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 PyTorch Nat model. """
16

17
import collections
18
import unittest
19

20
from transformers import NatConfig
21
from transformers.testing_utils import require_natten, require_torch, require_vision, slow, torch_device
22
from transformers.utils import cached_property, is_torch_available, is_vision_available
23

24
from ...test_backbone_common import BackboneTesterMixin
25
from ...test_configuration_common import ConfigTester
26
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
27
from ...test_pipeline_mixin import PipelineTesterMixin
28

29

30
if is_torch_available():
31
    import torch
32
    from torch import nn
33

34
    from transformers import NatBackbone, NatForImageClassification, NatModel
35
    from transformers.models.nat.modeling_nat import NAT_PRETRAINED_MODEL_ARCHIVE_LIST
36

37
if is_vision_available():
38
    from PIL import Image
39

40
    from transformers import AutoImageProcessor
41

42

43
class NatModelTester:
44
    def __init__(
45
        self,
46
        parent,
47
        batch_size=13,
48
        image_size=64,
49
        patch_size=4,
50
        num_channels=3,
51
        embed_dim=16,
52
        depths=[1, 2, 1],
53
        num_heads=[2, 4, 8],
54
        kernel_size=3,
55
        mlp_ratio=2.0,
56
        qkv_bias=True,
57
        hidden_dropout_prob=0.0,
58
        attention_probs_dropout_prob=0.0,
59
        drop_path_rate=0.1,
60
        hidden_act="gelu",
61
        patch_norm=True,
62
        initializer_range=0.02,
63
        layer_norm_eps=1e-5,
64
        is_training=True,
65
        scope=None,
66
        use_labels=True,
67
        num_labels=10,
68
        out_features=["stage1", "stage2"],
69
        out_indices=[1, 2],
70
    ):
71
        self.parent = parent
72
        self.batch_size = batch_size
73
        self.image_size = image_size
74
        self.patch_size = patch_size
75
        self.num_channels = num_channels
76
        self.embed_dim = embed_dim
77
        self.depths = depths
78
        self.num_heads = num_heads
79
        self.kernel_size = kernel_size
80
        self.mlp_ratio = mlp_ratio
81
        self.qkv_bias = qkv_bias
82
        self.hidden_dropout_prob = hidden_dropout_prob
83
        self.attention_probs_dropout_prob = attention_probs_dropout_prob
84
        self.drop_path_rate = drop_path_rate
85
        self.hidden_act = hidden_act
86
        self.patch_norm = patch_norm
87
        self.layer_norm_eps = layer_norm_eps
88
        self.initializer_range = initializer_range
89
        self.is_training = is_training
90
        self.scope = scope
91
        self.use_labels = use_labels
92
        self.num_labels = num_labels
93
        self.out_features = out_features
94
        self.out_indices = out_indices
95

96
    def prepare_config_and_inputs(self):
97
        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
98

99
        labels = None
100
        if self.use_labels:
101
            labels = ids_tensor([self.batch_size], self.num_labels)
102

103
        config = self.get_config()
104

105
        return config, pixel_values, labels
106

107
    def get_config(self):
108
        return NatConfig(
109
            num_labels=self.num_labels,
110
            image_size=self.image_size,
111
            patch_size=self.patch_size,
112
            num_channels=self.num_channels,
113
            embed_dim=self.embed_dim,
114
            depths=self.depths,
115
            num_heads=self.num_heads,
116
            kernel_size=self.kernel_size,
117
            mlp_ratio=self.mlp_ratio,
118
            qkv_bias=self.qkv_bias,
119
            hidden_dropout_prob=self.hidden_dropout_prob,
120
            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
121
            drop_path_rate=self.drop_path_rate,
122
            hidden_act=self.hidden_act,
123
            patch_norm=self.patch_norm,
124
            layer_norm_eps=self.layer_norm_eps,
125
            initializer_range=self.initializer_range,
126
            out_features=self.out_features,
127
            out_indices=self.out_indices,
128
        )
129

130
    def create_and_check_model(self, config, pixel_values, labels):
131
        model = NatModel(config=config)
132
        model.to(torch_device)
133
        model.eval()
134
        result = model(pixel_values)
135

