File size: 15,758 Bytes
a9bd396
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
<!--Copyright 2022 The HuggingFace 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.

โš ๏ธ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.

-->

# ์˜๋ฏธ์  ๋ถ„ํ• (Semantic segmentation)[[semantic-segmentation]]

[[open-in-colab]]

<Youtube id="dKE8SIt9C-w"/>

์˜๋ฏธ์  ๋ถ„ํ• (semantic segmentation)์€ ์ด๋ฏธ์ง€์˜ ๊ฐ ํ”ฝ์…€์— ๋ ˆ์ด๋ธ” ๋˜๋Š” ํด๋ž˜์Šค๋ฅผ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋ถ„ํ• (segmentation)์—๋Š” ์—ฌ๋Ÿฌ ์ข…๋ฅ˜๊ฐ€ ์žˆ์œผ๋ฉฐ, ์˜๋ฏธ์  ๋ถ„ํ• ์˜ ๊ฒฝ์šฐ ๋™์ผํ•œ ๋ฌผ์ฒด์˜ ๊ณ ์œ  ์ธ์Šคํ„ด์Šค๋ฅผ ๊ตฌ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฌผ์ฒด ๋ชจ๋‘ ๋™์ผํ•œ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋ฉ๋‹ˆ๋‹ค(์˜ˆ์‹œ๋กœ, "car-1" ๊ณผ "car-2" ๋Œ€์‹  "car"๋กœ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค).
์‹ค์ƒํ™œ์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ๋Š” ์˜๋ฏธ์  ๋ถ„ํ• ์˜ ์ ์šฉ ์‚ฌ๋ก€๋กœ๋Š” ๋ณดํ–‰์ž์™€ ์ค‘์š”ํ•œ ๊ตํ†ต ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•˜๋Š” ์ž์œจ ์ฃผํ–‰ ์ž๋™์ฐจ ํ•™์Šต, ์˜๋ฃŒ ์ด๋ฏธ์ง€์˜ ์„ธํฌ์™€ ์ด์ƒ ์ง•ํ›„ ์‹๋ณ„, ๊ทธ๋ฆฌ๊ณ  ์œ„์„ฑ ์ด๋ฏธ์ง€์˜ ํ™˜๊ฒฝ ๋ณ€ํ™” ๋ชจ๋‹ˆํ„ฐ๋ง๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

์ด๋ฒˆ ๊ฐ€์ด๋“œ์—์„œ ๋ฐฐ์šธ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

1. [SceneParse150](https://huggingface.co/datasets/scene_parse_150) ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ด์šฉํ•ด [SegFormer](https://huggingface.co/docs/transformers/main/en/model_doc/segformer#segformer) ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ธฐ.
2. ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์„ ์ถ”๋ก ์— ์‚ฌ์šฉํ•˜๊ธฐ.

<Tip>

์ด ์ž‘์—…๊ณผ ํ˜ธํ™˜๋˜๋Š” ๋ชจ๋“  ์•„ํ‚คํ…์ฒ˜์™€ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋ณด๋ ค๋ฉด [์ž‘์—… ํŽ˜์ด์ง€](https://huggingface.co/tasks/image-segmentation)๋ฅผ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.

</Tip>

์‹œ์ž‘ํ•˜๊ธฐ ์ „์— ํ•„์š”ํ•œ ๋ชจ๋“  ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๊ฐ€ ์„ค์น˜๋˜์—ˆ๋Š”์ง€ ํ™•์ธํ•˜์„ธ์š”:

```bash
pip install -q datasets transformers evaluate
```
์ปค๋ฎค๋‹ˆํ‹ฐ์— ๋ชจ๋ธ์„ ์—…๋กœ๋“œํ•˜๊ณ  ๊ณต์œ ํ•  ์ˆ˜ ์žˆ๋„๋ก Hugging Face ๊ณ„์ •์— ๋กœ๊ทธ์ธํ•˜๋Š” ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๊ฐ€ ๋‚˜ํƒ€๋‚˜๋ฉด ํ† ํฐ์„ ์ž…๋ ฅํ•˜์—ฌ ๋กœ๊ทธ์ธํ•˜์„ธ์š”:

```py
>>> from huggingface_hub import notebook_login

>>> notebook_login()
```

