Transformers documentation

๐Ÿค— Transformers๋กœ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ

You are viewing main version, which requires installation from source. If you'd like regular pip install, checkout the latest stable version (v4.46.2).
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

๐Ÿค— Transformers๋กœ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ

๐Ÿค— Transformers๋Š” ์ž์—ฐ์–ด์ฒ˜๋ฆฌ(NLP), ์ปดํ“จํ„ฐ ๋น„์ „, ์˜ค๋””์˜ค ๋ฐ ์Œ์„ฑ ์ฒ˜๋ฆฌ ์ž‘์—…์— ๋Œ€ํ•œ ์‚ฌ์ „ํ›ˆ๋ จ๋œ ์ตœ์ฒจ๋‹จ ๋ชจ๋ธ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์ž…๋‹ˆ๋‹ค. ์ด ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์„ ์œ„ํ•œ ํ˜„๋Œ€์ ์ธ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง๊ณผ ๊ฐ™์€ ํŠธ๋žœ์Šคํฌ๋จธ๊ฐ€ ์•„๋‹Œ ๋ชจ๋ธ๋„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์Šค๋งˆํŠธํฐ, ์•ฑ, ํ…”๋ ˆ๋น„์ „๊ณผ ๊ฐ™์€ ์˜ค๋Š˜๋‚  ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ์†Œ๋น„์ž ์ œํ’ˆ์„ ์‚ดํŽด๋ณด๋ฉด, ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ์ˆ ์ด ๊ทธ ๋’ค์— ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์„ ํ™•๋ฅ ์ด ๋†’์Šต๋‹ˆ๋‹ค. ์Šค๋งˆํŠธํฐ์œผ๋กœ ์ดฌ์˜ํ•œ ์‚ฌ์ง„์—์„œ ๋ฐฐ๊ฒฝ ๊ฐ์ฒด๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  ์‹ถ๋‹ค๋ฉด ์–ด๋–ป๊ฒŒ ํ• ๊นŒ์š”? ์ด๋Š” ํŒŒ๋†‰ํ‹ฑ ์„ธ๊ทธ๋ฉ˜ํ…Œ์ด์…˜ ์ž‘์—…์˜ ์˜ˆ์ž…๋‹ˆ๋‹ค(์•„์ง ์ด๊ฒŒ ๋ฌด์—‡์ธ์ง€ ๋ชจ๋ฅธ๋‹ค๋ฉด, ๋‹ค์Œ ์„น์…˜์—์„œ ์„ค๋ช…ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค!).

์ด ํŽ˜์ด์ง€๋Š” ๋‹ค์–‘ํ•œ ์Œ์„ฑ ๋ฐ ์˜ค๋””์˜ค, ์ปดํ“จํ„ฐ ๋น„์ „, NLP ์ž‘์—…์„ ๐Ÿค— Transformers ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋‹ค๋ฃจ๋Š” ๊ฐ„๋‹จํ•œ ์˜ˆ์ œ๋ฅผ 3์ค„์˜ ์ฝ”๋“œ๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.

์˜ค๋””์˜ค

์Œ์„ฑ ๋ฐ ์˜ค๋””์˜ค ์ฒ˜๋ฆฌ ์ž‘์—…์€ ๋‹ค๋ฅธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์™€ ์•ฝ๊ฐ„ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์ด๋Š” ์ฃผ๋กœ ์˜ค๋””์˜ค๊ฐ€ ์—ฐ์†์ ์ธ ์‹ ํ˜ธ๋กœ ์ž…๋ ฅ๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ํ…์ŠคํŠธ์™€ ๋‹ฌ๋ฆฌ ์›๋ณธ ์˜ค๋””์˜ค ํŒŒํ˜•(waveform)์€ ๋ฌธ์žฅ์ด ๋‹จ์–ด๋กœ ๋‚˜๋ˆ ์ง€๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๊น”๋”ํ•˜๊ฒŒ ์ด์‚ฐ์ ์ธ ๋ฌถ์Œ์œผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์›๋ณธ ์˜ค๋””์˜ค ์‹ ํ˜ธ๋Š” ์ผ์ •ํ•œ ๊ฐ„๊ฒฉ์œผ๋กœ ์ƒ˜ํ”Œ๋ง๋ฉ๋‹ˆ๋‹ค. ํ•ด๋‹น ๊ฐ„๊ฒฉ ๋‚ด์—์„œ ๋” ๋งŽ์€ ์ƒ˜ํ”Œ์„ ์ทจํ•  ๊ฒฝ์šฐ ์ƒ˜ํ”Œ๋ง๋ฅ ์ด ๋†’์•„์ง€๋ฉฐ, ์˜ค๋””์˜ค๋Š” ์›๋ณธ ์˜ค๋””์˜ค ์†Œ์Šค์— ๋” ๊ฐ€๊นŒ์›Œ์ง‘๋‹ˆ๋‹ค.

๊ณผ๊ฑฐ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์˜ค๋””์˜ค์—์„œ ์œ ์šฉํ•œ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ์˜ค๋””์˜ค๋ฅผ ์ „์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ํ˜„์žฌ๋Š” ์›๋ณธ ์˜ค๋””์˜ค ํŒŒํ˜•์„ ํŠน์„ฑ ์ธ์ฝ”๋”์— ์ง์ ‘ ๋„ฃ์–ด์„œ ์˜ค๋””์˜ค ํ‘œํ˜„(representation)์„ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ์ด ๋” ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์ „์ฒ˜๋ฆฌ ๋‹จ๊ณ„๊ฐ€ ๋‹จ์ˆœํ•ด์ง€๊ณ  ๋ชจ๋ธ์ด ๊ฐ€์žฅ ์ค‘์š”ํ•œ ํŠน์ง•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์˜ค๋””์˜ค ๋ถ„๋ฅ˜

์˜ค๋””์˜ค ๋ถ„๋ฅ˜๋Š” ์˜ค๋””์˜ค ๋ฐ์ดํ„ฐ์— ๋ฏธ๋ฆฌ ์ •์˜๋œ ํด๋ž˜์Šค ์ง‘ํ•ฉ์˜ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๋งŽ์€ ๊ตฌ์ฒด์ ์ธ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์„ ํฌํ•จํ•œ ๋„“์€ ๋ฒ”์ฃผ์ž…๋‹ˆ๋‹ค.

