File size: 5,976 Bytes
17c6d62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
<!--Copyright 2024 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.

-->

# ์ด๋ฏธ์ง€ ํŠน์ง• ์ถ”์ถœ[[image-feature-extraction]]

[[open-in-colab]]

์ด๋ฏธ์ง€ ํŠน์ง• ์ถ”์ถœ์€ ์ฃผ์–ด์ง„ ์ด๋ฏธ์ง€์—์„œ ์˜๋ฏธ๋ก ์ ์œผ๋กœ ์˜๋ฏธ ์žˆ๋Š” ํŠน์ง•์„ ์ถ”์ถœํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ด๋ฏธ์ง€ ์œ ์‚ฌ์„ฑ ๋ฐ ์ด๋ฏธ์ง€ ๊ฒ€์ƒ‰ ๋“ฑ ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฒŒ๋‹ค๊ฐ€ ๋Œ€๋ถ€๋ถ„์˜ ์ปดํ“จํ„ฐ ๋น„์ „ ๋ชจ๋ธ์€ ์ด๋ฏธ์ง€ ํŠน์ง• ์ถ”์ถœ์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ž‘์—… ํŠนํ™” ํ—ค๋“œ(์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜, ๋ฌผ์ฒด ๊ฐ์ง€ ๋“ฑ)๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  ํŠน์ง•์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŠน์ง•์€ ๊ฐ€์žฅ์ž๋ฆฌ ๊ฐ์ง€, ๋ชจ์„œ๋ฆฌ ๊ฐ์ง€ ๋“ฑ ๊ณ ์ฐจ์› ์ˆ˜์ค€์—์„œ ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ ๋ชจ๋ธ์˜ ๊นŠ์ด์— ๋”ฐ๋ผ ์‹ค์ œ ์„ธ๊ณ„์— ๋Œ€ํ•œ ์ •๋ณด(์˜ˆ: ๊ณ ์–‘์ด๊ฐ€ ์–ด๋–ป๊ฒŒ ์ƒ๊ฒผ๋Š”์ง€)๋ฅผ ํฌํ•จํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ์ถœ๋ ฅ์€ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ƒˆ๋กœ์šด ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ด ๊ฐ€์ด๋“œ์—์„œ๋Š”:

- `image-feature-extraction` ํŒŒ์ดํ”„๋ผ์ธ์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ์ด๋ฏธ์ง€ ์œ ์‚ฌ์„ฑ ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ฐฐ์›๋‹ˆ๋‹ค.
- ๊ธฐ๋ณธ ๋ชจ๋ธ ์ถ”๋ก ์œผ๋กœ ๋™์ผํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.

## `image-feature-extraction` ํŒŒ์ดํ”„๋ผ์ธ์„ ์ด์šฉํ•œ ์ด๋ฏธ์ง€ ์œ ์‚ฌ์„ฑ[[image-similarity-using-image-feature-extraction-pipeline]]

๋ฌผ๊ณ ๊ธฐ ๊ทธ๋ฌผ ์œ„์— ์•‰์•„ ์žˆ๋Š” ๋‘ ์žฅ์˜ ๊ณ ์–‘์ด ์‚ฌ์ง„์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ ํ•˜๋‚˜๋Š” ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์ž…๋‹ˆ๋‹ค.

```python
from PIL import Image
import requests

img_urls = ["https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png", "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.jpeg"]
image_real = Image.open(requests.get(img_urls[0], stream=True).raw).convert("RGB")
image_gen = Image.open(requests.get(img_urls[1], stream=True).raw).convert("RGB")
```

ํŒŒ์ดํ”„๋ผ์ธ์„ ์‹คํ–‰ํ•ด ๋ด…์‹œ๋‹ค. ๋จผ์ € ํŒŒ์ดํ”„๋ผ์ธ์„ ์ดˆ๊ธฐํ™”ํ•˜์„ธ์š”. ๋ชจ๋ธ์„ ์ง€์ •ํ•˜์ง€ ์•Š์œผ๋ฉด, ํŒŒ์ดํ”„๋ผ์ธ์€ ์ž๋™์œผ๋กœ [google/vit-base-patch16-224](google/vit-base-patch16-224) ๋ชจ๋ธ๋กœ ์ดˆ๊ธฐํ™”๋ฉ๋‹ˆ๋‹ค. ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•˜๋ ค๋ฉด `pool`์„ True๋กœ ์„ค์ •ํ•˜์„ธ์š”. 


```python
import torch
from transformers import pipeline

DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
pipe = pipeline(task="image-feature-extraction", model_name="google/vit-base-patch16-384", device=DEVICE, pool=True)
```

`pipe`๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก ํ•˜๋ ค๋ฉด ๋‘ ์ด๋ฏธ์ง€๋ฅผ ๋ชจ๋‘ ์ „๋‹ฌํ•˜์„ธ์š”.

```python
outputs = pipe([image_real, image_gen])
```