## SceneParse150 ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ถˆ๋Ÿฌ์˜ค๊ธฐ[[load-sceneparse150-dataset]]

๐Ÿค— Datasets ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ SceneParse150 ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋” ์ž‘์€ ๋ถ€๋ถ„ ์ง‘ํ•ฉ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ „์ฒด์— ๋Œ€ํ•œ ํ›ˆ๋ จ์— ๋งŽ์€ ์‹œ๊ฐ„์„ ํ• ์• ํ•˜๊ธฐ ์ „์— ์‹คํ—˜์„ ํ†ตํ•ด ๋ชจ๋“  ๊ฒƒ์ด ์ œ๋Œ€๋กœ ์ž‘๋™ํ•˜๋Š”์ง€ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

```py
>>> from datasets import load_dataset

>>> ds = load_dataset("scene_parse_150", split="train[:50]")
```

๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ `train`์„ [`~datasets.Dataset.train_test_split`] ๋ฉ”์†Œ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ๋ฐ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ๋ถ„ํ• ํ•˜์„ธ์š”:

```py
>>> ds = ds.train_test_split(test_size=0.2)
>>> train_ds = ds["train"]
>>> test_ds = ds["test"]
```

๊ทธ๋ฆฌ๊ณ  ์˜ˆ์‹œ๋ฅผ ์‚ดํŽด๋ณด์„ธ์š”:

```py
>>> train_ds[0]
{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=512x683 at 0x7F9B0C201F90>,
 'annotation': <PIL.PngImagePlugin.PngImageFile image mode=L size=512x683 at 0x7F9B0C201DD0>,
 'scene_category': 368}
```

- `image`: ์žฅ๋ฉด์˜ PIL ์ด๋ฏธ์ง€์ž…๋‹ˆ๋‹ค.
- `annotation`: ๋ถ„ํ•  ์ง€๋„(segmentation map)์˜ PIL ์ด๋ฏธ์ง€์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ํƒ€๊ฒŸ์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค.
- `scene_category`: "์ฃผ๋ฐฉ" ๋˜๋Š” "์‚ฌ๋ฌด์‹ค"๊ณผ ๊ฐ™์ด ์ด๋ฏธ์ง€ ์žฅ๋ฉด์„ ์„ค๋ช…ํ•˜๋Š” ์นดํ…Œ๊ณ ๋ฆฌ ID์ž…๋‹ˆ๋‹ค. ์ด ๊ฐ€์ด๋“œ์—์„œ๋Š” ๋‘˜ ๋‹ค PIL ์ด๋ฏธ์ง€์ธ `image`์™€ `annotation`๋งŒ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.

๋‚˜์ค‘์— ๋ชจ๋ธ์„ ์„ค์ •ํ•  ๋•Œ ์œ ์šฉํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ ˆ์ด๋ธ” ID๋ฅผ ๋ ˆ์ด๋ธ” ํด๋ž˜์Šค์— ๋งคํ•‘ํ•˜๋Š” ์‚ฌ์ „๋„ ๋งŒ๋“ค๊ณ  ์‹ถ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. Hub์—์„œ ๋งคํ•‘์„ ๋‹ค์šด๋กœ๋“œํ•˜๊ณ  `id2label` ๋ฐ `label2id` ์‚ฌ์ „์„ ๋งŒ๋“œ์„ธ์š”:

```py
>>> import json
>>> from pathlib import Path
>>> from huggingface_hub import hf_hub_download

>>> repo_id = "huggingface/label-files"
>>> filename = "ade20k-id2label.json"
>>> id2label = json.loads(Path(hf_hub_download(repo_id, filename, repo_type="dataset")).read_text())
>>> id2label = {int(k): v for k, v in id2label.items()}
>>> label2id = {v: k for k, v in id2label.items()}
>>> num_labels = len(id2label)
```