์ผ๋ถ€ ์˜ˆ์‹œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ์Œํ–ฅ ์žฅ๋ฉด ๋ถ„๋ฅ˜: ์˜ค๋””์˜ค์— ์žฅ๋ฉด ๋ ˆ์ด๋ธ”(โ€œ์‚ฌ๋ฌด์‹คโ€, โ€œํ•ด๋ณ€โ€, โ€œ๊ฒฝ๊ธฐ์žฅโ€)์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ์Œํ–ฅ ์ด๋ฒคํŠธ ๊ฐ์ง€: ์˜ค๋””์˜ค์— ์†Œ๋ฆฌ ์ด๋ฒคํŠธ ๋ ˆ์ด๋ธ”(โ€œ์ฐจ ๊ฒฝ์ โ€, โ€œ๊ณ ๋ž˜ ์šธ์Œ์†Œ๋ฆฌโ€, โ€œ์œ ๋ฆฌ ํŒŒ์†โ€)์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ํƒœ๊น…: ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์†Œ๋ฆฌ(์ƒˆ ์ง€์ €๊ท, ํšŒ์˜์—์„œ์˜ ํ™”์ž ์‹๋ณ„)๊ฐ€ ํฌํ•จ๋œ ์˜ค๋””์˜ค์— ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ์Œ์•… ๋ถ„๋ฅ˜: ์Œ์•…์— ์žฅ๋ฅด ๋ ˆ์ด๋ธ”(โ€œ๋ฉ”ํƒˆโ€, โ€œํž™ํ•ฉโ€, โ€œ์ปจํŠธ๋ฆฌโ€)์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.
>>> from transformers import pipeline

>>> classifier = pipeline(task="audio-classification", model="superb/hubert-base-superb-er")
>>> preds = classifier("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
>>> preds = [{"score": round(pred["score"], 4), "label": pred["label"]} for pred in preds]
>>> preds
[{'score': 0.4532, 'label': 'hap'},
 {'score': 0.3622, 'label': 'sad'},
 {'score': 0.0943, 'label': 'neu'},
 {'score': 0.0903, 'label': 'ang'}]

์ž๋™ ์Œ์„ฑ ์ธ์‹

์ž๋™ ์Œ์„ฑ ์ธ์‹(ASR)์€ ์Œ์„ฑ์„ ํ…์ŠคํŠธ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์Œ์„ฑ์€ ์ธ๊ฐ„์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ์˜์‚ฌ์†Œํ†ต ํ˜•ํƒœ์ด๊ธฐ ๋•Œ๋ฌธ์— ASR์€ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์˜ค๋””์˜ค ์ž‘์—… ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์˜ค๋Š˜๋‚  ASR ์‹œ์Šคํ…œ์€ ์Šคํ”ผ์ปค, ์ „ํ™” ๋ฐ ์ž๋™์ฐจ์™€ ๊ฐ™์€ โ€œ์Šค๋งˆํŠธโ€ ๊ธฐ์ˆ  ์ œํ’ˆ์— ๋‚ด์žฅ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ฐ€์ƒ ๋น„์„œ์—๊ฒŒ ์Œ์•… ์žฌ์ƒ, ์•Œ๋ฆผ ์„ค์ • ๋ฐ ๋‚ ์”จ ์ •๋ณด๋ฅผ ์š”์ฒญํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

ํ•˜์ง€๋งŒ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค€ ํ•ต์‹ฌ ๋„์ „ ๊ณผ์ œ ์ค‘ ํ•˜๋‚˜๋Š” ์–‘์ด ๋ฐ์ดํ„ฐ ์–‘์ด ์ ์€ ์–ธ์–ด(low-resource language)์— ๋Œ€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋Œ€๋Ÿ‰์˜ ์Œ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•œ ํ›„ ๋ฐ์ดํ„ฐ ์–‘์ด ์ ์€ ์–ธ์–ด์—์„œ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์Œ์„ฑ ๋ฐ์ดํ„ฐ 1์‹œ๊ฐ„๋งŒ์œผ๋กœ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋ฉด ์ด์ „์˜ 100๋ฐฐ ๋งŽ์€ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ASR ์‹œ์Šคํ…œ๋ณด๋‹ค ํ›จ์”ฌ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

>>> from transformers import pipeline

>>> transcriber = pipeline(task="automatic-speech-recognition", model="openai/whisper-small")
>>> transcriber("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
{'text': ' I have a dream that one day this nation will rise up and live out the true meaning of its creed.'}

์ปดํ“จํ„ฐ ๋น„์ „

์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—… ์ค‘ ๊ฐ€์žฅ ์ดˆ๊ธฐ์˜ ์„ฑ๊ณต์ ์ธ ์ž‘์—… ์ค‘ ํ•˜๋‚˜๋Š” ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง(CNN)์„ ์‚ฌ์šฉํ•˜์—ฌ ์šฐํŽธ๋ฒˆํ˜ธ ์ˆซ์ž ์ด๋ฏธ์ง€๋ฅผ ์ธ์‹ํ•˜๋Š” ๊ฒƒ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€๋Š” ํ”ฝ์…€๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ ๊ฐ ํ”ฝ์…€์€ ์ˆซ์ž ๊ฐ’์œผ๋กœ ํ‘œํ˜„๋ฉ๋‹ˆ๋‹ค. ์ด๋กœ์จ ์ด๋ฏธ์ง€๋ฅผ ํ”ฝ์…€ ๊ฐ’์˜ ํ–‰๋ ฌ๋กœ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์ด ์‰ฌ์›Œ์ง‘๋‹ˆ๋‹ค. ํŠน์ •ํ•œ ํ”ฝ์…€ ๊ฐ’์˜ ์กฐํ•ฉ์€ ์ด๋ฏธ์ง€์˜ ์ƒ‰์ƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค์Œ ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์œผ๋กœ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค:

