transformers

1
# coding=utf-8
2
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
3
#
4
# Licensed under the Apache License, Version 2.0 (the "License");
5
# you may not use this file except in compliance with the License.
6
# You may obtain a copy of the License at
7
#
8
#     http://www.apache.org/licenses/LICENSE-2.0
9
#
10
# Unless required by applicable law or agreed to in writing, software
11
# distributed under the License is distributed on an "AS IS" BASIS,
12
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
# See the License for the specific language governing permissions and
14
# limitations under the License.
15
""" Testing suite for the Tensorflow ResNet model. """
16

17

18
from __future__ import annotations
19

20
import inspect
21
import unittest
22

23
import numpy as np
24

25
from transformers import ResNetConfig
26
from transformers.testing_utils import require_tf, require_vision, slow
27
from transformers.utils import cached_property, is_tf_available, is_vision_available
28

29
from ...test_configuration_common import ConfigTester
30
from ...test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor
31
from ...test_pipeline_mixin import PipelineTesterMixin
32

33

34
if is_tf_available():
35
    import tensorflow as tf
36

37
    from transformers import TFResNetForImageClassification, TFResNetModel
38
    from transformers.models.resnet.modeling_tf_resnet import TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST
39

40

41
if is_vision_available():
42
    from PIL import Image
43

44
    from transformers import AutoImageProcessor
45

46

47
class TFResNetModelTester:
48
    def __init__(
49
        self,
50
        parent,
51
        batch_size=3,
52
        image_size=32,
53
        num_channels=3,
54
        embeddings_size=10,
55
        hidden_sizes=[10, 20, 30, 40],
56
        depths=[1, 1, 2, 1],
57
        is_training=True,
58
        use_labels=True,
59
        hidden_act="relu",
60
        num_labels=3,
61
        scope=None,
62
    ):
63
        self.parent = parent
64
        self.batch_size = batch_size
65
        self.image_size = image_size
66
        self.num_channels = num_channels
67
        self.embeddings_size = embeddings_size
68
        self.hidden_sizes = hidden_sizes
69
        self.depths = depths
70
        self.is_training = is_training
71
        self.use_labels = use_labels
72
        self.hidden_act = hidden_act
73
        self.num_labels = num_labels
74
        self.scope = scope
75
        self.num_stages = len(hidden_sizes)
76

77
    def prepare_config_and_inputs(self):
78
        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
79

80
        labels = None
81
        if self.use_labels:
82
            labels = ids_tensor([self.batch_size], self.num_labels)
83

84
        config = self.get_config()
85

86
        return config, pixel_values, labels
87

88
    def get_config(self):
89
        return ResNetConfig(
90
            num_channels=self.num_channels,
91
            embeddings_size=self.embeddings_size,
92
            hidden_sizes=self.hidden_sizes,
93
            depths=self.depths,
94
            hidden_act=self.hidden_act,
95
            num_labels=self.num_labels,
96
            image_size=self.image_size,
97
        )
98

99
    def create_and_check_model(self, config, pixel_values, labels):
100
        model = TFResNetModel(config=config)
101
        result = model(pixel_values)
102
        # expected last hidden states: B, C, H // 32, W // 32
103
        self.parent.assertEqual(
104
            result.last_hidden_state.shape,
105
            (self.batch_size, self.hidden_sizes[-1], self.image_size // 32, self.image_size // 32),
106
        )
107

108
    def create_and_check_for_image_classification(self, config, pixel_values, labels):
109
        config.num_labels = self.num_labels
110
        model = TFResNetForImageClassification(config)
111
        result = model(pixel_values, labels=labels)
112
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
113

114
    def prepare_config_and_inputs_for_common(self):
115
        config_and_inputs = self.prepare_config_and_inputs()
116
        config, pixel_values, labels = config_and_inputs
117
        inputs_dict = {"pixel_values": pixel_values}
118
        return config, inputs_dict
119

120

121
@require_tf
122
class TFResNetModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
123
    """
124
    Here we also overwrite some of the tests of test_modeling_common.py, as ResNet does not use input_ids, inputs_embeds,
125
    attention_mask and seq_length.
126
    """
127

128
    all_model_classes = (TFResNetModel, TFResNetForImageClassification) if is_tf_available() else ()
129
    pipeline_model_mapping = (
130
        {"feature-extraction": TFResNetModel, "image-classification": TFResNetForImageClassification}
131
        if is_tf_available()
132
        else {}
133
    )
134