## ์ „์ฒ˜๋ฆฌํ•˜๊ธฐ[[preprocess]

๋‹ค์Œ ๋‹จ๊ณ„๋Š” ๋ชจ๋ธ์— ์‚ฌ์šฉํ•  ์ด๋ฏธ์ง€์™€ ์ฃผ์„์„ ์ค€๋น„ํ•˜๊ธฐ ์œ„ํ•ด SegFormer ์ด๋ฏธ์ง€ ํ”„๋กœ์„ธ์„œ๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฐ™์€ ์ผ๋ถ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ฐฐ๊ฒฝ ํด๋ž˜์Šค๋กœ ์ œ๋กœ ์ธ๋ฑ์Šค๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ฐฐ๊ฒฝ ํด๋ž˜์Šค๋Š” 150๊ฐœ์˜ ํด๋ž˜์Šค์— ์‹ค์ œ๋กœ๋Š” ํฌํ•จ๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— `do_reduce_labels=True` ๋ฅผ ์„ค์ •ํ•ด ๋ชจ๋“  ๋ ˆ์ด๋ธ”์—์„œ ๋ฐฐ๊ฒฝ ํด๋ž˜์Šค๋ฅผ ์ œ๊ฑฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ œ๋กœ ์ธ๋ฑ์Šค๋Š” `255`๋กœ ๋Œ€์ฒด๋˜๋ฏ€๋กœ SegFormer์˜ ์†์‹ค ํ•จ์ˆ˜์—์„œ ๋ฌด์‹œ๋ฉ๋‹ˆ๋‹ค:

```py
>>> from transformers import AutoImageProcessor

>>> checkpoint = "nvidia/mit-b0"
>>> image_processor = AutoImageProcessor.from_pretrained(checkpoint, do_reduce_labels=True)
```


์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ์ ์šฉํ•˜์—ฌ ๊ณผ์ ํ•ฉ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๋ณด๋‹ค ๊ฐ•๊ฑดํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ์ด ๊ฐ€์ด๋“œ์—์„œ๋Š” [torchvision](https://pytorch.org/vision/stable/index.html)์˜ [`ColorJitter`](https://pytorch.org/vision/stable/generated/torchvision.transforms.ColorJitter.html)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€์˜ ์ƒ‰์ƒ ์†์„ฑ์„ ์ž„์˜๋กœ ๋ณ€๊ฒฝํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ, ์ž์‹ ์ด ์›ํ•˜๋Š” ์ด๋ฏธ์ง€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.

```py
>>> from torchvision.transforms import ColorJitter

>>> jitter = ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.1)
```

์ด์ œ ๋ชจ๋ธ์— ์‚ฌ์šฉํ•  ์ด๋ฏธ์ง€์™€ ์ฃผ์„์„ ์ค€๋น„ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐœ์˜ ์ „์ฒ˜๋ฆฌ ํ•จ์ˆ˜๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ด ํ•จ์ˆ˜๋“ค์€ ์ด๋ฏธ์ง€๋ฅผ `pixel_values`๋กœ, ์ฃผ์„์„ `labels`๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์„ธํŠธ์˜ ๊ฒฝ์šฐ ์ด๋ฏธ์ง€ ํ”„๋กœ์„ธ์„œ์— ์ด๋ฏธ์ง€๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์ „์— `jitter`๋ฅผ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๊ฒฝ์šฐ ์ด๋ฏธ์ง€ ํ”„๋กœ์„ธ์„œ๋Š” `images`๋ฅผ ์ž๋ฅด๊ณ  ์ •๊ทœํ™”ํ•˜๋ฉฐ, ํ…Œ์ŠคํŠธ ์ค‘์—๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์ด ์ ์šฉ๋˜์ง€ ์•Š์œผ๋ฏ€๋กœ `labels`๋งŒ ์ž๋ฆ…๋‹ˆ๋‹ค.

```py
>>> def train_transforms(example_batch):
...     images = [jitter(x) for x in example_batch["image"]]
...     labels = [x for x in example_batch["annotation"]]
...     inputs = image_processor(images, labels)
...     return inputs


>>> def val_transforms(example_batch):
...     images = [x for x in example_batch["image"]]
...     labels = [x for x in example_batch["annotation"]]
...     inputs = image_processor(images, labels)
...     return inputs
```

๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— `jitter`๋ฅผ ์ ์šฉํ•˜๋ ค๋ฉด, ๐Ÿค— Datasets [`~datasets.Dataset.set_transform`] ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์„ธ์š”. ์ฆ‰์‹œ ๋ณ€ํ™˜์ด ์ ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์— ๋” ๋น ๋ฅด๊ณ  ๋””์Šคํฌ ๊ณต๊ฐ„์„ ๋œ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค:

```py
>>> train_ds.set_transform(train_transforms)
>>> test_ds.set_transform(val_transforms)
```