  1. ํ•ฉ์„ฑ๊ณฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€์˜ ๋‚ฎ์€ ์ˆ˜์ค€ ํŠน์ง•์—์„œ ๋†’์€ ์ˆ˜์ค€์˜ ์ถ”์ƒ์ ์ธ ์š”์†Œ๊นŒ์ง€ ๊ณ„์ธต์ ์œผ๋กœ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค.

  2. ์ด๋ฏธ์ง€๋ฅผ ํŒจ์น˜๋กœ ๋‚˜๋ˆ„๊ณ  ํŠธ๋žœ์Šคํฌ๋จธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ ์ง„์ ์œผ๋กœ ๊ฐ ์ด๋ฏธ์ง€ ํŒจ์น˜๊ฐ€ ์„œ๋กœ ์–ด๋– ํ•œ ๋ฐฉ์‹์œผ๋กœ ์—ฐ๊ด€๋˜์–ด ์ด๋ฏธ์ง€๋ฅผ ํ˜•์„ฑํ•˜๋Š”์ง€ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. CNN์—์„œ ์„ ํ˜ธํ•˜๋Š” ์ƒํ–ฅ์‹ ์ ‘๊ทผ๋ฒ•๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์ด ๋ฐฉ์‹์€ ํ๋ฆฟํ•œ ์ด๋ฏธ์ง€๋กœ ์ดˆ์•ˆ์„ ๊ทธ๋ฆฌ๊ณ  ์ ์ง„์ ์œผ๋กœ ์„ ๋ช…ํ•œ ์ด๋ฏธ์ง€๋กœ ๋งŒ๋“ค์–ด๊ฐ€๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜

์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๋Š” ํ•œ ๊ฐœ์˜ ์ „์ฒด ์ด๋ฏธ์ง€์— ๋ฏธ๋ฆฌ ์ •์˜๋œ ํด๋ž˜์Šค ์ง‘ํ•ฉ์˜ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค.

๋Œ€๋ถ€๋ถ„์˜ ๋ถ„๋ฅ˜ ์ž‘์—…๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜์—๋Š” ๋‹ค์–‘ํ•œ ์‹ค์šฉ์ ์ธ ์šฉ๋„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ผ๋ถ€ ์˜ˆ์‹œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ์˜๋ฃŒ: ์งˆ๋ณ‘์„ ๊ฐ์ง€ํ•˜๊ฑฐ๋‚˜ ํ™˜์ž ๊ฑด๊ฐ•์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด ์˜๋ฃŒ ์ด๋ฏธ์ง€์— ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ํ™˜๊ฒฝ: ์œ„์„ฑ ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•˜์—ฌ ์‚ฐ๋ฆผ ๋ฒŒ์ฑ„๋ฅผ ๊ฐ์‹œํ•˜๊ณ  ์•ผ์ƒ ์ง€์—ญ ๊ด€๋ฆฌ๋ฅผ ์œ„ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๊ฑฐ๋‚˜ ์‚ฐ๋ถˆ์„ ๊ฐ์ง€ํ•ฉ๋‹ˆ๋‹ค.
  • ๋†์—…: ์ž‘๋ฌผ ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•˜์—ฌ ์‹๋ฌผ ๊ฑด๊ฐ•์„ ํ™•์ธํ•˜๊ฑฐ๋‚˜ ์œ„์„ฑ ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•˜์—ฌ ํ† ์ง€ ์ด์šฉ ๊ด€์ฐฐ์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
  • ์ƒํƒœํ•™: ๋™๋ฌผ์ด๋‚˜ ์‹๋ฌผ ์ข… ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•˜์—ฌ ์•ผ์ƒ ๋™๋ฌผ ๊ฐœ์ฒด๊ตฐ์„ ์กฐ์‚ฌํ•˜๊ฑฐ๋‚˜ ๋ฉธ์ข… ์œ„๊ธฐ์— ์ฒ˜ํ•œ ์ข…์„ ์ถ”์ ํ•ฉ๋‹ˆ๋‹ค.
>>> from transformers import pipeline

>>> classifier = pipeline(task="image-classification")
>>> preds = classifier(
...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
... )
>>> preds = [{"score": round(pred["score"], 4), "label": pred["label"]} for pred in preds]
>>> print(*preds, sep="\n")
{'score': 0.4335, 'label': 'lynx, catamount'}
{'score': 0.0348, 'label': 'cougar, puma, catamount, mountain lion, painter, panther, Felis concolor'}
{'score': 0.0324, 'label': 'snow leopard, ounce, Panthera uncia'}
{'score': 0.0239, 'label': 'Egyptian cat'}
{'score': 0.0229, 'label': 'tiger cat'}

๊ฐ์ฒด ํƒ์ง€

์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜์™€ ๋‹ฌ๋ฆฌ ๊ฐ์ฒด ํƒ์ง€๋Š” ์ด๋ฏธ์ง€ ๋‚ด์—์„œ ์—ฌ๋Ÿฌ ๊ฐ์ฒด๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค๋กœ ์ •์˜๋œ ๊ฐ์ฒด์˜ ์œ„์น˜๋ฅผ ํŒŒ์•…ํ•ฉ๋‹ˆ๋‹ค.

