File size: 12,074 Bytes
455a40f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
#
# 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.
""" Testing suite for the PyTorch UperNet framework. """


import inspect
import unittest

from huggingface_hub import hf_hub_download

from transformers import ConvNextConfig, UperNetConfig
from transformers.testing_utils import require_torch, require_torch_multi_gpu, require_vision, slow, torch_device
from transformers.utils import is_torch_available, is_vision_available

from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin


if is_torch_available():
    import torch

    from transformers import UperNetForSemanticSegmentation
    from transformers.models.upernet.modeling_upernet import UPERNET_PRETRAINED_MODEL_ARCHIVE_LIST


if is_vision_available():
    from PIL import Image

    from transformers import AutoImageProcessor


class UperNetModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        image_size=32,
        num_channels=3,
        num_stages=4,
        hidden_sizes=[10, 20, 30, 40],
        depths=[2, 2, 3, 2],
        is_training=True,
        use_labels=True,
        intermediate_size=37,
        hidden_act="gelu",
        type_sequence_label_size=10,
        initializer_range=0.02,
        out_features=["stage2", "stage3", "stage4"],
        num_labels=3,
        scope=None,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.image_size = image_size
        self.num_channels = num_channels
        self.num_stages = num_stages
        self.hidden_sizes = hidden_sizes
        self.depths = depths
        self.is_training = is_training
        self.use_labels = use_labels
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.type_sequence_label_size = type_sequence_label_size
        self.initializer_range = initializer_range
        self.out_features = out_features
        self.num_labels = num_labels
        self.scope = scope
        self.num_hidden_layers = num_stages

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

        labels = None
        if self.use_labels:
            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)

        config = self.get_config()

        return config, pixel_values, labels

    def get_backbone_config(self):
        return ConvNextConfig(
            num_channels=self.num_channels,
            num_stages=self.num_stages,
            hidden_sizes=self.hidden_sizes,
            depths=self.depths,
            is_training=self.is_training,
            intermediate_size=self.intermediate_size,
            hidden_act=self.hidden_act,
            out_features=self.out_features,
        )

    def get_config(self):
        return UperNetConfig(
            backbone_config=self.get_backbone_config(),
            hidden_size=512,
            pool_scales=[1, 2, 3, 6],
            use_auxiliary_head=True,
            auxiliary_loss_weight=0.4,
            auxiliary_in_channels=40,
            auxiliary_channels=256,
            auxiliary_num_convs=1,
            auxiliary_concat_input=False,
            loss_ignore_index=255,
            num_labels=self.num_labels,
        )

    def create_and_check_for_semantic_segmentation(self, config, pixel_values, labels):
        model = UperNetForSemanticSegmentation(config=config)
        model.to(torch_device)
        model.eval()
        result = model(pixel_values)
        self.parent.assertEqual(
            result.logits.shape, (self.batch_size, self.num_labels, self.image_size, self.image_size)
        )

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        (
            config,
            pixel_values,
            labels,
        ) = config_and_inputs
        inputs_dict = {"pixel_values": pixel_values}
        return config, inputs_dict


@require_torch
class UperNetModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
    """
    Here we also overwrite some of the tests of test_modeling_common.py, as UperNet does not use input_ids, inputs_embeds,
    attention_mask and seq_length.
    """

    all_model_classes = (UperNetForSemanticSegmentation,) if is_torch_available() else ()
    pipeline_model_mapping = {"image-segmentation": UperNetForSemanticSegmentation} if is_torch_available() else {}
    fx_compatible = False
    test_pruning = False
    test_resize_embeddings = False
    test_head_masking = False
    test_torchscript = False
    has_attentions = False

    def setUp(self):
        self.model_tester = UperNetModelTester(self)
        self.config_tester = ConfigTester(self, config_class=UperNetConfig, has_text_modality=False, hidden_size=37)

    def test_config(self):
        self.create_and_test_config_common_properties()
        self.config_tester.create_and_test_config_to_json_string()
        self.config_tester.create_and_test_config_to_json_file()
        self.config_tester.create_and_test_config_from_and_save_pretrained()
        self.config_tester.create_and_test_config_with_num_labels()
        self.config_tester.check_config_can_be_init_without_params()
        self.config_tester.check_config_arguments_init()

    def create_and_test_config_common_properties(self):
        return

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

        for model_class in self.all_model_classes:
            model = model_class(config)
            signature = inspect.signature(model.forward)
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
            arg_names = [*signature.parameters.keys()]

