Buckets:

hf-doc-build
/

doc

Files

xet

hf-doc-build/doc / transformers /main /ko /glossary.html

HuggingFaceDocBuilder

about 12 hours ago

download

raw

162 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"용어집(Glossary)","local":"용어집glossary","sections":[{"title":"A","local":"a","sections":[{"title":"어텐션 마스크 (attention mask)","local":"어텐션-마스크-attention-mask","sections":[],"depth":3},{"title":"오토인코딩 모델 (autoencoding models)","local":"오토인코딩-모델-autoencoding-models","sections":[],"depth":3},{"title":"자기회귀 모델 (autoregressive models)","local":"자기회귀-모델-autoregressive-models","sections":[],"depth":3}],"depth":2},{"title":"B","local":"b","sections":[{"title":"백본 (backbone)","local":"백본-backbone","sections":[],"depth":3}],"depth":2},{"title":"C","local":"c","sections":[{"title":"인과적 언어 모델링 (causal language modeling)","local":"인과적-언어-모델링-causal-language-modeling","sections":[],"depth":3},{"title":"채널 (channel)","local":"채널-channel","sections":[],"depth":3},{"title":"연결 시간분류(connectionist temporal classification, CTC)","local":"연결-시간분류connectionist-temporal-classification-ctc","sections":[],"depth":3},{"title":"컨볼루션 (convolution)","local":"컨볼루션-convolution","sections":[],"depth":3}],"depth":2},{"title":"D","local":"d","sections":[{"title":"데이터 병렬화 (DataParallel)","local":"데이터-병렬화-dataparallel","sections":[],"depth":3},{"title":"디코더 입력 ID (decoder input IDs)","local":"디코더-입력-id-decoder-input-ids","sections":[],"depth":3},{"title":"디코더 모델 (decoder models)","local":"디코더-모델-decoder-models","sections":[],"depth":3},{"title":"딥러닝 (deep learning)","local":"딥러닝-deep-learning","sections":[],"depth":3}],"depth":2},{"title":"E","local":"e","sections":[{"title":"인코더 모델 (encoder models)","local":"인코더-모델-encoder-models","sections":[],"depth":3}],"depth":2},{"title":"F","local":"f","sections":[{"title":"특징 추출 (feature extraction)","local":"특징-추출-feature-extraction","sections":[],"depth":3},{"title":"피드 포워드 청킹 (feed forward chunking)","local":"피드-포워드-청킹-feed-forward-chunking","sections":[],"depth":3},{"title":"파인튜닝 모델 (finetuned models)","local":"파인튜닝-모델-finetuned-models","sections":[],"depth":3}],"depth":2},{"title":"H","local":"h","sections":[{"title":"헤드 (head)","local":"헤드-head","sections":[],"depth":3}],"depth":2},{"title":"I","local":"i","sections":[{"title":"이미지 패치 (image patch)","local":"이미지-패치-image-patch","sections":[],"depth":3},{"title":"인퍼런스 (inference)","local":"인퍼런스-inference","sections":[],"depth":3},{"title":"입력 ID (input IDs)","local":"입력-id-input-ids","sections":[],"depth":3}],"depth":2},{"title":"L","local":"l","sections":[{"title":"레이블 (labels)","local":"레이블-labels","sections":[],"depth":3},{"title":"대규모 언어 모델 (LLM)","local":"대규모-언어-모델-llm","sections":[],"depth":3}],"depth":2},{"title":"M","local":"m","sections":[{"title":"마스킹된 언어 모델링 (MLM)","local":"마스킹된-언어-모델링-mlm","sections":[],"depth":3},{"title":"멀티모달 (multimodal)","local":"멀티모달-multimodal","sections":[],"depth":3}],"depth":2},{"title":"N","local":"n","sections":[{"title":"자연어 생성 (NLG)","local":"자연어-생성-nlg","sections":[],"depth":3},{"title":"자연어 처리 (NLP)","local":"자연어-처리-nlp","sections":[],"depth":3},{"title":"자연어 이해 (NLU)","local":"자연어-이해-nlu","sections":[],"depth":3}],"depth":2},{"title":"P","local":"p","sections":[{"title":"파이프라인 (pipeline)","local":"파이프라인-pipeline","sections":[],"depth":3},{"title":"파이프라인 병렬화 (PP)","local":"파이프라인-병렬화-pp","sections":[],"depth":3},{"title":"픽셀 값 (pixel values)","local":"픽셀-값-pixel-values","sections":[],"depth":3},{"title":"풀링 (pooling)","local":"풀링-pooling","sections":[],"depth":3},{"title":"포지션 ID (position IDs)","local":"포지션-id-position-ids","sections":[],"depth":3},{"title":"전처리 (preprocessing)","local":"전처리-preprocessing","sections":[],"depth":3},{"title":"사전 학습된 모델 (pretrained model)","local":"사전-학습된-모델-pretrained-model","sections":[],"depth":3}],"depth":2},{"title":"R","local":"r","sections":[{"title":"순환 신경망 (RNN)","local":"순환-신경망-rnn","sections":[],"depth":3},{"title":"표현학습 (representation learning)","local":"표현학습-representation-learning","sections":[],"depth":3}],"depth":2},{"title":"S","local":"s","sections":[{"title":"샘플링 속도 (sampling rate)","local":"샘플링-속도-sampling-rate","sections":[],"depth":3},{"title":"셀프 어텐션 (self-attention)","local":"셀프-어텐션-self-attention","sections":[],"depth":3},{"title":"자기지도 학습 (self-supervised learning)","local":"자기지도-학습-self-supervised-learning","sections":[],"depth":3},{"title":"준지도 학습 (semi-supervised learning)","local":"준지도-학습-semi-supervised-learning","sections":[],"depth":3},{"title":"시퀀스 투 시퀀스 (seq2seq)","local":"시퀀스-투-시퀀스-seq2seq","sections":[],"depth":3},{"title":"분할 DDP (Sharded DDP)","local":"분할-ddp-sharded-ddp","sections":[],"depth":3},{"title":"스트라이드 (stride)","local":"스트라이드-stride","sections":[],"depth":3},{"title":"지도학습 (supervised learning)","local":"지도학습-supervised-learning","sections":[],"depth":3}],"depth":2},{"title":"T","local":"t","sections":[{"title":"텐서 병렬화 (TP)","local":"텐서-병렬화-tp","sections":[],"depth":3},{"title":"토큰 (token)","local":"토큰-token","sections":[],"depth":3},{"title":"토큰 타입 ID (token type IDs)","local":"토큰-타입-id-token-type-ids","sections":[],"depth":3},{"title":"전이학습 (transfer learning)","local":"전이학습-transfer-learning","sections":[],"depth":3},{"title":"트랜스포머 (transformer)","local":"트랜스포머-transformer","sections":[],"depth":3}],"depth":2},{"title":"U","local":"u","sections":[{"title":"비지도 학습 (unsupervised learning)","local":"비지도-학습-unsupervised-learning","sections":[],"depth":3}],"depth":2},{"title":"Z","local":"z","sections":[{"title":"Zero Redundancy Optimizer (ZeRO)","local":"zero-redundancy-optimizer-zero","sections":[],"depth":3}],"depth":2}],"depth":1}">
	<link href="/docs/transformers/main/ko/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/entry/start.fac1c405.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/scheduler.53228c21.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/singletons.7988aa03.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.e93d0901.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/paths.1e80752f.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/entry/app.fb40223f.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/preload-helper.cb103237.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.3db2ce32.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/nodes/0.eb559a36.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/nodes/20.2a66bfc4.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/Tip.d62b0755.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/CopyLLMTxtMenu.1327b590.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.49b88d99.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/Youtube.b27f3a4a.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/CodeBlock.ada04ea6.