๊ฐ์ฒด ํƒ์ง€์˜ ๋ช‡ ๊ฐ€์ง€ ์‘์šฉ ์˜ˆ์‹œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ์ž์œจ ์ฃผํ–‰ ์ฐจ๋Ÿ‰: ๋‹ค๋ฅธ ์ฐจ๋Ÿ‰, ๋ณดํ–‰์ž ๋ฐ ์‹ ํ˜ธ๋“ฑ๊ณผ ๊ฐ™์€ ์ผ์ƒ์ ์ธ ๊ตํ†ต ๊ฐ์ฒด๋ฅผ ๊ฐ์ง€ํ•ฉ๋‹ˆ๋‹ค.
  • ์›๊ฒฉ ๊ฐ์ง€: ์žฌ๋‚œ ๋ชจ๋‹ˆํ„ฐ๋ง, ๋„์‹œ ๊ณ„ํš ๋ฐ ๊ธฐ์ƒ ์˜ˆ์ธก ๋“ฑ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
  • ๊ฒฐํ•จ ํƒ์ง€: ๊ฑด๋ฌผ์˜ ๊ท ์—ด์ด๋‚˜ ๊ตฌ์กฐ์  ์†์ƒ, ์ œ์กฐ ๊ฒฐํ•จ ๋“ฑ์„ ํƒ์ง€ํ•ฉ๋‹ˆ๋‹ค.
>>> from transformers import pipeline

>>> detector = pipeline(task="object-detection")
>>> preds = detector(
...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
... )
>>> preds = [{"score": round(pred["score"], 4), "label": pred["label"], "box": pred["box"]} for pred in preds]
>>> preds
[{'score': 0.9865,
  'label': 'cat',
  'box': {'xmin': 178, 'ymin': 154, 'xmax': 882, 'ymax': 598}}]

์ด๋ฏธ์ง€ ๋ถ„ํ• 

์ด๋ฏธ์ง€ ๋ถ„ํ• ์€ ํ”ฝ์…€ ์ฐจ์›์˜ ์ž‘์—…์œผ๋กœ, ์ด๋ฏธ์ง€ ๋‚ด์˜ ๋ชจ๋“  ํ”ฝ์…€์„ ํด๋ž˜์Šค์— ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐ์ฒด ํƒ์ง€์™€ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ๊ฐ์ฒด ํƒ์ง€๋Š” ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€ ๋‚ด์˜ ๊ฐ์ฒด๋ฅผ ๋ ˆ์ด๋ธ”๋งํ•˜๊ณ  ์˜ˆ์ธกํ•˜๋Š” ๋ฐ˜๋ฉด, ๋ถ„ํ• ์€ ๋” ์„ธ๋ถ„ํ™”๋œ ์ž‘์—…์ž…๋‹ˆ๋‹ค. ๋ถ„ํ• ์€ ํ”ฝ์…€ ์ˆ˜์ค€์—์„œ ๊ฐ์ฒด๋ฅผ ๊ฐ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ด๋ฏธ์ง€ ๋ถ„ํ• ์—๋Š” ์—ฌ๋Ÿฌ ์œ ํ˜•์ด ์žˆ์Šต๋‹ˆ๋‹ค:

  • ์ธ์Šคํ„ด์Šค ๋ถ„ํ• : ๊ฐœ์ฒด์˜ ํด๋ž˜์Šค๋ฅผ ๋ ˆ์ด๋ธ”๋งํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, ๊ฐœ์ฒด์˜ ๊ฐ ๊ตฌ๋ถ„๋œ ์ธ์Šคํ„ด์Šค์—๋„ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค (โ€œ๊ฐœ-1โ€, โ€œ๊ฐœ-2โ€ ๋“ฑ).
  • ํŒŒ๋†‰ํ‹ฑ ๋ถ„ํ• : ์˜๋ฏธ์  ๋ถ„ํ• ๊ณผ ์ธ์Šคํ„ด์Šค ๋ถ„ํ• ์˜ ์กฐํ•ฉ์ž…๋‹ˆ๋‹ค. ๊ฐ ํ”ฝ์…€์„ ์˜๋ฏธ์  ํด๋ž˜์Šค๋กœ ๋ ˆ์ด๋ธ”๋งํ•˜๋Š” ๋™์‹œ์— ๊ฐœ์ฒด์˜ ๊ฐ๊ฐ ๊ตฌ๋ถ„๋œ ์ธ์Šคํ„ด์Šค๋กœ๋„ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.

๋ถ„ํ•  ์ž‘์—…์€ ์ž์œจ ์ฃผํ–‰ ์ฐจ๋Ÿ‰์—์„œ ์œ ์šฉํ•˜๋ฉฐ, ์ฃผ๋ณ€ ํ™˜๊ฒฝ์˜ ํ”ฝ์…€ ์ˆ˜์ค€ ์ง€๋„๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๋ณดํ–‰์ž์™€ ๋‹ค๋ฅธ ์ฐจ๋Ÿ‰ ์ฃผ๋ณ€์—์„œ ์•ˆ์ „ํ•˜๊ฒŒ ํƒ์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜๋ฃŒ ์˜์ƒ์—์„œ๋„ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ถ„ํ•  ์ž‘์—…์ด ํ”ฝ์…€ ์ˆ˜์ค€์—์„œ ๊ฐ์ฒด๋ฅผ ๊ฐ์ง€ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋น„์ •์ƒ์ ์ธ ์„ธํฌ๋‚˜ ์žฅ๊ธฐ์˜ ํŠน์ง•์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ๋ถ„ํ• ์€ ์˜๋ฅ˜ ๊ฐ€์ƒ ์‹œ์ฐฉ์ด๋‚˜ ์นด๋ฉ”๋ผ๋ฅผ ํ†ตํ•ด ์‹ค์ œ ์„ธ๊ณ„์— ๊ฐ€์ƒ ๊ฐœ์ฒด๋ฅผ ๋ง์”Œ์›Œ ์ฆ๊ฐ• ํ˜„์‹ค ๊ฒฝํ—˜์„ ๋งŒ๋“œ๋Š” ๋“ฑ ์ „์ž ์ƒ๊ฑฐ๋ž˜ ๋ถ„์•ผ์—์„œ๋„ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