136
        expected_height = expected_width = (config.image_size // config.patch_size) // (2 ** (len(config.depths) - 1))
137
        expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
138

139
        self.parent.assertEqual(
140
            result.last_hidden_state.shape, (self.batch_size, expected_height, expected_width, expected_dim)
141
        )
142

143
    def create_and_check_for_image_classification(self, config, pixel_values, labels):
144
        model = NatForImageClassification(config)
145
        model.to(torch_device)
146
        model.eval()
147
        result = model(pixel_values, labels=labels)
148
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
149

150
        # test greyscale images
151
        config.num_channels = 1
152
        model = NatForImageClassification(config)
153
        model.to(torch_device)
154
        model.eval()
155

156
        pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
157
        result = model(pixel_values)
158
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
159

160
    def create_and_check_backbone(self, config, pixel_values, labels):
161
        model = NatBackbone(config=config)
162
        model.to(torch_device)
163
        model.eval()
164
        result = model(pixel_values)
165

166
        # verify hidden states
167
        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
168
        self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, model.channels[0], 16, 16])
169

170
        # verify channels
171
        self.parent.assertEqual(len(model.channels), len(config.out_features))
172

173
        # verify backbone works with out_features=None
174
        config.out_features = None
175
        model = NatBackbone(config=config)
176
        model.to(torch_device)
177
        model.eval()
178
        result = model(pixel_values)
179

180
        # verify feature maps
181
        self.parent.assertEqual(len(result.feature_maps), 1)
182
        self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, model.channels[-1], 4, 4])
183

184
        # verify channels
185
        self.parent.assertEqual(len(model.channels), 1)
186

187
    def prepare_config_and_inputs_for_common(self):
188
        config_and_inputs = self.prepare_config_and_inputs()
189
        config, pixel_values, labels = config_and_inputs
190
        inputs_dict = {"pixel_values": pixel_values}
191
        return config, inputs_dict
192

193

194
@require_natten
195
@require_torch
196
class NatModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
197
    all_model_classes = (
198
        (
199
            NatModel,
200
            NatForImageClassification,
201
            NatBackbone,
202
        )
203
        if is_torch_available()
204
        else ()
205
    )
206
    pipeline_model_mapping = (
207
        {"image-feature-extraction": NatModel, "image-classification": NatForImageClassification}
208
        if is_torch_available()
209
        else {}
210
    )
211
    fx_compatible = False
212

213
    test_torchscript = False
214
    test_pruning = False
215
    test_resize_embeddings = False
216
    test_head_masking = False
217

218
    def setUp(self):
219
        self.model_tester = NatModelTester(self)
220
        self.config_tester = ConfigTester(self, config_class=NatConfig, embed_dim=37)
221

222
    def test_config(self):
223
        self.create_and_test_config_common_properties()
224
        self.config_tester.create_and_test_config_to_json_string()
225
        self.config_tester.create_and_test_config_to_json_file()
226
        self.config_tester.create_and_test_config_from_and_save_pretrained()
227
        self.config_tester.create_and_test_config_with_num_labels()
228
        self.config_tester.check_config_can_be_init_without_params()
229
        self.config_tester.check_config_arguments_init()
230

231
    def create_and_test_config_common_properties(self):
232
        return
233

234
    def test_model(self):
235
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
236
        self.model_tester.create_and_check_model(*config_and_inputs)
237

238
    def test_for_image_classification(self):
239
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
240
        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
241

242
    def test_backbone(self):
243
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
244
        self.model_tester.create_and_check_backbone(*config_and_inputs)
245

246
    @unittest.skip(reason="Nat does not use inputs_embeds")
247
    def test_inputs_embeds(self):
248
        pass
249

250
    @unittest.skip(reason="Nat does not use feedforward chunking")
251
    def test_feed_forward_chunking(self):
252
        pass
253

254
    def test_model_common_attributes(self):
255
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
256

257
        for model_class in self.all_model_classes:
258
            model = model_class(config)
259
            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
260
            x = model.get_output_embeddings()
261
            self.assertTrue(x is None or isinstance(x, nn.Linear))
262