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"용어집(Glossary)","local":"용어집glossary","sections":[{"title":"A","local":"a","sections":[{"title":"어텐션 마스크 (attention mask)","local":"어텐션-마스크-attention-mask","sections":[],"depth":3},{"title":"오토인코딩 모델 (autoencoding models)","local":"오토인코딩-모델-autoencoding-models","sections":[],"depth":3},{"title":"자기회귀 모델 (autoregressive models)","local":"자기회귀-모델-autoregressive-models","sections":[],"depth":3}],"depth":2},{"title":"B","local":"b","sections":[{"title":"백본 (backbone)","local":"백본-backbone","sections":[],"depth":3}],"depth":2},{"title":"C","local":"c","sections":[{"title":"인과적 언어 모델링 (causal language modeling)","local":"인과적-언어-모델링-causal-language-modeling","sections":[],"depth":3},{"title":"채널 (channel)","local":"채널-channel","sections":[],"depth":3},{"title":"연결 시간분류(connectionist temporal classification, CTC)","local":"연결-시간분류connectionist-temporal-classification-ctc","sections":[],"depth":3},{"title":"컨볼루션 (convolution)","local":"컨볼루션-convolution","sections":[],"depth":3}],"depth":2},{"title":"D","local":"d","sections":[{"title":"데이터 병렬화 (DataParallel)","local":"데이터-병렬화-dataparallel","sections":[],"depth":3},{"title":"디코더 입력 ID (decoder input IDs)","local":"디코더-입력-id-decoder-input-ids","sections":[],"depth":3},{"title":"디코더 모델 (decoder models)","local":"디코더-모델-decoder-models","sections":[],"depth":3},{"title":"딥러닝 (deep learning)","local":"딥러닝-deep-learning","sections":[],"depth":3}],"depth":2},{"title":"E","local":"e","sections":[{"title":"인코더 모델 (encoder models)","local":"인코더-모델-encoder-models","sections":[],"depth":3}],"depth":2},{"title":"F","local":"f","sections":[{"title":"특징 추출 (feature extraction)","local":"특징-추출-feature-extraction","sections":[],"depth":3},{"title":"피드 포워드 청킹 (feed forward chunking)","local":"피드-포워드-청킹-feed-forward-chunking","sections":[],"depth":3},{"title":"파인튜닝 모델 (finetuned models)","local":"파인튜닝-모델-finetuned-models","sections":[],"depth":3}],"depth":2},{"title":"H","local":"h","sections":[{"title":"헤드 (head)","local":"헤드-head","sections":[],"depth":3}],"depth":2},{"title":"I","local":"i","sections":[{"title":"이미지 패치 (image patch)","local":"이미지-패치-image-patch","sections":[],"depth":3},{"title":"인퍼런스 (inference)","local":"인퍼런스-inference","sections":[],"depth":3},{"title":"입력 ID (input IDs)","local":"입력-id-input-ids","sections":[],"depth":3}],"depth":2},{"title":"L","local":"l","sections":[{"title":"레이블 (labels)","local":"레이블-labels","sections":[],"depth":3},{"title":"대규모 언어 모델 (LLM)","local":"대규모-언어-모델-llm","sections":[],"depth":3}],"depth":2},{"title":"M","local":"m","sections":[{"title":"마스킹된 언어 모델링 (MLM)","local":"마스킹된-언어-모델링-mlm","sections":[],"depth":3},{"title":"멀티모달 (multimodal)","local":"멀티모달-multimodal","sections":[],"depth":3}],"depth":2},{"title":"N","local":"n","sections":[{"title":"자연어 생성 (NLG)","local":"자연어-생성-nlg","sections":[],"depth":3},{"title":"자연어 처리 (NLP)","local":"자연어-처리-nlp","sections":[],"depth":3},{"title":"자연어 이해 (NLU)","local":"자연어-이해-nlu","sections":[],"depth":3}],"depth":2},{"title":"P","local":"p","sections":[{"title":"파이프라인 (pipeline)","local":"파이프라인-pipeline","sections":[],"depth":3},{"title":"파이프라인 병렬화 (PP)","local":"파이프라인-병렬화-pp","sections":[],"depth":3},{"title":"픽셀 값 (pixel values)","local":"픽셀-값-pixel-values","sections":[],"depth":3},{"title":"풀링 (pooling)","local":"풀링-pooling","sections":[],"depth":3},{"title":"포지션 ID (position IDs)","local":"포지션-id-position-ids","sections":[],"depth":3},{"title":"전처리 (preprocessing)","local":"전처리-preprocessing","sections":[],"depth":3},{"title":"사전 학습된 모델 (pretrained model)","local":"사전-학습된-모델-pretrained-model","sections":[],"depth":3}],"depth":2},{"title":"R","local":"r","sections":[{"title":"순환 신경망 (RNN)","local":"순환-신경망-rnn","sections":[],"depth":3},{"title":"표현학습 (representation learning)","local":"표현학습-representation-learning","sections":[],"depth":3}],"depth":2},{"title":"S","local":"s","sections":[{"title":"샘플링 속도 (sampling rate)","local":"샘플링-속도-sampling-rate","sections":[],"depth":3},{"title":"셀프 어텐션 (self-attention)","local":"셀프-어텐션-self-attention","sections":[],"depth":3},{"title":"자기지도 학습 (self-supervised learning)","local":"자기지도-학습-self-supervised-learning","sections":[],"depth":3},{"title":"준지도 학습 (semi-supervised learning)","local":"준지도-학습-semi-supervised-learning","sections":[],"depth":3},{"title":"시퀀스 투 시퀀스 (seq2seq)","local":"시퀀스-투-시퀀스-seq2seq","sections":[],"depth":3},{"title":"분할 DDP (Sharded DDP)","local":"분할-ddp-sharded-ddp","sections":[],"depth":3},{"title":"스트라이드 (stride)","local":"스트라이드-stride","sections":[],"depth":3},{"title":"지도학습 (supervised learning)","local":"지도학습-supervised-learning","sections":[],"depth":3}],"depth":2},{"title":"T","local":"t","sections":[{"title":"텐서 병렬화 (TP)","local":"텐서-병렬화-tp","sections":[],"depth":3},{"title":"토큰 (token)","local":"토큰-token","sections":[],"depth":3},{"title":"토큰 타입 ID (token type IDs)","local":"토큰-타입-id-token-type-ids","sections":[],"depth":3},{"title":"전이학습 (transfer learning)","local":"전이학습-transfer-learning","sections":[],"depth":3},{"title":"트랜스포머 (transformer)","local":"트랜스포머-transformer","sections":[],"depth":3}],"depth":2},{"title":"U","local":"u","sections":[{"title":"비지도 학습 (unsupervised learning)","local":"비지도-학습-unsupervised-learning","sections":[],"depth":3}],"depth":2},{"title":"Z","local":"z","sections":[{"title":"Zero Redundancy Optimizer (ZeRO)","local":"zero-redundancy-optimizer-zero","sections":[],"depth":3}],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg></span> <span>Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg></button></div> </div> <h1 class="relative group"><a id="용어집glossary" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#용어집glossary"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>용어집(Glossary)</span></h1> <p data-svelte-h="svelte-10603yx">이 용어집은 전반적인 머신러닝 및 🤗 Transformers 관련 용어를 정의하여 문서를 더 잘 이해하는 데 도움을 줍니다.</p> <h2 class="relative group"><a id="a" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#a"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>A</span></h2> <h3 class="relative group"><a id="어텐션-마스크-attention-mask" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#어텐션-마스크-attention-mask"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>어텐션 마스크 (attention mask)</span></h3> <p data-svelte-h="svelte-6k3bwk">어텐션 마스크(attention mask)는 여러 시퀀스를 배치(batch)로 처리할 때 사용되는 선택적 인자입니다.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/M6adb1j2jPI" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-pz8mtv">이 인자는 모델에게 어떤 토큰에 주의를 기울여야 하는지, 그리고 어떤 토큰은 무시해야 하는지를 알려줍니다.</p> <p data-svelte-h="svelte-we7nak">예를 들어, 다음 두 개의 시퀀스가 있다고 가정해 봅시다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BertTokenizer