>>> from transformers import pipeline

>>> segmenter = pipeline(task="image-segmentation")
>>> preds = segmenter(
...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
... )
>>> preds = [{"score": round(pred["score"], 4), "label": pred["label"]} for pred in preds]
>>> print(*preds, sep="\n")
{'score': 0.9879, 'label': 'LABEL_184'}
{'score': 0.9973, 'label': 'snow'}
{'score': 0.9972, 'label': 'cat'}

๊นŠ์ด ์ถ”์ •

๊นŠ์ด ์ถ”์ •์€ ์นด๋ฉ”๋ผ๋กœ๋ถ€ํ„ฐ ์ด๋ฏธ์ง€ ๋‚ด๋ถ€์˜ ๊ฐ ํ”ฝ์…€์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. ์ด ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์€ ํŠนํžˆ ์žฅ๋ฉด ์ดํ•ด์™€ ์žฌ๊ตฌ์„ฑ์— ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ž์œจ ์ฃผํ–‰ ์ฐจ๋Ÿ‰์€ ๋ณดํ–‰์ž, ๊ตํ†ต ํ‘œ์ง€ํŒ ๋ฐ ๋‹ค๋ฅธ ์ฐจ๋Ÿ‰๊ณผ ๊ฐ™์€ ๊ฐ์ฒด์™€์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์ดํ•ดํ•˜์—ฌ ์žฅ์• ๋ฌผ๊ณผ ์ถฉ๋Œ์„ ํ”ผํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊นŠ์ด ์ •๋ณด๋Š” ๋˜ํ•œ 2D ์ด๋ฏธ์ง€์—์„œ 3D ํ‘œํ˜„์„ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋ฉฐ ์ƒ๋ฌผํ•™์  ๊ตฌ์กฐ๋‚˜ ๊ฑด๋ฌผ์˜ ๊ณ ํ’ˆ์งˆ 3D ํ‘œํ˜„์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๊นŠ์ด ์ถ”์ •์—๋Š” ๋‘ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์žˆ์Šต๋‹ˆ๋‹ค:

  • ์Šคํ…Œ๋ ˆ์˜ค: ์•ฝ๊ฐ„ ๋‹ค๋ฅธ ๊ฐ๋„์—์„œ ์ดฌ์˜๋œ ๋™์ผํ•œ ์ด๋ฏธ์ง€ ๋‘ ์žฅ์„ ๋น„๊ตํ•˜์—ฌ ๊นŠ์ด๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ๋‹จ์•ˆ: ๋‹จ์ผ ์ด๋ฏธ์ง€์—์„œ ๊นŠ์ด๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค.
>>> from transformers import pipeline

>>> depth_estimator = pipeline(task="depth-estimation")
>>> preds = depth_estimator(
...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
... )

์ž์—ฐ์–ด์ฒ˜๋ฆฌ

ํ…์ŠคํŠธ๋Š” ์ธ๊ฐ„์ด ์˜์‚ฌ ์†Œํ†ตํ•˜๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ๋ฐฉ์‹ ์ค‘ ํ•˜๋‚˜์ด๊ธฐ ๋•Œ๋ฌธ์— ์ž์—ฐ์–ด์ฒ˜๋ฆฌ ์—ญ์‹œ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์ž‘์—… ์œ ํ˜• ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ์ธ์‹ํ•˜๋Š” ํ˜•์‹์œผ๋กœ ํ…์ŠคํŠธ๋ฅผ ๋ณ€ํ™˜ํ•˜๋ ค๋ฉด ํ† ํฐํ™”ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ๊ฐœ๋ณ„ ๋‹จ์–ด ๋˜๋Š” ํ•˜์œ„ ๋‹จ์–ด(ํ† ํฐ)๋กœ ๋ถ„ํ• ํ•œ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ํ† ํฐ์„ ์ˆซ์ž๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์ˆซ์ž ์‹œํ€€์Šค๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ˆซ์ž ์‹œํ€€์Šค๋ฅผ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด์ฒ˜๋ฆฌ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๋ชจ๋ธ์— ์ž…๋ ฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค!