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

    def test_for_semantic_segmentation(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_for_semantic_segmentation(*config_and_inputs)

    @unittest.skip(reason="UperNet does not use inputs_embeds")
    def test_inputs_embeds(self):
        pass

    @unittest.skip(reason="UperNet does not support input and output embeddings")
    def test_model_common_attributes(self):
        pass

    @unittest.skip(reason="UperNet does not have a base model")
    def test_save_load_fast_init_from_base(self):
        pass

    @unittest.skip(reason="UperNet does not have a base model")
    def test_save_load_fast_init_to_base(self):
        pass

    @require_torch_multi_gpu
    @unittest.skip(reason="UperNet has some layers using `add_module` which doesn't work well with `nn.DataParallel`")
    def test_multi_gpu_data_parallel_forward(self):
        pass

    def test_hidden_states_output(self):
        def check_hidden_states_output(inputs_dict, config, model_class):
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

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

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

            # ConvNext's feature maps are of shape (batch_size, num_channels, height, width)
            self.assertListEqual(
                list(hidden_states[0].shape[-2:]),
                [self.model_tester.image_size // 4, self.model_tester.image_size // 4],
            )

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            inputs_dict["output_hidden_states"] = True
            check_hidden_states_output(inputs_dict, config, model_class)

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

            check_hidden_states_output(inputs_dict, config, model_class)

    def test_initialization(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        configs_no_init = _config_zero_init(config)
        configs_no_init.backbone_config = _config_zero_init(configs_no_init.backbone_config)
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            for name, param in model.named_parameters():
                if param.requires_grad:
                    self.assertIn(
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
                        [0.0, 1.0],
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
                    )

    @unittest.skip(reason="UperNet does not have tied weights")
    def test_tied_model_weights_key_ignore(self):
        pass

    @slow
    def test_model_from_pretrained(self):
        for model_name in UPERNET_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = UperNetForSemanticSegmentation.from_pretrained(model_name)
            self.assertIsNotNone(model)


# We will verify our results on an image of ADE20k
def prepare_img():
    filepath = hf_hub_download(
        repo_id="hf-internal-testing/fixtures_ade20k", repo_type="dataset", filename="ADE_val_00000001.jpg"
    )
    image = Image.open(filepath).convert("RGB")
    return image


@require_torch
@require_vision
@slow
class UperNetModelIntegrationTest(unittest.TestCase):
    def test_inference_swin_backbone(self):
        processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-swin-tiny")
        model = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-swin-tiny").to(torch_device)

        image = prepare_img()
        inputs = processor(images=image, return_tensors="pt").to(torch_device)

        with torch.no_grad():
            outputs = model(**inputs)

        expected_shape = torch.Size((1, model.config.num_labels, 512, 512))
        self.assertEqual(outputs.logits.shape, expected_shape)

        expected_slice = torch.tensor(
            [[-7.5958, -7.5958, -7.4302], [-7.5958, -7.5958, -7.4302], [-7.4797, -7.4797, -7.3068]]
        ).to(torch_device)
        self.assertTrue(torch.allclose(outputs.logits[0, 0, :3, :3], expected_slice, atol=1e-4))

    def test_inference_convnext_backbone(self):
        processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-convnext-tiny")
        model = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-convnext-tiny").to(torch_device)

        image = prepare_img()
        inputs = processor(images=image, return_tensors="pt").to(torch_device)

        with torch.no_grad():
            outputs = model(**inputs)

        expected_shape = torch.Size((1, model.config.num_labels, 512, 512))
        self.assertEqual(outputs.logits.shape, expected_shape)

        expected_slice = torch.tensor(
            [[-8.8110, -8.8110, -8.6521], [-8.8110, -8.8110, -8.6521], [-8.7746, -8.7746, -8.6130]]
        ).to(torch_device)
        self.assertTrue(torch.allclose(outputs.logits[0, 0, :3, :3], expected_slice, atol=1e-4))