	<span class="hljs-meta">>>> </span>tokenizer = BertTokenizer.from_pretrained(<span class="hljs-string">"google-bert/bert-base-cased"</span>)

	<span class="hljs-meta">>>> </span>sequence_a = <span class="hljs-string">"This is a short sequence."</span>
	<span class="hljs-meta">>>> </span>sequence_b = <span class="hljs-string">"This is a rather long sequence. It is at least longer than the sequence A."</span>

	<span class="hljs-meta">>>> </span>encoded_sequence_a = tokenizer(sequence_a)[<span class="hljs-string">"input_ids"</span>]
	<span class="hljs-meta">>>> </span>encoded_sequence_b = tokenizer(sequence_b)[<span class="hljs-string">"input_ids"</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-7pyee8">인코딩된 버전들의 길이가 다릅니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-built_in">len</span>(encoded_sequence_a), <span class="hljs-built_in">len</span>(encoded_sequence_b)
	(<span class="hljs-number">8</span>, <span class="hljs-number">19</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1xbtvp2">따라서 이 두 시퀀스를 그대로 하나의 텐서에 넣을 수는 없습니다. 첫 번째 시퀀스를 두 번째 길이에 맞춰 패딩 하거나, 반대로 두 번째 시퀀스를 첫 번째 길이에 맞춰 잘라내야 합니다.</p> <p data-svelte-h="svelte-xy2kd1">첫 번째 경우에는 ID 목록이 패딩 인덱스로 확장됩니다. 이렇게 패딩을 적용하려면 토크나이저에 리스트를 전달하고 다음과 같이 요청할 수 있습니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>padded_sequences = tokenizer([sequence_a, sequence_b], padding=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-16qcd97">첫 번째 문장 오른쪽에 0이 추가되어 두 번째 문장과 길이가 같아진 것을 볼 수 있습니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>padded_sequences[<span class="hljs-string">"input_ids"</span>]
	[[<span class="hljs-number">101</span>, <span class="hljs-number">1188</span>, <span class="hljs-number">1110</span>, <span class="hljs-number">170</span>, <span class="hljs-number">1603</span>, <span class="hljs-number">4954</span>, <span class="hljs-number">119</span>, <span class="hljs-number">102</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>], [<span class="hljs-number">101</span>, <span class="hljs-number">1188</span>, <span class="hljs-number">1110</span>, <span class="hljs-number">170</span>, <span class="hljs-number">1897</span>, <span class="hljs-number">1263</span>, <span class="hljs-number">4954</span>, <span class="hljs-number">119</span>, <span class="hljs-number">1135</span>, <span class="hljs-number">1110</span>, <span class="hljs-number">1120</span>, <span class="hljs-number">1655</span>, <span class="hljs-number">2039</span>, <span class="hljs-number">1190</span>, <span class="hljs-number">1103</span>, <span class="hljs-number">4954</span>, <span class="hljs-number">138</span>, <span class="hljs-number">119</span>, <span class="hljs-number">102</span>]]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-9lhnw">이것은 PyTorch나 TensorFlow의 텐서로 변환될 수 있습니다. 어텐션 마스크는 모델이 패딩 된 인덱스를 참조하지 않도록 해당 위치를 나타내는 이진 텐서입니다. <a href="/docs/transformers/main/ko/model_doc/bert#transformers.BertTokenizer">BertTokenizer</a>의 경우, <code>1</code>은 어텐션이 필요한 값을 나타내고, <code>0</code>은 패딩 된 값을 나타냅니다. 이 어텐션 마스크는 토크나이저가 반환되는 딕셔너리의 “attention_mask” 키 아래에 포함되어 있습니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>padded_sequences[<span class="hljs-string">"attention_mask"</span>]
	[[<span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>], [<span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>]]<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="오토인코딩-모델-autoencoding-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#오토인코딩-모델-autoencoding-models"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>오토인코딩 모델 (autoencoding models)</span></h3> <p data-svelte-h="svelte-1g57jvw"><a href="#encoder-models">인코더 모델</a>과 <a href="#masked-language-modeling-mlm">마스킹된 언어 모델링</a>을 참고하세요.</p> <h3 class="relative group"><a id="자기회귀-모델-autoregressive-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#자기회귀-모델-autoregressive-models"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>자기회귀 모델 (autoregressive models)</span></h3> <p data-svelte-h="svelte-38j25d"><a href="#causal-language-modeling">인과적 언어 모델링</a>과 <a href="#decoder-models">디코더 모델</a>을 참고하세요.</p> <h2 class="relative group"><a id="b" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#b"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>B</span></h2> <h3 class="relative group"><a id="백본-backbone" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#백본-backbone"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>백본 (backbone)</span></h3> <p data-svelte-h="svelte-2jv5ue">백본(backbone)은 원시(hidden) 은닉 상태(hidden state) 또는 특징(feature)을 출력하는 네트워크(임베딩과 레이어)입니다. 일반적으로 이 백본은 해당 특징을 입력으로 받아 예측을 수행하는 <a href="#head">헤드</a>와 연결됩니다. 예를 들어, <a href="/docs/transformers/main/ko/model_doc/vit#transformers.ViTModel">ViTModel</a>은 특정 헤드가 없는 백본입니다. 다른 모델들도<code>VitModel</code>을 백본으로 사용할 수 있으며, <a href="model_doc/dpt">DPT</a>등이 그 예시입니다.</p> <h2 class="relative group"><a id="c" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#c"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>C</span></h2> <h3 class="relative group"><a id="인과적-언어-모델링-causal-language-modeling" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#인과적-언어-모델링-causal-language-modeling"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>인과적 언어 모델링 (causal language modeling)</span></h3> <p data-svelte-h="svelte-bprfr1">모델이 텍스트를 순서대로 읽으며 다음 단어를 예측해야 하는 사전 학습(pretraining) 작업입니다. 