## ํ‰๊ฐ€ํ•˜๊ธฐ[[evaluate]]

ํ›ˆ๋ จ ์ค‘์— ๋ฉ”ํŠธ๋ฆญ์„ ํฌํ•จํ•˜๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๐Ÿค— [Evaluate](https://huggingface.co/docs/evaluate/index) ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ๋น ๋ฅด๊ฒŒ ๋กœ๋“œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํƒœ์Šคํฌ์—์„œ๋Š” [mean Intersection over Union](https://huggingface.co/spaces/evaluate-metric/accuracy) (IoU) ๋ฉ”ํŠธ๋ฆญ์„ ๋กœ๋“œํ•˜์„ธ์š” (๋ฉ”ํŠธ๋ฆญ์„ ๋กœ๋“œํ•˜๊ณ  ๊ณ„์‚ฐํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์•Œ์•„๋ณด๋ ค๋ฉด ๐Ÿค— Evaluate [quick tour](https://huggingface.co/docs/evaluate/a_quick_tour)๋ฅผ ์‚ดํŽด๋ณด์„ธ์š”).

```py
>>> import evaluate

>>> metric = evaluate.load("mean_iou")
```

๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ฉ”ํŠธ๋ฆญ์„ [`~evaluate.EvaluationModule.compute`]ํ•˜๋Š” ํ•จ์ˆ˜๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์˜ˆ์ธก์„ ๋จผ์ € ๋กœ์ง“์œผ๋กœ ๋ณ€ํ™˜ํ•œ ๋‹ค์Œ, ๋ ˆ์ด๋ธ”์˜ ํฌ๊ธฐ์— ๋งž๊ฒŒ ๋ชจ์–‘์„ ๋‹ค์‹œ ์ง€์ •ํ•ด์•ผ [`~evaluate.EvaluationModule.compute`]๋ฅผ ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:


```py
>>> import numpy as np
>>> import torch
>>> from torch import nn

>>> def compute_metrics(eval_pred):
...     with torch.no_grad():
...         logits, labels = eval_pred
...         logits_tensor = torch.from_numpy(logits)
...         logits_tensor = nn.functional.interpolate(
...             logits_tensor,
...             size=labels.shape[-2:],
...             mode="bilinear",
...             align_corners=False,
...         ).argmax(dim=1)

...         pred_labels = logits_tensor.detach().cpu().numpy()
...         metrics = metric.compute(
...             predictions=pred_labels,
...             references=labels,
...             num_labels=num_labels,
...             ignore_index=255,
...             reduce_labels=False,
...         )
...         for key, value in metrics.items():
...             if isinstance(value, np.ndarray):
...                 metrics[key] = value.tolist()
...         return metrics
```



์ด์ œ `compute_metrics` ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•  ์ค€๋น„๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํŠธ๋ ˆ์ด๋‹์„ ์„ค์ •ํ•  ๋•Œ ์ด ํ•จ์ˆ˜๋กœ ๋Œ์•„๊ฐ€๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

## ํ•™์Šตํ•˜๊ธฐ[[train]]
<Tip>

๋งŒ์•ฝ [`Trainer`]๋ฅผ ์‚ฌ์šฉํ•ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์— ์ต์ˆ™ํ•˜์ง€ ์•Š๋‹ค๋ฉด, [์—ฌ๊ธฐ](../training#finetune-with-trainer)์—์„œ ๊ธฐ๋ณธ ํŠœํ† ๋ฆฌ์–ผ์„ ์‚ดํŽด๋ณด์„ธ์š”!

</Tip>

์ด์ œ ๋ชจ๋ธ ํ•™์Šต์„ ์‹œ์ž‘ํ•  ์ค€๋น„๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค! [`AutoModelForSemanticSegmentation`]๋กœ SegFormer๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๊ณ , ๋ชจ๋ธ์— ๋ ˆ์ด๋ธ” ID์™€ ๋ ˆ์ด๋ธ” ํด๋ž˜์Šค ๊ฐ„์˜ ๋งคํ•‘์„ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค:

```py
>>> from transformers import AutoModelForSemanticSegmentation, TrainingArguments, Trainer

>>> model = AutoModelForSemanticSegmentation.from_pretrained(checkpoint, id2label=id2label, label2id=label2id)
```

์ด์ œ ์„ธ ๋‹จ๊ณ„๋งŒ ๋‚จ์•˜์Šต๋‹ˆ๋‹ค:

1. ํ•™์Šต ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ [`TrainingArguments`]์— ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. `image` ์—ด์ด ์‚ญ์ œ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉํ•˜์ง€ ์•Š๋Š” ์—ด์„ ์ œ๊ฑฐํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. `image` ์—ด์ด ์—†์œผ๋ฉด `pixel_values`์„ ์ƒ์„ฑํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฐ ๊ฒฝ์šฐ๋ฅผ ๋ฐฉ์ง€ํ•˜๋ ค๋ฉด `remove_unused_columns=False`๋กœ ์„ค์ •ํ•˜์„ธ์š”! ์œ ์ผํ•˜๊ฒŒ ํ•„์š”ํ•œ ๋‹ค๋ฅธ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ๋ชจ๋ธ์„ ์ €์žฅํ•  ์œ„์น˜๋ฅผ ์ง€์ •ํ•˜๋Š” `output_dir`์ž…๋‹ˆ๋‹ค. `push_to_hub=True`๋ฅผ ์„ค์ •ํ•˜์—ฌ ์ด ๋ชจ๋ธ์„ Hub์— ํ‘ธ์‹œํ•ฉ๋‹ˆ๋‹ค(๋ชจ๋ธ์„ ์—…๋กœ๋“œํ•˜๋ ค๋ฉด Hugging Face์— ๋กœ๊ทธ์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค). ๊ฐ ์—ํฌํฌ๊ฐ€ ๋๋‚  ๋•Œ๋งˆ๋‹ค [`Trainer`]๊ฐ€ IoU ๋ฉ”ํŠธ๋ฆญ์„ ํ‰๊ฐ€ํ•˜๊ณ  ํ•™์Šต ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์ €์žฅํ•ฉ๋‹ˆ๋‹ค.
2. ๋ชจ๋ธ, ๋ฐ์ดํ„ฐ ์„ธํŠธ, ํ† ํฌ๋‚˜์ด์ €, ๋ฐ์ดํ„ฐ ์ฝœ๋ ˆ์ดํ„ฐ, `compute_metrics` ํ•จ์ˆ˜์™€ ํ•จ๊ป˜ ํ•™์Šต ์ธ์ž๋ฅผ [`Trainer`]์— ์ „๋‹ฌํ•˜์„ธ์š”.
3. ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด [`~Trainer.train`]๋ฅผ ํ˜ธ์ถœํ•˜์„ธ์š”.

```py
>>> training_args = TrainingArguments(
...     output_dir="segformer-b0-scene-parse-150",
...     learning_rate=6e-5,
...     num_train_epochs=50,
...     per_device_train_batch_size=2,
...     per_device_eval_batch_size=2,
...     save_total_limit=3,
...     eval_strategy="steps",
...     save_strategy="steps",
...     save_steps=20,
...     eval_steps=20,
...     logging_steps=1,
...     eval_accumulation_steps=5,
...     remove_unused_columns=False,
...     push_to_hub=True,
... )

>>> trainer = Trainer(
...     model=model,
...     args=training_args,
...     train_dataset=train_ds,
...     eval_dataset=test_ds,
...     compute_metrics=compute_metrics,
... )

>>> trainer.train()
```
ํ•™์Šต์ด ์™„๋ฃŒ๋˜๋ฉด, ๋ˆ„๊ตฌ๋‚˜ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก [`~transformers.Trainer.push_to_hub`] ๋ฉ”์„œ๋“œ๋ฅผ ์‚ฌ์šฉํ•ด Hub์— ๋ชจ๋ธ์„ ๊ณต์œ ํ•˜์„ธ์š”:

```py
>>> trainer.push_to_hub()
```


## ์ถ”๋ก ํ•˜๊ธฐ[[inference]]

์ด์ œ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ–ˆ์œผ๋‹ˆ ์ถ”๋ก ์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค!