ํ…์ŠคํŠธ ๋ถ„๋ฅ˜

๋‹ค๋ฅธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์—์„œ์˜ ๋ถ„๋ฅ˜ ์ž‘์—…๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ํ…์ŠคํŠธ ๋ถ„๋ฅ˜๋Š” ๋ฏธ๋ฆฌ ์ •์˜๋œ ํด๋ž˜์Šค ์ง‘ํ•ฉ์—์„œ ํ…์ŠคํŠธ ์‹œํ€€์Šค(๋ฌธ์žฅ ์ˆ˜์ค€, ๋‹จ๋ฝ ๋˜๋Š” ๋ฌธ์„œ ๋“ฑ)์— ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ๋ถ„๋ฅ˜์—๋Š” ๋‹ค์–‘ํ•œ ์‹ค์šฉ์ ์ธ ์‘์šฉ ์‚ฌ๋ก€๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ผ๋ถ€ ์˜ˆ์‹œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ๊ฐ์„ฑ ๋ถ„์„: ํ…์ŠคํŠธ๋ฅผ ๊ธ์ • ๋˜๋Š” ๋ถ€์ •๊ณผ ๊ฐ™์€ ์–ด๋–ค ๊ทน์„ฑ์— ๋”ฐ๋ผ ๋ ˆ์ด๋ธ”๋งํ•˜์—ฌ ์ •์น˜, ๊ธˆ์œต, ๋งˆ์ผ€ํŒ…๊ณผ ๊ฐ™์€ ๋ถ„์•ผ์—์„œ ์˜์‚ฌ ๊ฒฐ์ •์— ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์ง€์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ์ฝ˜ํ…์ธ  ๋ถ„๋ฅ˜: ํ…์ŠคํŠธ๋ฅผ ์ฃผ์ œ์— ๋”ฐ๋ผ ๋ ˆ์ด๋ธ”๋ง(๋‚ ์”จ, ์Šคํฌ์ธ , ๊ธˆ์œต ๋“ฑ)ํ•˜์—ฌ ๋‰ด์Šค ๋ฐ ์†Œ์…œ ๋ฏธ๋””์–ด ํ”ผ๋“œ์—์„œ ์ •๋ณด๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ  ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
>>> from transformers import pipeline

>>> classifier = pipeline(task="sentiment-analysis")
>>> preds = classifier("Hugging Face is the best thing since sliced bread!")
>>> preds = [{"score": round(pred["score"], 4), "label": pred["label"]} for pred in preds]
>>> preds
[{'score': 0.9991, 'label': 'POSITIVE'}]

ํ† ํฐ ๋ถ„๋ฅ˜

๋ชจ๋“  ์ž์—ฐ์–ด์ฒ˜๋ฆฌ ์ž‘์—…์—์„œ๋Š” ํ…์ŠคํŠธ๊ฐ€ ๊ฐœ๋ณ„ ๋‹จ์–ด๋‚˜ ํ•˜์œ„ ๋‹จ์–ด๋กœ ๋ถ„๋ฆฌ๋˜์–ด ์ „์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค. ๋ถ„๋ฆฌ๋œ ๋‹จ์–ด๋ฅผ ํ† ํฐ์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ํ† ํฐ ๋ถ„๋ฅ˜๋Š” ๊ฐ ํ† ํฐ์— ๋ฏธ๋ฆฌ ์ •์˜๋œ ํด๋ž˜์Šค ์ง‘ํ•ฉ์˜ ๋ ˆ์ด๋ธ”์„ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค.

ํ† ํฐ ๋ถ„๋ฅ˜์˜ ๋‘ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ์œ ํ˜•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ๊ฐœ์ฒด๋ช… ์ธ์‹ (NER): ํ† ํฐ์„ ์กฐ์ง, ์ธ๋ฌผ, ์œ„์น˜ ๋˜๋Š” ๋‚ ์งœ์™€ ๊ฐ™์€ ๊ฐœ์ฒด ๋ฒ”์ฃผ์— ๋”ฐ๋ผ ๋ ˆ์ด๋ธ”๋งํ•ฉ๋‹ˆ๋‹ค. NER์€ ํŠนํžˆ ์œ ์ „์ฒดํ•™์ ์ธ ํ™˜๊ฒฝ์—์„œ ์œ ์ „์ž, ๋‹จ๋ฐฑ์งˆ ๋ฐ ์•ฝ๋ฌผ ์ด๋ฆ„์— ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
  • ํ’ˆ์‚ฌ ํƒœ๊น… (POS): ๋ช…์‚ฌ, ๋™์‚ฌ, ํ˜•์šฉ์‚ฌ์™€ ๊ฐ™์€ ํ’ˆ์‚ฌ์— ๋”ฐ๋ผ ํ† ํฐ์— ๋ ˆ์ด๋ธ”์„ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค. POS๋Š” ๋ฒˆ์—ญ ์‹œ์Šคํ…œ์ด ๋™์ผํ•œ ๋‹จ์–ด๊ฐ€ ๋ฌธ๋ฒ•์ ์œผ๋กœ ์–ด๋–ป๊ฒŒ ๋‹ค๋ฅธ์ง€ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค (๋ช…์‚ฌ๋กœ ์‚ฌ์šฉ๋˜๋Š” โ€œbank(์€ํ–‰)โ€œ๊ณผ ๋™์‚ฌ๋กœ ์‚ฌ์šฉ๋˜๋Š” โ€œbank(์˜ˆ๊ธˆ์„ ์˜ˆ์น˜ํ•˜๋‹ค)โ€œ๊ณผ ๊ฐ™์€ ๊ฒฝ์šฐ).
>>> from transformers import pipeline

>>> classifier = pipeline(task="ner")
>>> preds = classifier("Hugging Face is a French company based in New York City.")
>>> preds = [
...     {
...         "entity": pred["entity"],
...         "score": round(pred["score"], 4),
...         "index": pred["index"],
...         "word": pred["word"],
...         "start": pred["start"],
...         "end": pred["end"],
...     }
...     for pred in preds
... ]
>>> print(*preds, sep="\n")
{'entity': 'I-ORG', 'score': 0.9968, 'index': 1, 'word': 'Hu', 'start': 0, 'end': 2}
{'entity': 'I-ORG', 'score': 0.9293, 'index': 2, 'word': '##gging', 'start': 2, 'end': 7}
{'entity': 'I-ORG', 'score': 0.9763, 'index': 3, 'word': 'Face', 'start': 8, 'end': 12}
{'entity': 'I-MISC', 'score': 0.9983, 'index': 6, 'word': 'French', 'start': 18, 'end': 24}
{'entity': 'I-LOC', 'score': 0.999, 'index': 10, 'word': 'New', 'start': 42, 'end': 45}
{'entity': 'I-LOC', 'score': 0.9987, 'index': 11, 'word': 'York', 'start': 46, 'end': 50}
{'entity': 'I-LOC', 'score': 0.9992, 'index': 12, 'word': 'City', 'start': 51, 'end': 55}