일반적으로 문장을 전체로 읽되, 모델 내부에서 특징 시점 이후의 토큰을 마스킹(masking)하여 다음 단어를 예측하게 됩니다.</p> <h3 class="relative group"><a id="채널-channel" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#채널-channel"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>채널 (channel)</span></h3> <p data-svelte-h="svelte-1vylj3v">컬러 이미지는 빨간색(R), 초록색(G), 파란색(B)의 세 채널 값을 조합하여 구성되며, 흑백 이미지는 단일 채널만을 가집니다. 🤗 Transformers에서는 이미지 텐서의 채널이 첫 번째 또는 마지막 차원에 위치할 수 있습니다:[<code>n_channels</code>, <code>height</code>, <code>width</code>] 또는 [<code>height</code>, <code>width</code>, <code>n_channels</code>]와 같은 형식입니다.</p> <h3 class="relative group"><a id="연결-시간분류connectionist-temporal-classification-ctc" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#연결-시간분류connectionist-temporal-classification-ctc"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>연결 시간분류(connectionist temporal classification, CTC)</span></h3> <p data-svelte-h="svelte-1tly9lk">입력과 출력의 정렬 상태를 정확히 몰라도 모델이 학습할 수 있도록 돕는 알고리즘입니다. CTC는 주어진 입력에 대해 가능한 모든 출력의 확률 분포를 계산하고, 그중 가장 가능성이 높은 출력을 선택합니다. CTC는 말하는 속도의 차이 등 여러 이유로 음성과 텍스트가 항상 정확하게 일치하지 않기 때문에 음성 인식 작업에서 자주 사용됩니다.</p> <h3 class="relative group"><a id="컨볼루션-convolution" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#컨볼루션-convolution"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>컨볼루션 (convolution)</span></h3> <p data-svelte-h="svelte-22ouow">신경망에서 사용되는 레이어의 한 종류로, 입력 행렬에 대해 더 작은 행렬(커널 또는 필터)을 원소별로 곱한 뒤 그 값을 합산해 새로운 행렬을 만드는 연산입니다. 이 연산을 컨볼루션 연산이라고 하며, 입력 행렬 전체에 걸쳐 반복적으로 수행됩니다. 각 연산은 입력 행렬의 서로 다른 구간에 적용됩니다. 컨볼루션 신경망(CNN)은 컴퓨터 비전 분야에서 널리 사용됩니다.</p> <h2 class="relative group"><a id="d" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#d"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>D</span></h2> <h3 class="relative group"><a id="데이터-병렬화-dataparallel" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#데이터-병렬화-dataparallel"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>데이터 병렬화 (DataParallel)</span></h3> <p data-svelte-h="svelte-150svwp">여러 개의 GPU에서 훈련을 수행할 때 사용하는 병렬화 기법으로, 동일한 모델 구성이 여러 번 복제되며 각 인스턴스는 서로 다른 데이터 조각을 받습니다. 모든 인스턴스는 병렬로 처리를 수행하며, 각 훈련 단계가 끝난 후 결과를 동기화합니다.</p> <p data-svelte-h="svelte-1bhb3a2">DataParallel 방식에 대해 더 알아보려면 <a href="perf_train_gpu_many#dataparallel-vs-distributeddataparallel">여기</a>를 참고하세요.</p> <h3 class="relative group"><a id="디코더-입력-id-decoder-input-ids" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#디코더-입력-id-decoder-input-ids"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>디코더 입력 ID (decoder input IDs)</span></h3> <p data-svelte-h="svelte-mwg3i2">이 입력은 인코더-디코더 모델에 특화된 것으로, 디코더에 전달될 input ID 들을 포함합니다. 이러한 입력은 번역이나 요약과 같은 시퀀스-투-시퀀스(sequence-to-sequence) 작업에 사용되며, 일반적으로 모델마다 고유한 방식으로 구성됩니다.</p> <p data-svelte-h="svelte-16860sq">대부분의 인코더-디코더 모델(BART, T5 등)은 <code>labels</code>로부터 자동으로 <code>decoder_input_ids</code>를 생성합니다. 이러한 모델에서는 학습 시 <code>labels</code>를 전달하는 것이 일반적으로 권장됩니다.</p> <p data-svelte-h="svelte-14cayds">시퀀스-투-시퀀스 학습에서 각 모델이 이러한 input ID를 어떻게 처리하는지는 모델 문서를 참고하시기를 바랍니다.</p> <h3 class="relative group"><a id="디코더-모델-decoder-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#디코더-모델-decoder-models"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>디코더 모델 (decoder models)</span></h3> <p data-svelte-h="svelte-17irnog">자기회귀 모델(Autoregressive models)이라고도 불리는 디코더 모델은 인과 언어 모델링(causal language modeling)이라 불리는 사전 학습 작업을 수행합니다. 이 작업에서는 모델이 텍스트를 순서대로 읽고 다음 단어를 예측해야 합니다. 일반적으로 문장의 전체를 읽되, 특정 시점 이후의 토큰은 마스크로 가려 예측하게 합니다.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/d_ixlCubqQw" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <h3 class="relative group"><a id="딥러닝-deep-learning" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#딥러닝-deep-learning"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>딥러닝 (deep learning)</span></h3> <p data-svelte-h="svelte-11c8qyy">여러 층의 신경망(neural network)을 사용하는 머신러닝 알고리즘입니다.</p> <h2 class="relative group"><a id="e" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#e"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>E</span></h2> <h3 class="relative group"><a id="인코더-모델-encoder-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#인코더-모델-encoder-models"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>인코더 모델 (encoder models)</span></h3> <p data-svelte-h="svelte-1p9pyd3">자동 인코딩 모델(Autoencoding models)이라고도 불리는 인코더 모델은 텍스트나 이미지와 같은 입력을 받아 임베딩이라 불리는 압축된 수치 표현으로 반환합니다. 일반적으로 인코더 모델은 입력 시퀀스의 일부를 마스킹하고 더 의미 있는 표현을 생성하도록 학습하는 <a href="#masked-language-modeling-mlm">masked language modeling</a>과 같은 기술을 사용하여 사전 학습됩니다.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/H39Z_720T5s" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <h2 class="relative group"><a id="f" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#f"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>F</span></h2> <h3 class="relative group"><a id="특징-추출-feature-extraction" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#특징-추출-feature-extraction"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>특징 추출 (feature extraction)</span></h3> <p data-svelte-h="svelte-19httxw">머신러닝 알고리즘이 더 효과적으로 학습할 수 있도록, 원시 데이터를 선택하고 변환하여 더 유용한 특징(feature) 집합으로 만드는 과정입니다. 예를 들어, 원시 텍스트를 워드 임베딩으로 변환하거나 이미지나 비디오 데이터에서 윤곽선이나 형태와 같은 중요한 특징을 추출하는 것이 있습니다.</p> <h3 class="relative group"><a id="피드-포워드-청킹-feed-forward-chunking" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#피드-포워드-청킹-feed-forward-chunking"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>피드 포워드 청킹 (feed forward chunking)</span></h3> <p data-svelte-h="svelte-135hzqn">트랜스포머의 각 residual attention Block에서는 self-Attention Layer 다음에 보통 두 개의 Feed Forward Layer가 이어집니다. 이 Feed Forward Layers의 중간 임베딩 크기는 종종 모델의 히든 사이즈(hidden size)보다 큽니다(예:
	<code>google-bert/bert-base-uncased</code> 모델의 경우).</p> <p data-svelte-h="svelte-1cut73q">입력 크기가 <code>[batch_size, sequence_length]</code>일 경우, 중간 Feed Forward 임베딩
	<code>[batch_size, sequence_length, config.intermediate_size]</code>을 저장하는 데 필요한 메모리는 전체 메모리 사용량의 큰 부분을 차지할 수 있습니다.
	<a href="https://huggingface.co/papers/2001.04451" rel="nofollow">Reformer: The Efficient Transformer</a> 논문의 저자들은 이 연산이 <code>sequence_length</code> 차원에 대해 독립적이기 때문에,토큰마다 Feed Forward Layer의 출력 임베딩을 각 토큰별로 <code>[batch_size, config.hidden_size]</code>을 개별적으로 계산한 뒤, 이를 이어 붙여 <code>[batch_size, sequence_length, config.hidden_size]</code> 형태로 만들 수 있습니다.<code>n = sequence_length</code>. 이 방식은 계산 시간은 늘어나지만, 메모리 사용량은 줄어들게 됩니다.</p> <p data-svelte-h="svelte-10llef"><a href="/docs/transformers/main/ko/internal/modeling_utils#transformers.apply_chunking_to_forward">apply_chunking_to_forward()</a> 함수를 사용하는 모델의 경우, <code>chunk_size</code>는 병렬로 계산되는 출력 임베딩의 개수를 정의하며, 이는 메모리 사용량과 계산 시간 간의 트레이드오프를 결정합니다.
	<code>chunk_size</code>가 0으로 설정되면, 피드 포워드 청킹(Feed Forward Chunking)은 수행되지 않습니다.</p> <h3 class="relative group"><a id="파인튜닝-모델-finetuned-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#파인튜닝-모델-finetuned-models"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>파인튜닝 모델 (finetuned models)</span></h3> <p data-svelte-h="svelte-ejdek1">파인튜닝(Finetuning)은 전이 학습(transfer learning)의 한 형태로, 사전 학습된 (pretrained) 모델을 사용하여 가중치를 고정(freeze)하고, 출력층을 새롭게 추가된 <a href="#head">모델 헤드</a>로 교체한 뒤, 해당 모델 헤드를 목표 데이터셋에 맞게 학습시키는 방식입니다.</p> <p data-svelte-h="svelte-1ocw32o">자세한 내용은 <a href="https://huggingface.co/docs/transformers/training" rel="nofollow">Fine-tune a pretrained model</a> 튜토리얼을 참고하시고, 🤗 Transformers를 사용해 모델을 파인 튜닝하는 방법도 함께 확인해 보세요.</p> <h2 class="relative group"><a id="h" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#h"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>H</span></h2> <h3 class="relative group"><a id="헤드-head" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#헤드-head"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>헤드 (head)</span></h3> <p data-svelte-h="svelte-1nvkrqh">모델 헤드(model head)란 신경망의 마지막 층을 의미하며, 이 층은 이전 층에서 나온 히든 상태(hidden states)를 받아 다른 차원으로 변환합니다. 각 작업(task)에 따라 서로 다른 모델 헤드가 사용됩니다. 예를 들어:</p> <ul data-svelte-h="svelte-1f4pbyz"><li><a href="/docs/transformers/main/ko/model_doc/gpt2#transformers.GPT2ForSequenceClassification">GPT2ForSequenceClassification</a>은 기본 <a href="/docs/transformers/main/ko/model_doc/gpt2#transformers.GPT2Model">GPT2Model</a> 위에 시퀀스 분류를 위한 선형계층(linear layer)을 추가한 모델 헤드입니다.</li> <li><a href="/docs/transformers/main/ko/model_doc/vit#transformers.ViTForImageClassification">ViTForImageClassification</a>은 이미지 분류를 위한 모델 헤드로, 기본 <a href="/docs/transformers/main/ko/model_doc/vit#transformers.ViTModel">ViTModel</a> 위에 <code>CLS</code> 토큰의 마지막 히든 상태에 선형 계층(linear layer)을 추가한 구조입니다.</li> <li><code>Wav2Vec2ForCTC</code>는 기본 <code>Wav2Vec2Model</code> 위에 <a href="#connectionist-temporal-classification-ctc">CTC</a>를 적용한 언어 모델링 헤드입니다.</li></ul> <h2 class="relative group"><a id="i" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#i"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>I</span></h2> <h3 class="relative group"><a id="이미지-패치-image-patch" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#이미지-패치-image-patch"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>이미지 패치 (image patch)</span></h3> <p data-svelte-h="svelte-a6rxhs">비전 기반 Transformer 모델은 이미지를 작은 패치로 분할한 후, 각 패치를 선형 임베딩하여 시퀀스로 모델에 입력합니다. 모델의 구성 파일에서 <code>patch_size</code>(또는 해상도)를 확인할 수 있습니다.</p> <h3 class="relative group"><a id="인퍼런스-inference" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#인퍼런스-inference"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>인퍼런스 (inference)</span></h3> <p data-svelte-h="svelte-8ju112">인퍼런스는 학습이 완료된 모델에 새로운 데이터를 입력하여 예측을 수행하는 과정입니다. 🤗 Transformer에서 인퍼런스를 수행하는 방법은 <a href="https://huggingface.co/docs/transformers/pipeline_tutorial" rel="nofollow">Pipeline for inference</a> 튜토리얼을 참고하세요.</p> <h3 class="relative group"><a id="입력-id-input-ids" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#입력-id-input-ids"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>입력 ID (input IDs)</span></h3> <p data-svelte-h="svelte-1exvk7z">입력 ID는 종종 모델에 입력으로 전달해야 하는 유일한 필수 파라미터입니다. 이들은 토큰의 인덱스로, 모델이 입력으로 사용할 시퀀스를 구성하는 토큰들의 숫자 표현입니다.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/VFp38yj8h3A" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-s0hsuz">토크나이저마다 작동 방식은 다르지만, 기본 메커니즘은 동일합니다. 다음은 <a href="https://huggingface.co/papers/1609.08144" rel="nofollow">WordPiece</a> 토크나이저인 BERT 토크나이저를 사용한 예시입니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BertTokenizer

