual(expected, [convert_opus_name_to_hf_name(x) for x in old_names])

    def test_undoing_renaming(self):
        hf_names = ["opus-mt-ZH-fi", "opus-mt-cmn_cn-fi", "opus-mt-en-de", "opus-mt-en-de"]
        converted_opus_names = [convert_hf_name_to_opus_name(x) for x in hf_names]
        expected_opus_names = [
            "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-fi",
            "cmn+cn-fi",
            "en-de",  # standard name
            "en-de",
        ]
        self.assertListEqual(expected_opus_names, converted_opus_names)