์ถœ๋ ฅ์—๋Š” ๋‘ ์ด๋ฏธ์ง€์˜ ํ’€๋ง๋œ(pooled) ์ž„๋ฒ ๋”ฉ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.

```python
# ๋‹จ์ผ ์ถœ๋ ฅ์˜ ๊ธธ์ด ๊ตฌํ•˜๊ธฐ
print(len(outputs[0][0]))
# ์ถœ๋ ฅ ๊ฒฐ๊ณผ ํ‘œ์‹œํ•˜๊ธฐ
print(outputs)

# 768
# [[[-0.03909236937761307, 0.43381670117378235, -0.06913255900144577,
```

์œ ์‚ฌ๋„ ์ ์ˆ˜๋ฅผ ์–ป์œผ๋ ค๋ฉด, ์ด๋“ค์„ ์œ ์‚ฌ๋„ ํ•จ์ˆ˜์— ์ „๋‹ฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

```python
from torch.nn.functional import cosine_similarity

similarity_score = cosine_similarity(torch.Tensor(outputs[0]),
                                     torch.Tensor(outputs[1]), dim=1)

print(similarity_score)

# tensor([0.6043])
```

ํ’€๋ง ์ด์ „์˜ ๋งˆ์ง€๋ง‰ ์€๋‹‰ ์ƒํƒœ๋ฅผ ์–ป๊ณ  ์‹ถ๋‹ค๋ฉด, `pool` ๋งค๊ฐœ๋ณ€์ˆ˜์— ์•„๋ฌด ๊ฐ’๋„ ์ „๋‹ฌํ•˜์ง€ ๋งˆ์„ธ์š”. ๋˜ํ•œ, ๊ธฐ๋ณธ๊ฐ’์€ `False`๋กœ ์„ค์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์€๋‹‰ ์ƒํƒœ๋Š” ๋ชจ๋ธ์˜ ํŠน์ง•์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒˆ๋กœ์šด ๋ถ„๋ฅ˜๊ธฐ๋‚˜ ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๋ฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.

```python
pipe = pipeline(task="image-feature-extraction", model_name="google/vit-base-patch16-224", device=DEVICE)
output = pipe(image_real)
```

์•„์ง ์ถœ๋ ฅ์ด ํ’€๋ง๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์—, ์ฒซ ๋ฒˆ์งธ ์ฐจ์›์€ ๋ฐฐ์น˜ ํฌ๊ธฐ์ด๊ณ  ๋งˆ์ง€๋ง‰ ๋‘ ์ฐจ์›์€ ์ž„๋ฒ ๋”ฉ ํ˜•ํƒœ์ธ ๋งˆ์ง€๋ง‰ ์€๋‹‰ ์ƒํƒœ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

```python
import numpy as np
print(np.array(outputs).shape)
# (1, 197, 768)
```

## `AutoModel`์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ง•๊ณผ ์œ ์‚ฌ์„ฑ ์–ป๊ธฐ[[getting-features-and-similarities-using-automodel]]

transformers์˜ `AutoModel` ํด๋ž˜์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ง•์„ ์–ป์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. `AutoModel`์€ ์ž‘์—… ํŠนํ™” ํ—ค๋“œ ์—†์ด ๋ชจ๋“  transformers ๋ชจ๋ธ์„ ๋กœ๋“œํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ํŠน์ง•์„ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

```python
from transformers import AutoImageProcessor, AutoModel

processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
model = AutoModel.from_pretrained("google/vit-base-patch16-224").to(DEVICE)
```

์ถ”๋ก ์„ ์œ„ํ•œ ๊ฐ„๋‹จํ•œ ํ•จ์ˆ˜๋ฅผ ์ž‘์„ฑํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋จผ์ € ์ž…๋ ฅ๊ฐ’์„ `processor`์— ์ „๋‹ฌํ•œ ๋‹ค์Œ, ๊ทธ ์ถœ๋ ฅ๊ฐ’์„ `model`์— ์ „๋‹ฌํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.

```python
def infer(image):
  inputs = processor(image, return_tensors="pt").to(DEVICE)
  outputs = model(**inputs)
  return outputs.pooler_output
```

์ด ํ•จ์ˆ˜์— ์ด๋ฏธ์ง€๋ฅผ ์ง์ ‘ ์ „๋‹ฌํ•˜์—ฌ ์ž„๋ฒ ๋”ฉ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

```python
embed_real = infer(image_real)
embed_gen = infer(image_gen)
```

๊ทธ๋ฆฌ๊ณ  ์ด ์ž„๋ฒ ๋”ฉ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์‹œ ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

```python
from torch.nn.functional import cosine_similarity

similarity_score = cosine_similarity(embed_real, embed_gen, dim=1)
print(similarity_score)

# tensor([0.6061], device='cuda:0', grad_fn=<SumBackward1>)
```