	<span class="hljs-meta">>>> </span>tokenizer = BertTokenizer.from_pretrained(<span class="hljs-string">"google-bert/bert-base-cased"</span>)

	<span class="hljs-meta">>>> </span>sequence = <span class="hljs-string">"A Titan RTX has 24GB of VRAM"</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1wfvp85">토크나이저는 시퀀스를 토크나이저의 토큰 목록에 있는 항목으로 분리합니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>tokenized_sequence = tokenizer.tokenize(sequence)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-13q6nrh">토큰은 단어이거나 서브 워드(subword)입니다. 예를 들어, “VRAM”은 모델의 어휘 사전에 없는 단어이기 때문에 “V”, “RA”, “M”으로 나뉘었습니다. 이 토큰들이 개별 단어가 아니라 같은 단어의 일부임을 나타내기 위해 “RA”와 “M” 앞에 더블 해시(<code>##</code>)가 추가 됩니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(tokenized_sequence)
	[<span class="hljs-string">'A'</span>, <span class="hljs-string">'Titan'</span>, <span class="hljs-string">'R'</span>, <span class="hljs-string">'##T'</span>, <span class="hljs-string">'##X'</span>, <span class="hljs-string">'has'</span>, <span class="hljs-string">'24'</span>, <span class="hljs-string">'##GB'</span>, <span class="hljs-string">'of'</span>, <span class="hljs-string">'V'</span>, <span class="hljs-string">'##RA'</span>, <span class="hljs-string">'##M'</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ighplf">이러한 토큰들은 모델이 이해할 수 있는 ID로 변환될 수 있습니다. 이 과정은 문장을 바로 토크나이저에 입력함으로써 수행되며, 성능 최적화를 위해 <a href="https://github.com/huggingface/tokenizers" rel="nofollow">🤗 Tokenizers</a>의 Rust 구현을 활용합니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>inputs = tokenizer(sequence)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-179xfhk">토크나이저는 해당 모델이 올바르게 작동하는 데 필요한 모든 인자를 포함한 딕셔너리를 반환합니다. 토큰 인덱스는 <code>input_ids</code>라는 키에 저장됩니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>encoded_sequence = inputs[<span class="hljs-string">"input_ids"</span>]
	<span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(encoded_sequence)
	[<span class="hljs-number">101</span>, <span class="hljs-number">138</span>, <span class="hljs-number">18696</span>, <span class="hljs-number">155</span>, <span class="hljs-number">1942</span>, <span class="hljs-number">3190</span>, <span class="hljs-number">1144</span>, <span class="hljs-number">1572</span>, <span class="hljs-number">13745</span>, <span class="hljs-number">1104</span>, <span class="hljs-number">159</span>, <span class="hljs-number">9664</span>, <span class="hljs-number">2107</span>, <span class="hljs-number">102</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-p195jj">토크나이저는 (연결된 모델이 이를 사용하는 경우) 자동으로 “특수 토큰”을 추가합니다. 이들은 모델이 특정 상황에서 사용하는 특별한 ID입니다.</p> <p data-svelte-h="svelte-1y6398">이전의 ID 시퀀스를 디코딩하면,</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>decoded_sequence = tokenizer.decode(encoded_sequence)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1o08719">우리는 다음과 같은 결과를 보게 될 것입니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(decoded_sequence)
	[CLS] A Titan RTX has 24GB of VRAM [SEP]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-p9k9wc">이는 <a href="/docs/transformers/main/ko/model_doc/bert#transformers.BertModel">BertModel</a>이 입력값을 기대하는 방식이기 때문입니다.</p> <h2 class="relative group"><a id="l" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#l"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>L</span></h2> <h3 class="relative group"><a id="레이블-labels" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#레이블-labels"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>레이블 (labels)</span></h3> <p data-svelte-h="svelte-1yb1em1">레이블은 모델이 손실(loss)을 직접 계산할 수 있도록 전달되는 선택적 인자입니다. 이 레이블은 모델이 예측해야 할 정답 값을 의미하며, 모델은 예측값과 이 정답(label) 사이의 차이를 표준 손실 함수를 이용해 계산하게 됩니다.</p> <p data-svelte-h="svelte-1a6jcfm">이 레이블(label)의 형태는 모델 헤드(model head)의 종류에 따라 달라집니다. 예를 들어:</p> <ul data-svelte-h="svelte-1k8nc94"><li>시퀀스 분류 모델(<a href="/docs/transformers/main/ko/model_doc/bert#transformers.BertForSequenceClassification">BertForSequenceClassification</a> 등)의 경우, 모델은
	<code>(batch_size)</code> 차원의 텐서를 입력으로 받으며, 배치의 각 값은 전체 시퀀스에 대한 예상 레이블을 나타냅니다.</li> <li>토큰 분류 모델(<a href="/docs/transformers/main/ko/model_doc/bert#transformers.BertForTokenClassification">BertForTokenClassification</a> 등)의 경우, 모델은 <code>(batch_size, seq_length)</code> 차원의 텐서를 입력으로 받으며, 각 값은 개별 토큰에 대한 예상 레이블을 나타냅니다.</li> <li>마스킹 언어 모델(<a href="/docs/transformers/main/ko/model_doc/bert#transformers.BertForMaskedLM">BertForMaskedLM</a>)의 경우, 모델은 <code>(batch_size,seq_length)</code> 차원의 텐서를 입력으로 받으며, 각 값은 개별 토큰에 대한 예상 레이블을 나타냅니다. 레이블은 마스킹 된 토큰의 토큰 ID이며, 나머지 토큰에 대해서는 무시할 값을 사용합니다(일반적으로 -100).</li> <li>시퀀스 투 시퀀스 작업(<a href="/docs/transformers/main/ko/model_doc/bart#transformers.BartForConditionalGeneration">BartForConditionalGeneration</a>, <code>MBartForConditionalGeneration</code>등)의 경우, 모델은 <code>(batch_size, tgt_seq_length)</code> 차원의 텐서를 입력으로 받으며, 각 값은 입력 시퀀스에 대응하는 타겟 시퀀스를 나타냅니다. 학습 중에는 BART와 T5가 적절한 <code>decoder_input_ids</code>와 디코더 attention 마스크를 내부적으로 생성하므로, 일반적으로 따로 제공할 필요가 없습니다. 단, 이는 Encoder-Decoder 프레임워크를 직접 활용하는 모델에는 적용되지 않습니다.</li> <li>이미지 분류 모델(<a href="/docs/transformers/main/ko/model_doc/vit#transformers.ViTForImageClassification">ViTForImageClassification</a> 등)의 경우, 모델은 <code>(batch_size)</code> 차원의 텐서를 입력으로 받으며, 배치의 각 값은 개별 이미지에 대한 예상 레이블을 나타냅니다.</li> <li>시멘틱 세그멘테이션 모델(<code>SegformerForSemanticSegmentation</code> 등)의 경우, 모델은 <code>(batch_size, height, width)</code> 차원의 텐서를 입력으로 받으며, 배치의 각 값은 개별 픽셀에 대한 예상 레이블을 나타냅니다.</li> <li>객체 탐지 모델(<code>DetrForObjectDetection</code> 등)의 경우, 모델은 <code>class_labels</code>와 <code>boxes</code> 키를 포함하는 딕셔너리들의 리스트를 입력으로 받습니다. 배치의 각 값은 개별 이미지에 대한 예상 클래스 레이블과 바운딩 박스 정보를 나타냅니다.</li> <li>자동 음성 인식 모델(<code>Wav2Vec2ForCTC</code> 등)의 경우 모델은 <code>(batch_size,target_length)</code> 차원의 텐서를 입력으로 받으며, 각 값은 개별 토큰에 대한 예상 레이블을 나타냅니다.</li></ul> <blockquote class="tip"><p data-svelte-h="svelte-1pxtqt1">모델마다 요구하는 레이블 형식이 다를 수 있으므로, 각 모델의 문서를 확인하여 해당 모델에 맞는 레이블 형식을 반드시 확인하세요!</p></blockquote> <p data-svelte-h="svelte-12go13m">기본 모델(<a href="/docs/transformers/main/ko/model_doc/bert#transformers.BertModel">BertModel</a> 등)은 레이블을 입력으로 받지 않습니다. 