135
    test_pruning = False
136
    test_resize_embeddings = False
137
    test_head_masking = False
138
    test_onnx = False
139
    has_attentions = False
140

141
    def setUp(self):
142
        self.model_tester = TFResNetModelTester(self)
143
        self.config_tester = ConfigTester(self, config_class=ResNetConfig, has_text_modality=False)
144

145
    def test_config(self):
146
        self.create_and_test_config_common_properties()
147
        self.config_tester.create_and_test_config_to_json_string()
148
        self.config_tester.create_and_test_config_to_json_file()
149
        self.config_tester.create_and_test_config_from_and_save_pretrained()
150
        self.config_tester.create_and_test_config_with_num_labels()
151
        self.config_tester.check_config_can_be_init_without_params()
152
        self.config_tester.check_config_arguments_init()
153

154
    def create_and_test_config_common_properties(self):
155
        return
156

157
    @unittest.skip(reason="ResNet does not use inputs_embeds")
158
    def test_inputs_embeds(self):
159
        pass
160

161
    @unittest.skip(reason="ResNet does not support input and output embeddings")
162
    def test_model_common_attributes(self):
163
        pass
164

165
    def test_forward_signature(self):
166
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
167

168
        for model_class in self.all_model_classes:
169
            model = model_class(config)
170
            signature = inspect.signature(model.call)
171
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
172
            arg_names = [*signature.parameters.keys()]
173

174
            expected_arg_names = ["pixel_values"]
175
            self.assertListEqual(arg_names[:1], expected_arg_names)
176

177
    def test_model(self):
178
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
179
        self.model_tester.create_and_check_model(*config_and_inputs)
180

181
    def test_hidden_states_output(self):
182
        def check_hidden_states_output(inputs_dict, config, model_class):
183
            model = model_class(config)
184
            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
185

186
            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
187

188
            expected_num_stages = self.model_tester.num_stages
189
            self.assertEqual(len(hidden_states), expected_num_stages + 1)
190

191
            # ResNet's feature maps are of shape (batch_size, num_channels, height, width)
192
            self.assertListEqual(
193
                list(hidden_states[0].shape[-2:]),
194
                [self.model_tester.image_size // 4, self.model_tester.image_size // 4],
195
            )
196

197
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
198
        layers_type = ["basic", "bottleneck"]
199
        for model_class in self.all_model_classes:
200
            for layer_type in layers_type:
201
                config.layer_type = layer_type
202
                inputs_dict["output_hidden_states"] = True
203
                check_hidden_states_output(inputs_dict, config, model_class)
204

205
                # check that output_hidden_states also work using config
206
                del inputs_dict["output_hidden_states"]
207
                config.output_hidden_states = True
208

209
                check_hidden_states_output(inputs_dict, config, model_class)
210

211
    def test_for_image_classification(self):
212
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
213
        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
214

215
    @slow
216
    def test_model_from_pretrained(self):
217
        for model_name in TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
218
            model = TFResNetModel.from_pretrained(model_name)
219
            self.assertIsNotNone(model)
220

221

222
# We will verify our results on an image of cute cats
223
def prepare_img():
224
    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
225
    return image
226

227

228
@require_tf
229
@require_vision
230
class TFResNetModelIntegrationTest(unittest.TestCase):
231
    @cached_property
232
    def default_image_processor(self):
233
        return (
234
            AutoImageProcessor.from_pretrained(TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST[0])
235
            if is_vision_available()
236
            else None
237
        )
238

239
    @slow
240
    def test_inference_image_classification_head(self):
241
        model = TFResNetForImageClassification.from_pretrained(TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST[0])
242

243
        image_processor = self.default_image_processor
244
        image = prepare_img()
245
        inputs = image_processor(images=image, return_tensors="tf")
246

247
        # forward pass
248
        outputs = model(**inputs)
249

250
        # verify the logits
251
        expected_shape = tf.TensorShape((1, 1000))
252
        self.assertEqual(outputs.logits.shape, expected_shape)
253

254
        expected_slice = tf.constant([-11.1069, -9.7877, -8.3777])
255

256
        self.assertTrue(np.allclose(outputs.logits[0, :3].numpy(), expected_slice, atol=1e-4))
257