transformers

1
# Copyright 2023 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
import shutil
15
import tempfile
16
import unittest
17

18
import numpy as np
19
import pytest
20

21
from transformers.testing_utils import require_vision
22
from transformers.utils import is_vision_available
23

24

25
if is_vision_available():
26
    from PIL import Image
27

28
    from transformers import (
29
        AutoProcessor,
30
        BertTokenizerFast,
31
        BlipImageProcessor,
32
        GPT2Tokenizer,
33
        InstructBlipProcessor,
34
        PreTrainedTokenizerFast,
35
    )
36

37

38
@require_vision
39
class InstructBlipProcessorTest(unittest.TestCase):
40
    def setUp(self):
41
        self.tmpdirname = tempfile.mkdtemp()
42

43
        image_processor = BlipImageProcessor()
44
        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-GPT2Model")
45
        qformer_tokenizer = BertTokenizerFast.from_pretrained("hf-internal-testing/tiny-random-bert")
46

47
        processor = InstructBlipProcessor(image_processor, tokenizer, qformer_tokenizer)
48

49
        processor.save_pretrained(self.tmpdirname)
50

51
    def get_tokenizer(self, **kwargs):
52
        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
53

54
    def get_image_processor(self, **kwargs):
55
        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
56

57
    def get_qformer_tokenizer(self, **kwargs):
58
        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).qformer_tokenizer
59

60
    def tearDown(self):
61
        shutil.rmtree(self.tmpdirname)
62

63
    def prepare_image_inputs(self):
64
        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
65
        or a list of PyTorch tensors if one specifies torchify=True.
66
        """
67

68
        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
69

70
        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
71

72
        return image_inputs
73

74
    def test_save_load_pretrained_additional_features(self):
75
        processor = InstructBlipProcessor(
76
            tokenizer=self.get_tokenizer(),
77
            image_processor=self.get_image_processor(),
78
            qformer_tokenizer=self.get_qformer_tokenizer(),
79
        )
80
        processor.save_pretrained(self.tmpdirname)
81

82
        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
83
        image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
84

85
        processor = InstructBlipProcessor.from_pretrained(
86
            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
87
        )
88

89
        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
90
        self.assertIsInstance(processor.tokenizer, PreTrainedTokenizerFast)
91

92
        self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
93
        self.assertIsInstance(processor.image_processor, BlipImageProcessor)
94
        self.assertIsInstance(processor.qformer_tokenizer, BertTokenizerFast)
95

96
    def test_image_processor(self):
97
        image_processor = self.get_image_processor()
98
        tokenizer = self.get_tokenizer()
99
        qformer_tokenizer = self.get_qformer_tokenizer()
100

101
        processor = InstructBlipProcessor(
102
            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
103
        )
104

105
        image_input = self.prepare_image_inputs()
106

107
        input_feat_extract = image_processor(image_input, return_tensors="np")
108
        input_processor = processor(images=image_input, return_tensors="np")
109

110
        for key in input_feat_extract.keys():
111
            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
112

113
    def test_tokenizer(self):
114
        image_processor = self.get_image_processor()
115
        tokenizer = self.get_tokenizer()
116
        qformer_tokenizer = self.get_qformer_tokenizer()
117

118
        processor = InstructBlipProcessor(
119
            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
120
        )
121

122
        input_str = "lower newer"
123

124
        encoded_processor = processor(text=input_str)
125

126
        encoded_tokens = tokenizer(input_str, return_token_type_ids=False)
127
        encoded_tokens_qformer = qformer_tokenizer(input_str, return_token_type_ids=False)
128

129
        for key in encoded_tokens.keys():
130
            self.assertListEqual(encoded_tokens[key], encoded_processor[key])
131

132
        for key in encoded_tokens_qformer.keys():
133
            self.assertListEqual(encoded_tokens_qformer[key], encoded_processor["qformer_" + key])
134

135
    def test_processor(self):
136
        image_processor = self.get_image_processor()
137
        tokenizer = self.get_tokenizer()
138
        qformer_tokenizer = self.get_qformer_tokenizer()
139

140
        processor = InstructBlipProcessor(
141
            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
142
        )
143

144
        input_str = "lower newer"
145
        image_input = self.prepare_image_inputs()
146

147
        inputs = processor(text=input_str, images=image_input)
148

149
        self.assertListEqual(
150
            list(inputs.keys()),
151
            ["input_ids", "attention_mask", "qformer_input_ids", "qformer_attention_mask", "pixel_values"],
152
        )
153

154
        # test if it raises when no input is passed
155
        with pytest.raises(ValueError):
156
            processor()
157

158
    def test_tokenizer_decode(self):
159
        image_processor = self.get_image_processor()
160
        tokenizer = self.get_tokenizer()
161
        qformer_tokenizer = self.get_qformer_tokenizer()
162

163
        processor = InstructBlipProcessor(
164
            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
165
        )
166

167
        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
168

169
        decoded_processor = processor.batch_decode(predicted_ids)
170
        decoded_tok = tokenizer.batch_decode(predicted_ids)
171

172
        self.assertListEqual(decoded_tok, decoded_processor)
173

174
    def test_model_input_names(self):
175
        image_processor = self.get_image_processor()
176
        tokenizer = self.get_tokenizer()
177
        qformer_tokenizer = self.get_qformer_tokenizer()
178

179
        processor = InstructBlipProcessor(
180
            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
181
        )
182

183
        input_str = "lower newer"
184
        image_input = self.prepare_image_inputs()
185

186
        inputs = processor(text=input_str, images=image_input)
187

188
        self.assertListEqual(
189
            list(inputs.keys()),
190
            ["input_ids", "attention_mask", "qformer_input_ids", "qformer_attention_mask", "pixel_values"],
191
        )
192