이러한 모델은 단순히 특징(feature)을 출력하는 기본 트랜스포머 모델이기 때문입니다.</p> <h3 class="relative group"><a id="대규모-언어-모델-llm" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#대규모-언어-모델-llm"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>대규모 언어 모델 (LLM)</span></h3> <p data-svelte-h="svelte-5parey">대규모 데이터로 학습된 트랜스포머 언어 모델(GPT-3, BLOOM, OPT 등)을 지칭하는 일반적인 용어입니다. 이러한 모델은 학습할 수 있는 파라미터(parameter)의 수가 매우 많으며, 예를 들어 GPT-3는 약 1,750억 개의 파라미터를 가지고 있습니다.</p> <h2 class="relative group"><a id="m" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#m"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>M</span></h2> <h3 class="relative group"><a id="마스킹된-언어-모델링-mlm" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#마스킹된-언어-모델링-mlm"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>마스킹된 언어 모델링 (MLM)</span></h3> <p data-svelte-h="svelte-1s7p38z">사전 학습 단계 중 하나로, 모델은 일부 토큰이 무작위로 마스킹 된 손상된 문장을 입력받고, 원래의 문장을 예측해야 합니다.</p> <h3 class="relative group"><a id="멀티모달-multimodal" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#멀티모달-multimodal"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>멀티모달 (multimodal)</span></h3> <p data-svelte-h="svelte-18zfg45">텍스트와 이미지와 같은 다른 형태의 입력을 함께 사용하는 작업입니다.</p> <h2 class="relative group"><a id="n" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#n"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>N</span></h2> <h3 class="relative group"><a id="자연어-생성-nlg" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#자연어-생성-nlg"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>자연어 생성 (NLG)</span></h3> <p data-svelte-h="svelte-28egyw">텍스트를 생성하는 모든 작업을 의미합니다. (예: <a href="https://transformer.huggingface.co/" rel="nofollow">Write With Transformers</a>, 번역 등).</p> <h3 class="relative group"><a id="자연어-처리-nlp" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#자연어-처리-nlp"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>자연어 처리 (NLP)</span></h3> <p data-svelte-h="svelte-1swsdr">텍스트를 다루는 작업 전반을 지칭하는 일반적인 용어입니다.</p> <h3 class="relative group"><a id="자연어-이해-nlu" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#자연어-이해-nlu"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>자연어 이해 (NLU)</span></h3> <p data-svelte-h="svelte-cnhqjm">텍스트에 담긴 의미를 이해하는 모든 작업을 포함합니다. (예: 전체 문서 분류, 개별 단어 분류 등).</p> <h2 class="relative group"><a id="p" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#p"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>P</span></h2> <h3 class="relative group"><a id="파이프라인-pipeline" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#파이프라인-pipeline"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>파이프라인 (pipeline)</span></h3> <p data-svelte-h="svelte-9f8al2">🤗 Transformers에서 파이프라인은 데이터를 전처리하고 변환한 후, 모델을 통해 예측값을 반환하는 일련의 단계를 순차적으로 수행하는 추상화된 개념입니다. 파이프라인에 포함될 수 있는 단계로는 데이터 전처리, 특징 추출(feature extraction), 정규화(normalization) 등이 있습니다.</p> <p data-svelte-h="svelte-61t41p">자세한 내용은 <a href="https://huggingface.co/docs/transformers/pipeline_tutorial" rel="nofollow">Pipelines for inference</a> 문서를 참고하세요.</p> <h3 class="relative group"><a id="파이프라인-병렬화-pp" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#파이프라인-병렬화-pp"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>파이프라인 병렬화 (PP)</span></h3> <p data-svelte-h="svelte-11vmbba">모델을 수직 방향(레이어 단위)으로 여러 GPU에 분할하여 병렬로 처리하는 병렬화 기법입니다. 각 GPU는 모델의 하나 또는 여러 개의 레이어만을 담당하며, 전체 파이프라인의 서로 다른 단계를 병렬로 처리하게 됩니다. 또한 각 GPU는 배치(batch)의 일부 작은 조각만 처리합니다. Pipeline Parallel 방식에 대해 더 알아보려면 <a href="perf_train_gpu_many#from-naive-model-parallelism-to-pipeline-parallelism">이 문서</a>를 참고하세요.</p> <h3 class="relative group"><a id="픽셀-값-pixel-values" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#픽셀-값-pixel-values"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>픽셀 값 (pixel values)</span></h3> <p data-svelte-h="svelte-1n9ur0d">이미지를 수치상으로 표현한 텐서로, 모델에 입력으로 전달됩니다. 이 텐서는 이미지 프로세서를 통해 생성되면, 값은 [<code>batch_size</code>, <code>num_channels</code>, <code>height</code>, <code>width</code>] 형태의 차원을 가집니다.</p> <h3 class="relative group"><a id="풀링-pooling" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#풀링-pooling"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>풀링 (pooling)</span></h3> <p data-svelte-h="svelte-1iwtghh">행렬의 특정 차원에서 최댓값이나 평균값을 취하여 더 작은 행렬로 줄이는 연산입니다. 풀링 계층은 주로 합성곱 계층 사이에 위치하여 특징 표현을 다운샘플링 하는 데 사용됩니다.</p> <h3 class="relative group"><a id="포지션-id-position-ids" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#포지션-id-position-ids"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>포지션 ID (position IDs)</span></h3> <p data-svelte-h="svelte-1fsozoj">RNN 모델과 달리 트랜스포머는 각 토큰의 위치 정보를 내부적으로 가지고 있지 않습니다. 따라서 모델은 <code>position_ids</code>를 사용하여 각 토큰이 시퀀스 내에서 어느 위치에 있는지를 인식합니다. 이 값은 선택적인 파라미터입니다. 모델에 <code>position_ids</code>를 전달하지 않으면, 절대 위치 임베딩 방식으로 자동 생성됩니다. 절대 위치 임베딩은 <code>[0, config.max_position_embeddings - 1]</code> 범위 내에서 선택됩니다. 일부 모델은 사인파 형태의 위치 임베딩(sinusoidal position embeddings) 또는 상대 위치 임베딩(relative position embeddings)과 같은 다른 유형의 위치 임베딩을 사용하기도 합니다.</p> <h3 class="relative group"><a id="전처리-preprocessing" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#전처리-preprocessing"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>전처리 (preprocessing)</span></h3> <p data-svelte-h="svelte-q9hqm0">머신러닝 모델이 쉽게 처리할 수 있도록 가공되지 않은 데이터를 정제하는 작업입니다. 예를 들어, 텍스트는 일반적으로 토큰화(tokenization) 과정을 거칩니다. 다른 입력 유형에 대한 전처리 방식이 궁금하다면 <a href="https://huggingface.co/docs/transformers/preprocessing" rel="nofollow">Preprocess</a> 튜토리얼을 참고해 보세요.</p> <h3 class="relative group"><a id="사전-학습된-모델-pretrained-model" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#사전-학습된-모델-pretrained-model"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>사전 학습된 모델 (pretrained model)</span></h3> <p data-svelte-h="svelte-3mh1r0">일부 데이터(예: 위키피디아 전체)로 사전 학습(pretraining)된 모델입니다. 사전 학습은 자기 지도 학습(self-supervised learning)의 목표를 포함하며, 예를 들어 문장을 읽고 다음 단어를 예측하거나 (<a href="#causal-language-modeling">causal language modeling</a>) 참고, 일부 단어를 마스킹하고 이를 예측하는 방식(<a href="#masked-language-modeling-mlm">masked language modeling</a>)이 있습니다.