์ถ”๋ก ํ•  ์ด๋ฏธ์ง€๋ฅผ ๋กœ๋“œํ•˜์„ธ์š”:

```py
>>> image = ds[0]["image"]
>>> image
```

<div class="flex justify-center">
    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/semantic-seg-image.png" alt="Image of bedroom"/>
</div>


์ถ”๋ก ์„ ์œ„ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•œ ๋ชจ๋ธ์„ ์‹œํ—˜ํ•ด ๋ณด๋Š” ๊ฐ€์žฅ ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์€ [`pipeline`]์—์„œ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€ ๋ถ„ํ• ์„ ์œ„ํ•œ `pipeline`์„ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๊ณ  ์ด๋ฏธ์ง€๋ฅผ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค:

```py
>>> from transformers import pipeline

>>> segmenter = pipeline("image-segmentation", model="my_awesome_seg_model")
>>> segmenter(image)
[{'score': None,
  'label': 'wall',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062690>},
 {'score': None,
  'label': 'sky',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062A50>},
 {'score': None,
  'label': 'floor',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062B50>},
 {'score': None,
  'label': 'ceiling',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062A10>},
 {'score': None,
  'label': 'bed ',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062E90>},
 {'score': None,
  'label': 'windowpane',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062390>},
 {'score': None,
  'label': 'cabinet',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062550>},
 {'score': None,
  'label': 'chair',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062D90>},
 {'score': None,
  'label': 'armchair',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062E10>}]
```
์›ํ•˜๋Š” ๊ฒฝ์šฐ `pipeline`์˜ ๊ฒฐ๊ณผ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋ณต์ œํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ํ”„๋กœ์„ธ์„œ๋กœ ์ด๋ฏธ์ง€๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ณ  `pixel_values`์„ GPU์— ๋ฐฐ์น˜ํ•ฉ๋‹ˆ๋‹ค:

```py
>>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # ๊ฐ€๋Šฅํ•˜๋‹ค๋ฉด GPU๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ๊ทธ๋ ‡์ง€ ์•Š๋‹ค๋ฉด CPU๋ฅผ ์‚ฌ์šฉํ•˜์„ธ์š”
>>> encoding = image_processor(image, return_tensors="pt")
>>> pixel_values = encoding.pixel_values.to(device)
```

๋ชจ๋ธ์— ์ž…๋ ฅ์„ ์ „๋‹ฌํ•˜๊ณ  `logits`๋ฅผ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค:

```py
>>> outputs = model(pixel_values=pixel_values)
>>> logits = outputs.logits.cpu()
```
๊ทธ๋Ÿฐ ๋‹ค์Œ ๋กœ์ง“์˜ ํฌ๊ธฐ๋ฅผ ์›๋ณธ ์ด๋ฏธ์ง€ ํฌ๊ธฐ๋กœ ๋‹ค์‹œ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค:

```py
>>> upsampled_logits = nn.functional.interpolate(
...     logits,
...     size=image.size[::-1],
...     mode="bilinear",
...     align_corners=False,
... )

>>> pred_seg = upsampled_logits.argmax(dim=1)[0]
```


๊ฒฐ๊ณผ๋ฅผ ์‹œ๊ฐํ™”ํ•˜๋ ค๋ฉด [dataset color palette](https://github.com/tensorflow/models/blob/3f1ca33afe3c1631b733ea7e40c294273b9e406d/research/deeplab/utils/get_dataset_colormap.py#L51)๋ฅผ ๊ฐ ํด๋ž˜์Šค๋ฅผ RGB ๊ฐ’์— ๋งคํ•‘ํ•˜๋Š” `ade_palette()`๋กœ ๋กœ๋“œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋ฏธ์ง€์™€ ์˜ˆ์ธก๋œ ๋ถ„ํ•  ์ง€๋„(segmentation map)์„ ๊ฒฐํ•ฉํ•˜์—ฌ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:

```py
>>> import matplotlib.pyplot as plt
>>> import numpy as np

>>> color_seg = np.zeros((pred_seg.shape[0], pred_seg.shape[1], 3), dtype=np.uint8)
>>> palette = np.array(ade_palette())
>>> for label, color in enumerate(palette):
...     color_seg[pred_seg == label, :] = color
>>> color_seg = color_seg[..., ::-1]  # BGR๋กœ ๋ณ€ํ™˜

>>> img = np.array(image) * 0.5 + color_seg * 0.5  # ๋ถ„ํ•  ์ง€๋„์œผ๋กœ ์ด๋ฏธ์ง€ ๊ตฌ์„ฑ
>>> img = img.astype(np.uint8)

>>> plt.figure(figsize=(15, 10))
>>> plt.imshow(img)
>>> plt.show()
```

<div class="flex justify-center">
    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/semantic-seg-preds.png" alt="Image of bedroom overlaid with segmentation map"/>
</div>