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

15
import inspect
16
import unittest
17

18
from transformers import ResNetConfig, is_flax_available
19
from transformers.testing_utils import require_flax, slow
20
from transformers.utils import cached_property, is_vision_available
21

22
from ...test_configuration_common import ConfigTester
23
from ...test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor
24

25

26
if is_flax_available():
27
    import jax
28
    import jax.numpy as jnp
29

30
    from transformers.models.resnet.modeling_flax_resnet import FlaxResNetForImageClassification, FlaxResNetModel
31

32
if is_vision_available():
33
    from PIL import Image
34

35
    from transformers import AutoImageProcessor
36

37

38
class FlaxResNetModelTester(unittest.TestCase):
39
    def __init__(
40
        self,
41
        parent,
42
        batch_size=3,
43
        image_size=32,
44
        num_channels=3,
45
        embeddings_size=10,
46
        hidden_sizes=[10, 20, 30, 40],
47
        depths=[1, 1, 2, 1],
48
        is_training=True,
49
        use_labels=True,
50
        hidden_act="relu",
51
        num_labels=3,
52
        scope=None,
53
    ):
54
        self.parent = parent
55
        self.batch_size = batch_size
56
        self.image_size = image_size
57
        self.num_channels = num_channels
58
        self.embeddings_size = embeddings_size
59
        self.hidden_sizes = hidden_sizes
60
        self.depths = depths
61
        self.is_training = is_training
62
        self.use_labels = use_labels
63
        self.hidden_act = hidden_act
64
        self.num_labels = num_labels
65
        self.scope = scope
66
        self.num_stages = len(hidden_sizes)
67

68
    def prepare_config_and_inputs(self):
69
        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
70

71
        config = self.get_config()
72

73
        return config, pixel_values
74

75
    def get_config(self):
76
        return ResNetConfig(
77
            num_channels=self.num_channels,
78
            embeddings_size=self.embeddings_size,
79
            hidden_sizes=self.hidden_sizes,
80
            depths=self.depths,
81
            hidden_act=self.hidden_act,
82
            num_labels=self.num_labels,
83
            image_size=self.image_size,
84
        )
85

86
    def create_and_check_model(self, config, pixel_values):
87
        model = FlaxResNetModel(config=config)
88
        result = model(pixel_values)
89

90
        # Output shape (b, c, h, w)
91
        self.parent.assertEqual(
92
            result.last_hidden_state.shape,
93
            (self.batch_size, self.hidden_sizes[-1], self.image_size // 32, self.image_size // 32),
94
        )
95

96
    def create_and_check_for_image_classification(self, config, pixel_values):
97
        config.num_labels = self.num_labels
98
        model = FlaxResNetForImageClassification(config=config)
99
        result = model(pixel_values)
100
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
101

102
    def prepare_config_and_inputs_for_common(self):
103
        config_and_inputs = self.prepare_config_and_inputs()
104
        config, pixel_values = config_and_inputs
105
        inputs_dict = {"pixel_values": pixel_values}
106
        return config, inputs_dict
107

108

109
@require_flax
110
class FlaxResNetModelTest(FlaxModelTesterMixin, unittest.TestCase):
111
    all_model_classes = (FlaxResNetModel, FlaxResNetForImageClassification) if is_flax_available() else ()
112

113
    is_encoder_decoder = False
114
    test_head_masking = False
115
    has_attentions = False
116

117
    def setUp(self) -> None:
118
        self.model_tester = FlaxResNetModelTester(self)
119
        self.config_tester = ConfigTester(self, config_class=ResNetConfig, has_text_modality=False)
120

121
    def test_config(self):
122
        self.create_and_test_config_common_properties()
123
        self.config_tester.create_and_test_config_to_json_string()
124
        self.config_tester.create_and_test_config_to_json_file()
125
        self.config_tester.create_and_test_config_from_and_save_pretrained()
126
        self.config_tester.create_and_test_config_with_num_labels()
127
        self.config_tester.check_config_can_be_init_without_params()
128
        self.config_tester.check_config_arguments_init()
129

130
    def create_and_test_config_common_properties(self):
131
        return
132

133
    def test_model(self):
134
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
135
        self.model_tester.create_and_check_model(*config_and_inputs)
136

137
    def test_for_image_classification(self):
138
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
139
        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
140

141
    @unittest.skip(reason="ResNet does not use inputs_embeds")
142
    def test_inputs_embeds(self):
143
        pass
144

145
    @unittest.skip(reason="ResNet does not support input and output embeddings")
146
    def test_model_common_attributes(self):
147
        pass
148

149
    def test_forward_signature(self):
150
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
151

152
        for model_class in self.all_model_classes:
153
            model = model_class(config)
154
            signature = inspect.signature(model.__call__)
155
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
156
            arg_names = [*signature.parameters.keys()]
157

158
            expected_arg_names = ["pixel_values"]
159
            self.assertListEqual(arg_names[:1], expected_arg_names)
160

161
    def test_hidden_states_output(self):
162
        def check_hidden_states_output(inputs_dict, config, model_class):
163
            model = model_class(config)
164

165
            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
166

167
            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
168

169
            expected_num_stages = self.model_tester.num_stages
170
            self.assertEqual(len(hidden_states), expected_num_stages + 1)
171

172
    @unittest.skip(reason="ResNet does not use feedforward chunking")
173
    def test_feed_forward_chunking(self):
174
        pass
175

176
    def test_jit_compilation(self):
177
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
178

179
        for model_class in self.all_model_classes:
180
            with self.subTest(model_class.__name__):
181
                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
182
                model = model_class(config)
183

184
                @jax.jit
185
                def model_jitted(pixel_values, **kwargs):
186
                    return model(pixel_values=pixel_values, **kwargs)
187

188
                with self.subTest("JIT Enabled"):
189
                    jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()
190

191
                with self.subTest("JIT Disabled"):
192
                    with jax.disable_jit():
193
                        outputs = model_jitted(**prepared_inputs_dict).to_tuple()
194

195
                self.assertEqual(len(outputs), len(jitted_outputs))
196
                for jitted_output, output in zip(jitted_outputs, outputs):
197
                    self.assertEqual(jitted_output.shape, output.shape)
198

199

200
# We will verify our results on an image of cute cats
201
def prepare_img():
202
    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
203
    return image
204

205

206
@require_flax
207
class FlaxResNetModelIntegrationTest(unittest.TestCase):
208
    @cached_property
209
    def default_image_processor(self):
210
        return AutoImageProcessor.from_pretrained("microsoft/resnet-50") if is_vision_available() else None
211

212
    @slow
213
    def test_inference_image_classification_head(self):
214
        model = FlaxResNetForImageClassification.from_pretrained("microsoft/resnet-50")
215

216
        image_processor = self.default_image_processor
217
        image = prepare_img()
218
        inputs = image_processor(images=image, return_tensors="np")
219

220
        outputs = model(**inputs)
221

222
        # verify the logits
223
        expected_shape = (1, 1000)
224
        self.assertEqual(outputs.logits.shape, expected_shape)
225

226
        expected_slice = jnp.array([-11.1069, -9.7877, -8.3777])
227

228
        self.assertTrue(jnp.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
229