263
    def test_attention_outputs(self):
264
        self.skipTest("Nat's attention operation is handled entirely by NATTEN.")
265

266
    def check_hidden_states_output(self, inputs_dict, config, model_class, image_size):
267
        model = model_class(config)
268
        model.to(torch_device)
269
        model.eval()
270

271
        with torch.no_grad():
272
            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
273

274
        hidden_states = outputs.hidden_states
275

276
        expected_num_layers = getattr(
277
            self.model_tester, "expected_num_hidden_layers", len(self.model_tester.depths) + 1
278
        )
279
        self.assertEqual(len(hidden_states), expected_num_layers)
280

281
        # Nat has a different seq_length
282
        patch_size = (
283
            config.patch_size
284
            if isinstance(config.patch_size, collections.abc.Iterable)
285
            else (config.patch_size, config.patch_size)
286
        )
287

288
        height = image_size[0] // patch_size[0]
289
        width = image_size[1] // patch_size[1]
290

291
        self.assertListEqual(
292
            list(hidden_states[0].shape[-3:]),
293
            [height, width, self.model_tester.embed_dim],
294
        )
295

296
        if model_class.__name__ != "NatBackbone":
297
            reshaped_hidden_states = outputs.reshaped_hidden_states
298
            self.assertEqual(len(reshaped_hidden_states), expected_num_layers)
299

300
            batch_size, num_channels, height, width = reshaped_hidden_states[0].shape
301
            reshaped_hidden_states = (
302
                reshaped_hidden_states[0].view(batch_size, num_channels, height, width).permute(0, 2, 3, 1)
303
            )
304
            self.assertListEqual(
305
                list(reshaped_hidden_states.shape[-3:]),
306
                [height, width, self.model_tester.embed_dim],
307
            )
308

309
    def test_hidden_states_output(self):
310
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
311

312
        image_size = (
313
            self.model_tester.image_size
314
            if isinstance(self.model_tester.image_size, collections.abc.Iterable)
315
            else (self.model_tester.image_size, self.model_tester.image_size)
316
        )
317

318
        for model_class in self.all_model_classes:
319
            inputs_dict["output_hidden_states"] = True
320
            self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
321

322
            # check that output_hidden_states also work using config
323
            del inputs_dict["output_hidden_states"]
324
            config.output_hidden_states = True
325

326
            self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
327

328
    @slow
329
    def test_model_from_pretrained(self):
330
        for model_name in NAT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
331
            model = NatModel.from_pretrained(model_name)
332
            self.assertIsNotNone(model)
333

334
    def test_initialization(self):
335
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
336

337
        configs_no_init = _config_zero_init(config)
338
        for model_class in self.all_model_classes:
339
            model = model_class(config=configs_no_init)
340
            for name, param in model.named_parameters():
341
                if "embeddings" not in name and param.requires_grad:
342
                    self.assertIn(
343
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
344
                        [0.0, 1.0],
345
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
346
                    )
347

348

349
@require_natten
350
@require_vision
351
@require_torch
352
class NatModelIntegrationTest(unittest.TestCase):
353
    @cached_property
354
    def default_image_processor(self):
355
        return AutoImageProcessor.from_pretrained("shi-labs/nat-mini-in1k-224") if is_vision_available() else None
356

357
    @slow
358
    def test_inference_image_classification_head(self):
359
        model = NatForImageClassification.from_pretrained("shi-labs/nat-mini-in1k-224").to(torch_device)
360
        image_processor = self.default_image_processor
361

362
        image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
363
        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
364

365
        # forward pass
366
        with torch.no_grad():
367
            outputs = model(**inputs)
368

369
        # verify the logits
370
        expected_shape = torch.Size((1, 1000))
371
        self.assertEqual(outputs.logits.shape, expected_shape)
372
        expected_slice = torch.tensor([0.3805, -0.8676, -0.3912]).to(torch_device)
373
        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
374

375

376
@require_torch
377
@require_natten
378
class NatBackboneTest(unittest.TestCase, BackboneTesterMixin):
379
    all_model_classes = (NatBackbone,) if is_torch_available() else ()
380
    config_class = NatConfig
381

382
    def setUp(self):
383
        self.model_tester = NatModelTester(self)
384