test_modeling_marian.py 9.5 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_torch, slow, torch_device
23
24
25
26


if is_torch_available():
    import torch
27
28
29
30
31
32
33
34
    from transformers import (
        AutoTokenizer,
        MarianConfig,
        AutoConfig,
        AutoModelWithLMHead,
        MarianTokenizer,
        MarianMTModel,
    )
35
36
37
38
39
    from transformers.convert_marian_to_pytorch import (
        convert_hf_name_to_opus_name,
        convert_opus_name_to_hf_name,
        ORG_NAME,
    )
40
    from transformers.pipelines import TranslationPipeline
41
42
43
44


class ModelManagementTests(unittest.TestCase):
    @slow
45
    def test_model_names(self):
46
        model_list = HfApi().model_list()
47
48
49
50
        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)
51
52
53


@require_torch
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
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

75
76
    @classmethod
    def setUpClass(cls) -> None:
77
78
        cls.model_name = f"Helsinki-NLP/opus-mt-{cls.src}-{cls.tgt}"
        cls.tokenizer: MarianTokenizer = AutoTokenizer.from_pretrained(cls.model_name)
79
80
81
82
83
        cls.eos_token_id = cls.tokenizer.eos_token_id
        return cls

    @cached_property
    def model(self):
84
85
86
87
88
89
        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)

90
91
92
93
94
        if torch_device == "cuda":
            return model.half()
        else:
            return model

95
96
97
98
99
    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):
100
        model_inputs = self.tokenizer.prepare_seq2seq_batch(src_texts=self.src_text, **tokenizer_kwargs).to(
101
102
            torch_device
        )
103
        self.assertEqual(self.model.device, model_inputs.input_ids.device)
104
        generated_ids = self.model.generate(
105
            model_inputs.input_ids, attention_mask=model_inputs.attention_mask, num_beams=2
106
107
108
109
110
111
        )
        generated_words = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
        return generated_words


class TestMarian_EN_DE_More(MarianIntegrationTest):
112
113
    @slow
    def test_forward(self):
114
        src, tgt = ["I am a small frog"], ["Ich bin ein kleiner Frosch."]
115
        expected_ids = [38, 121, 14, 697, 38848, 0]
116

117
        model_inputs: dict = self.tokenizer.prepare_seq2seq_batch(src, tgt_texts=tgt).to(torch_device)
118
        self.assertListEqual(expected_ids, model_inputs.input_ids[0].tolist())
119
120
121
122
123
124
125
126
127
128
129

        desired_keys = {
            "input_ids",
            "attention_mask",
            "decoder_input_ids",
            "decoder_attention_mask",
        }
        self.assertSetEqual(desired_keys, set(model_inputs.keys()))
        with torch.no_grad():
            logits, *enc_features = self.model(**model_inputs)
        max_indices = logits.argmax(-1)
130
        self.tokenizer.batch_decode(max_indices)
131

132
133
    def test_unk_support(self):
        t = self.tokenizer
134
        ids = t.prepare_seq2seq_batch(["||"]).to(torch_device).input_ids[0].tolist()
135
136
137
        expected = [t.unk_token_id, t.unk_token_id, t.eos_token_id]
        self.assertEqual(expected, ids)

138
    def test_pad_not_split(self):
139
        input_ids_w_pad = self.tokenizer.prepare_seq2seq_batch(["I am a small frog <pad>"]).input_ids[0].tolist()
140
        expected_w_pad = [38, 121, 14, 697, 38848, self.tokenizer.pad_token_id, 0]  # pad
141
        self.assertListEqual(expected_w_pad, input_ids_w_pad)
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189

    @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)


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()


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()


class TestMarian_RU_FR(MarianIntegrationTest):
    src = "ru"
    tgt = "fr"
    src_text = ["Он показал мне рукопись своей новой пьесы."]
190
    expected_text = ["Il m'a montré le manuscrit de sa nouvelle pièce."]
191

192
    @slow
193
194
195
196
197
198
199
    def test_batch_generation_ru_fr(self):
        self._assert_generated_batch_equal_expected()


class TestMarian_MT_EN(MarianIntegrationTest):
    src = "mt"
    tgt = "en"
200
201
    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."]
202

203
    @slow
204
205
206
207
    def test_batch_generation_mt_en(self):
        self._assert_generated_batch_equal_expected()


208
209
210
211
212
213
214
215
216
217
218
class TestMarian_eng_zho(MarianIntegrationTest):
    src = "eng"
    tgt = "zho"
    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()


219
220
class TestMarian_en_ROMANCE(MarianIntegrationTest):
    """Multilingual on target side."""
221

222
223
224
225
226
227
228
229
230
231
232
233
    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.",
    ]
234

235
236
    @slow
    def test_batch_generation_en_ROMANCE_multi(self):
237
238
        self._assert_generated_batch_equal_expected()

239
240
241
242
    def test_tokenizer_handles_empty(self):
        normalized = self.tokenizer.normalize("")
        self.assertIsInstance(normalized, str)
        with self.assertRaises(ValueError):
243
            self.tokenizer.prepare_seq2seq_batch([""])
244

245
    def test_pipeline(self):
246
247
        device = 0 if torch_device == "cuda" else -1
        pipeline = TranslationPipeline(self.model, self.tokenizer, framework="pt", device=device)
248
249
250
        output = pipeline(self.src_text)
        self.assertEqual(self.expected_text, [x["translation_text"] for x in output])

251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273

@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)