์งˆ์˜์‘๋‹ต

์งˆ์˜์‘๋‹ต์€ ๋˜ ํ•˜๋‚˜์˜ ํ† ํฐ ์ฐจ์›์˜ ์ž‘์—…์œผ๋กœ, ๋ฌธ๋งฅ์ด ์žˆ์„ ๋•Œ(๊ฐœ๋ฐฉํ˜• ๋„๋ฉ”์ธ)์™€ ๋ฌธ๋งฅ์ด ์—†์„ ๋•Œ(ํ์‡„ํ˜• ๋„๋ฉ”์ธ) ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๊ฐ€์ƒ ๋น„์„œ์—๊ฒŒ ์‹๋‹น์ด ์˜์—… ์ค‘์ธ์ง€์™€ ๊ฐ™์€ ์งˆ๋ฌธ์„ ํ•  ๋•Œ๋งˆ๋‹ค ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ ๊ฐ ์ง€์› ๋˜๋Š” ๊ธฐ์ˆ  ์ง€์›์„ ์ œ๊ณตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰ ์—”์ง„์ด ์š”์ฒญํ•œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์งˆ๋ฌธ ๋‹ต๋ณ€์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์ด ์žˆ์Šต๋‹ˆ๋‹ค:

  • ์ถ”์ถœํ˜•: ์งˆ๋ฌธ๊ณผ ๋ฌธ๋งฅ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ๋ชจ๋ธ์ด ์ฃผ์–ด์ง„ ๋ฌธ๋งฅ์˜ ์ผ๋ถ€์—์„œ ๊ฐ€์ ธ์˜จ ํ…์ŠคํŠธ์˜ ๋ฒ”์œ„๋ฅผ ๋‹ต๋ณ€์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
  • ์ƒ์„ฑํ˜•: ์งˆ๋ฌธ๊ณผ ๋ฌธ๋งฅ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ฃผ์–ด์ง„ ๋ฌธ๋งฅ์„ ํ†ตํ•ด ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ QuestionAnsweringPipeline ๋Œ€์‹  Text2TextGenerationPipeline์„ ํ†ตํ•ด ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค.
>>> from transformers import pipeline

>>> question_answerer = pipeline(task="question-answering")
>>> preds = question_answerer(
...     question="What is the name of the repository?",
...     context="The name of the repository is huggingface/transformers",
... )
>>> print(
...     f"score: {round(preds['score'], 4)}, start: {preds['start']}, end: {preds['end']}, answer: {preds['answer']}"
... )
score: 0.9327, start: 30, end: 54, answer: huggingface/transformers

์š”์•ฝ

์š”์•ฝ์€ ์›๋ณธ ๋ฌธ์„œ์˜ ์˜๋ฏธ๋ฅผ ์ตœ๋Œ€ํ•œ ๋ณด์กดํ•˜๋ฉด์„œ ๊ธด ๋ฌธ์„œ๋ฅผ ์งง์€ ๋ฌธ์„œ๋กœ ๋งŒ๋“œ๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์š”์•ฝ์€ sequence-to-sequence ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ž…๋ ฅ๋ณด๋‹ค ์งง์€ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค. ์š”์•ฝ ์ž‘์—…์€ ๋…์ž๊ฐ€ ์žฅ๋ฌธ ๋ฌธ์„œ๋“ค์˜ ์ฃผ์š” ํฌ์ธํŠธ๋ฅผ ๋น ๋ฅด๊ฒŒ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž…๋ฒ•์•ˆ, ๋ฒ•๋ฅ  ๋ฐ ๊ธˆ์œต ๋ฌธ์„œ, ํŠนํ—ˆ ๋ฐ ๊ณผํ•™ ๋…ผ๋ฌธ์€ ์š”์•ฝ ์ž‘์—…์ด ๋…์ž์˜ ์‹œ๊ฐ„์„ ์ ˆ์•ฝํ•˜๊ณ  ๋…์„œ ๋ณด์กฐ ๋„๊ตฌ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค.

์งˆ๋ฌธ ๋‹ต๋ณ€๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์š”์•ฝ์—๋Š” ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์ด ์žˆ์Šต๋‹ˆ๋‹ค:

  • ์ถ”์ถœํ˜•: ์›๋ณธ ํ…์ŠคํŠธ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๋ฌธ์žฅ์„ ์‹๋ณ„ํ•˜๊ณ  ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.
  • ์ƒ์„ฑํ˜•: ์›๋ณธ ํ…์ŠคํŠธ์—์„œ ๋ชฉํ‘œ ์š”์•ฝ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ ๋ฌธ์„œ์— ์—†๋Š” ์ƒˆ๋กœ์šด ๋‹จ์–ด๋ฅผ ํฌํ•จํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. SummarizationPipeline์€ ์ƒ์„ฑํ˜• ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
>>> from transformers import pipeline

>>> summarizer = pipeline(task="summarization")
>>> summarizer(
...     "In this work, we presented the Transformer, the first sequence transduction model based entirely on attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention. For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art. In the former task our best model outperforms even all previously reported ensembles."
... )
[{'summary_text': ' The Transformer is the first sequence transduction model based entirely on attention . It replaces the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention . For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers .'}]

๋ฒˆ์—ญ

๋ฒˆ์—ญ์€ ํ•œ ์–ธ์–ด๋กœ ๋œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐฐ๊ฒฝ์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ๋“ค์ด ์„œ๋กœ ์†Œํ†ตํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ฃผ๋Š” ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋” ๋„“์€ ๋Œ€์ค‘์—๊ฒŒ ์ฝ˜ํ…์ธ ๋ฅผ ๋ฒˆ์—ญํ•˜์—ฌ ์ „๋‹ฌํ•˜๊ฑฐ๋‚˜, ์ƒˆ๋กœ์šด ์–ธ์–ด๋ฅผ ๋ฐฐ์šฐ๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํ•™์Šต ๋„๊ตฌ๊ฐ€ ๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์š”์•ฝ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ๋ฒˆ์—ญ์€ sequence-to-sequence ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋ธ์€ ์ž…๋ ฅ ์‹œํ€€์Šค๋ฅผ ๋ฐ›์•„์„œ ์ถœ๋ ฅ์ด ๋˜๋Š” ๋ชฉํ‘œ ์‹œํ€€์Šค๋ฅผ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค.

