transformers

1
# Copyright 2022 The HuggingFace Team. All rights reserved.
2
#
3
# Licensed under the Apache License, Version 2.0 (the "License");
4
# you may not use this file except in compliance with the License.
5
# You may obtain a copy of the License at
6
#
7
#     http://www.apache.org/licenses/LICENSE-2.0
8
#
9
# Unless required by applicable law or agreed to in writing, software
10
# distributed under the License is distributed on an "AS IS" BASIS,
11
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
# See the License for the specific language governing permissions and
13
# limitations under the License.
14

15
import shutil
16
import tempfile
17
import unittest
18

19
import pytest
20

21
from transformers import WhisperTokenizer, is_speech_available
22
from transformers.testing_utils import require_sentencepiece, require_torch, require_torchaudio
23

24
from .test_feature_extraction_whisper import floats_list
25

26

27
if is_speech_available():
28
    from transformers import WhisperFeatureExtractor, WhisperProcessor
29

30

31
TRANSCRIBE = 50358
32
NOTIMESTAMPS = 50362
33

34

35
@require_torch
36
@require_torchaudio
37
@require_sentencepiece
38
class WhisperProcessorTest(unittest.TestCase):
39
    def setUp(self):
40
        self.checkpoint = "openai/whisper-small.en"
41
        self.tmpdirname = tempfile.mkdtemp()
42

43
    def get_tokenizer(self, **kwargs):
44
        return WhisperTokenizer.from_pretrained(self.checkpoint, **kwargs)
45

46
    def get_feature_extractor(self, **kwargs):
47
        return WhisperFeatureExtractor.from_pretrained(self.checkpoint, **kwargs)
48

49
    def tearDown(self):
50
        shutil.rmtree(self.tmpdirname)
51

52
    def test_save_load_pretrained_default(self):
53
        tokenizer = self.get_tokenizer()
54
        feature_extractor = self.get_feature_extractor()
55

56
        processor = WhisperProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
57

58
        processor.save_pretrained(self.tmpdirname)
59
        processor = WhisperProcessor.from_pretrained(self.tmpdirname)
60

61
        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
62
        self.assertIsInstance(processor.tokenizer, WhisperTokenizer)
63

64
        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
65
        self.assertIsInstance(processor.feature_extractor, WhisperFeatureExtractor)
66

67
    def test_save_load_pretrained_additional_features(self):
68
        processor = WhisperProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
69
        processor.save_pretrained(self.tmpdirname)
70

71
        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
72
        feature_extractor_add_kwargs = self.get_feature_extractor(do_normalize=False, padding_value=1.0)
73

74
        processor = WhisperProcessor.from_pretrained(
75
            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
76
        )
77

78
        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
79
        self.assertIsInstance(processor.tokenizer, WhisperTokenizer)
80

81
        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
82
        self.assertIsInstance(processor.feature_extractor, WhisperFeatureExtractor)
83

84
    def test_feature_extractor(self):
85
        feature_extractor = self.get_feature_extractor()
86
        tokenizer = self.get_tokenizer()
87

88
        processor = WhisperProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
89

90
        raw_speech = floats_list((3, 1000))
91

92
        input_feat_extract = feature_extractor(raw_speech, return_tensors="np")
93
        input_processor = processor(raw_speech, return_tensors="np")
94

95
        for key in input_feat_extract.keys():
96
            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
97

98
    def test_tokenizer(self):
99
        feature_extractor = self.get_feature_extractor()
100
        tokenizer = self.get_tokenizer()
101

102
        processor = WhisperProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
103

104
        input_str = "This is a test string"
105

106
        encoded_processor = processor(text=input_str)
107

108
        encoded_tok = tokenizer(input_str)
109

110
        for key in encoded_tok.keys():
111
            self.assertListEqual(encoded_tok[key], encoded_processor[key])
112

113
    def test_tokenizer_decode(self):
114
        feature_extractor = self.get_feature_extractor()
115
        tokenizer = self.get_tokenizer()
116

117
        processor = WhisperProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
118

119
        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
120

121
        decoded_processor = processor.batch_decode(predicted_ids)
122
        decoded_tok = tokenizer.batch_decode(predicted_ids)
123

124
        self.assertListEqual(decoded_tok, decoded_processor)
125

126
    def test_model_input_names(self):
127
        feature_extractor = self.get_feature_extractor()
128
        tokenizer = self.get_tokenizer()
129

130
        processor = WhisperProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
131

132
        self.assertListEqual(
133
            processor.model_input_names,
134
            feature_extractor.model_input_names,
135
            msg="`processor` and `feature_extractor` model input names do not match",
136
        )
137

138
    def test_get_decoder_prompt_ids(self):
139
        feature_extractor = self.get_feature_extractor()
140
        tokenizer = self.get_tokenizer()
141

142
        processor = WhisperProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
143
        forced_decoder_ids = processor.get_decoder_prompt_ids(task="transcribe", no_timestamps=True)
144

145
        self.assertIsInstance(forced_decoder_ids, list)
146
        for ids in forced_decoder_ids:
147
            self.assertIsInstance(ids, (list, tuple))
148

149
        expected_ids = [TRANSCRIBE, NOTIMESTAMPS]
150
        self.assertListEqual([ids[-1] for ids in forced_decoder_ids], expected_ids)
151

152
    def test_get_prompt_ids(self):
153
        processor = WhisperProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
154
        prompt_ids = processor.get_prompt_ids("Mr. Quilter")
155
        decoded_prompt = processor.tokenizer.decode(prompt_ids)
156

157
        self.assertListEqual(prompt_ids.tolist(), [50360, 1770, 13, 2264, 346, 353])
158
        self.assertEqual(decoded_prompt, "<|startofprev|> Mr. Quilter")
159

160
    def test_empty_get_prompt_ids(self):
161
        processor = WhisperProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
162
        prompt_ids = processor.get_prompt_ids("")
163
        decoded_prompt = processor.tokenizer.decode(prompt_ids)
164

165
        self.assertListEqual(prompt_ids.tolist(), [50360, 220])
166
        self.assertEqual(decoded_prompt, "<|startofprev|> ")
167

168
    def test_get_prompt_ids_with_special_tokens(self):
169
        processor = WhisperProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
170

171
        def _test_prompt_error_raised_helper(prompt, special_token):
172
            with pytest.raises(ValueError) as excinfo:
173
                processor.get_prompt_ids(prompt)
174
            expected = f"Encountered text in the prompt corresponding to disallowed special token: {special_token}."
175
            self.assertEqual(expected, str(excinfo.value))
176

177
        _test_prompt_error_raised_helper("<|startofprev|> test", "<|startofprev|>")
178
        _test_prompt_error_raised_helper("test <|notimestamps|>", "<|notimestamps|>")
179
        _test_prompt_error_raised_helper("test <|zh|> test <|transcribe|>", "<|zh|>")
180