์ดˆ๊ธฐ์˜ ๋ฒˆ์—ญ ๋ชจ๋ธ์€ ๋Œ€๋ถ€๋ถ„ ๋‹จ์ผ ์–ธ์–ด๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ์—ˆ์ง€๋งŒ, ์ตœ๊ทผ์—๋Š” ๋งŽ์€ ์–ธ์–ด ์Œ ๊ฐ„์— ๋ฒˆ์—ญ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค์ค‘ ์–ธ์–ด ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

>>> from transformers import pipeline

>>> text = "translate English to French: Hugging Face is a community-based open-source platform for machine learning."
>>> translator = pipeline(task="translation", model="google-t5/t5-small")
>>> translator(text)
[{'translation_text': "Hugging Face est une tribune communautaire de l'apprentissage des machines."}]

์–ธ์–ด ๋ชจ๋ธ๋ง

์–ธ์–ด ๋ชจ๋ธ๋ง์€ ํ…์ŠคํŠธ ์‹œํ€€์Šค์—์„œ ๋‹จ์–ด๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋งŽ์€ ๋‹ค๋ฅธ ํ•˜์œ„ ์ž‘์—…์— ๋”ฐ๋ผ ๋ฏธ์„ธ ์กฐ์ •๋  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋งค์šฐ ์ธ๊ธฐ ์žˆ๋Š” ์ž์—ฐ์–ด์ฒ˜๋ฆฌ ์ž‘์—…์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ์ œ๋กœ ์ƒท(zero-shot) ๋˜๋Š” ํ“จ ์ƒท(few-shot) ํ•™์Šต์ด ๊ฐ€๋Šฅํ•œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(Large Language Models, LLM)์— ๋Œ€ํ•œ ๋งŽ์€ ๊ด€์‹ฌ์ด ๋ฐœ์ƒํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์ด ๋ช…์‹œ์ ์œผ๋กœ ํ›ˆ๋ จ๋˜์ง€ ์•Š์€ ์ž‘์—…๋„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค! ์–ธ์–ด ๋ชจ๋ธ์€ ์œ ์ฐฝํ•˜๊ณ  ์„ค๋“๋ ฅ ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ํ…์ŠคํŠธ๊ฐ€ ํ•ญ์ƒ ์ •ํ™•ํ•˜์ง€๋Š” ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ฃผ์˜๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

์–ธ์–ด ๋ชจ๋ธ๋ง์—๋Š” ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์ด ์žˆ์Šต๋‹ˆ๋‹ค:

  • ์ธ๊ณผ์  ์–ธ์–ด ๋ชจ๋ธ๋ง: ์ด ๋ชจ๋ธ์˜ ๋ชฉ์ ์€ ์‹œํ€€์Šค์—์„œ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ๋ฏธ๋ž˜ ํ† ํฐ์ด ๋งˆ์Šคํ‚น ๋ฉ๋‹ˆ๋‹ค.

    >>> from transformers import pipeline
    
    >>> prompt = "Hugging Face is a community-based open-source platform for machine learning."
    >>> generator = pipeline(task="text-generation")
    >>> generator(prompt)  # doctest: +SKIP
  • ๋งˆ์Šคํ‚น๋œ ์–ธ์–ด ๋ชจ๋ธ๋ง: ์ด ๋ชจ๋ธ์˜ ๋ชฉ์ ์€ ์‹œํ€€์Šค ๋‚ด์˜ ๋งˆ์Šคํ‚น๋œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ์‹œํ€€์Šค ๋‚ด์˜ ๋ชจ๋“  ํ† ํฐ์— ๋Œ€ํ•œ ์ ‘๊ทผ์ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.

    >>> text = "Hugging Face is a community-based open-source <mask> for machine learning."
    >>> fill_mask = pipeline(task="fill-mask")
    >>> preds = fill_mask(text, top_k=1)
    >>> preds = [
    ...     {
    ...         "score": round(pred["score"], 4),
    ...         "token": pred["token"],
    ...         "token_str": pred["token_str"],
    ...         "sequence": pred["sequence"],
    ...     }
    ...     for pred in preds
    ... ]
    >>> preds
    [{'score': 0.2236,
      'token': 1761,
      'token_str': ' platform',
      'sequence': 'Hugging Face is a community-based open-source platform for machine learning.'}]

์ด ํŽ˜์ด์ง€๋ฅผ ํ†ตํ•ด ๊ฐ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์˜ ๋‹ค์–‘ํ•œ ์ž‘์—… ์œ ํ˜•๊ณผ ๊ฐ ์ž‘์—…์˜ ์‹ค์šฉ์  ์ค‘์š”์„ฑ์— ๋Œ€ํ•ด ์ถ”๊ฐ€์ ์ธ ๋ฐฐ๊ฒฝ ์ •๋ณด๋ฅผ ์–ป์œผ์…จ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๋‹ค์Œ ์„น์…˜์—์„œ๋Š” ๐Ÿค— Transformer๊ฐ€ ์ด๋Ÿฌํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์•Œ์•„๋ณด์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

< > Update on GitHub