</p> <p data-svelte-h="svelte-1xyniwl">음성 및 비전 모델은 고유의 사전 학습 목표를 가지고 있습니다. 예를 들어, Wav2Vec2는 음성 표현 중 “진짜”를 “가짜” 중에서 구분하는 대조 학습(contrastive learning) 방식으로 사전 학습된 음성 모델입니다. 반면, BEiT는 이미지 패치 중 일부를 마스킹하고 이를 예측하는 마스킹 이미지 모델링 방식으로 사전 학습된 비전 모델입니다. 이는 마스킹 언어 모델링과 유사한 방식입니다.</p> <h2 class="relative group"><a id="r" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#r"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>R</span></h2> <h3 class="relative group"><a id="순환-신경망-rnn" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#순환-신경망-rnn"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>순환 신경망 (RNN)</span></h3> <p data-svelte-h="svelte-1ghvolf">텍스트와 같은 시퀀스 데이터를 처리하기 위해 레이어에 반복 구조(루프)를 사용하는 신경망 모델의 한 종류입니다.</p> <h3 class="relative group"><a id="표현학습-representation-learning" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#표현학습-representation-learning"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>표현학습 (representation learning)</span></h3> <p data-svelte-h="svelte-1kje9yb">머신러닝의 하위 분야로, 원시 데이터로부터 의미 있는 표현을 학습하는 데 중점을 둡니다. 대표적인 기법으로는 단어 임베딩, 오토인코더(autoencoder), 생성적 적대 신경망(GAN) 등이 있습니다.</p> <h2 class="relative group"><a id="s" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#s"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>S</span></h2> <h3 class="relative group"><a id="샘플링-속도-sampling-rate" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#샘플링-속도-sampling-rate"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>샘플링 속도 (sampling rate)</span></h3> <p data-svelte-h="svelte-1vnlrda">샘플링 속도는 1초에 추출하는 (오디오 신호) 샘플의 개수를 헤르츠(Hz) 단위로 나타낸 측정값입니다. 이는 음성처럼 연속적인 신호를 디지털화하여 이산적인 형태로 만드는 결과입니다.</p> <h3 class="relative group"><a id="셀프-어텐션-self-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#셀프-어텐션-self-attention"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>셀프 어텐션 (self-attention)</span></h3> <p data-svelte-h="svelte-1axcrbb">입력의 각 요소가 다른 어떤 요소에 주목해야 하는지를 스스로 판단하는 메커니즘입니다. 이는 모델이 문장에서 특정 단어만을 보는 것이 아니라, 다른 단어들과의 관계를 고려하여 어떤 정보에 더 집중해야 할지를 학습하게 합니다.</p> <h3 class="relative group"><a id="자기지도-학습-self-supervised-learning" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#자기지도-학습-self-supervised-learning"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>자기지도 학습 (self-supervised learning)</span></h3> <p data-svelte-h="svelte-1lebyoz">레이블이 없는 데이터로부터 모델이 스스로 학습 목표를 정의하여 학습하는 머신러닝 기법의 한 종류입니다. <a href="#unsupervised-learning">비지도 학습</a>이나 <a href="#supervised-learning">지도 학습</a>과 달리, 학습 과정 자체는 감독 방식 되지만, 라벨이 명시적으로 주어지는 것은 아닙니다.</p> <p data-svelte-h="svelte-v4vdp2">예시로는 <a href="#masked-language-modeling-mlm">마스크 언어 모델링</a>이 있으며, 이는 문장의 일부 토큰을 제거한 상태로 모델에 입력하고, 모델이 해당 토큰을 예측하도록 학습하는 방식입니다.</p> <h3 class="relative group"><a id="준지도-학습-semi-supervised-learning" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#준지도-학습-semi-supervised-learning"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>준지도 학습 (semi-supervised learning)</span></h3> <p data-svelte-h="svelte-o8znp3">소량의 라벨이 달린 데이터와 대량의 라벨이 없는 데이터를 함께 사용하여 모델의 정확도를 높이는 머신러닝 훈련 기법의 넓은 범주입니다. 이는 <a href="#supervised-learning">지도 학습</a>이나 <a href="#unsupervised-learning">비지도 학습</a>과는 다른 방식입니다.</p> <p data-svelte-h="svelte-1yja8qb">준지도 학습 기법의 예로는 “자기 학습(self-training)“이 있습니다. 이 방식은 먼저 라벨이 있는 데이터로 모델을 학습시키고, 그 모델을 사용해 라벨이 없는 데이터에 대한 예측을 수행합니다. 모델이 가장 높은 확신을 가지고 예측한 라벨이 없는 데이터 일부를 라벨이 있는 데이터로 추가하고, 이를 통해 모델을 다시 학습시킵니다.</p> <h3 class="relative group"><a id="시퀀스-투-시퀀스-seq2seq" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#시퀀스-투-시퀀스-seq2seq"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>시퀀스 투 시퀀스 (seq2seq)</span></h3> <p data-svelte-h="svelte-1nihi38">입력으로부터 새로운 시퀀스를 생성하는 모델입니다. 예를 들어 번역 모델이나 요약 모델이 이에 해당하며, 대표적인 예로는 <a href="model_doc/bart">Bart</a>나<a href="model_doc/t5">T5</a> 모델이 있습니다.</p> <h3 class="relative group"><a id="분할-ddp-sharded-ddp" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#분할-ddp-sharded-ddp"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>분할 DDP (Sharded DDP)</span></h3> <p data-svelte-h="svelte-1kdvnte"><a href="#zero-redundancy-optimizer-zero">ZeRO</a> 개념을 기반으로 다양한 구현에서 사용되는 다른 이름으로 불립니다.</p> <h3 class="relative group"><a id="스트라이드-stride" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#스트라이드-stride"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>스트라이드 (stride)</span></h3> <p data-svelte-h="svelte-2the5i"><a href="#convolution">convolution</a> 또는 <a href="#pooling">pooling</a>에서 스트라이드(stride)는 커널이 행렬 위를 이동하는 간격을 의미합니다. 스트라이드가 1이면 커널이 한 픽셀씩 이동하고, 2이면 두 픽셀씩 이동합니다.</p> <h3 class="relative group"><a id="지도학습-supervised-learning" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#지도학습-supervised-learning"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>지도학습 (supervised learning)</span></h3> <p data-svelte-h="svelte-162uuae">정답이 포함된 라벨링된 데이터를 직접 사용하여 모델의 성능을 개선하는 학습 방식입니다. 학습 중인 모델에 데이터를 입력하고, 예측 결과를 정답과 비교하여 오차를 계산합니다. 모델은 이 오차를 기반으로 가중치를 업데이트하며, 이러한 과정을 반복하여 성능을 최적화합니다.</p> <h2 class="relative group"><a id="t" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#t"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>T</span></h2> <h3 class="relative group"><a id="텐서-병렬화-tp" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#텐서-병렬화-tp"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>텐서 병렬화 (TP)</span></h3> <p data-svelte-h="svelte-1qmulqd">여러 GPU에서 훈련하기 위한 병렬화 기법으로, 각 텐서를 여러 덩어리(chunk)로 나눕니다. 따라서 전체 텐서가 단일 GPU에 상주하는 대신, 텐서의 각 조각(shard)이 지정된 GPU에 상주하게 됩니다. 이 조각들은 각각 다른 GPU에서 개별적으로 병렬 처리되며, 처리 단계가 끝날 때 결과가 동기화됩니다. 이러한 분할이 수평 방향으로 일어나기 때문에, 이는 때때로 수평적 병렬화라고 불립니다. Tensor Parallelism에 대해 더 알아보려면 <a href="perf_train_gpu_many#tensor-parallelism">여기</a>를 참고하세요.</p> <h3 class="relative group"><a id="토큰-token" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#토큰-token"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>토큰 (token)</span></h3> <p data-svelte-h="svelte-rb6dft">일반적인 단어 단위이지만, 때에 따라 서브 워드(자주 사용되지 않는 단어는 서브 워드로 분리됨)나 문장 부호도 포함될 수 있는 문장의 구성 요소입니다.</p> <h3 class="relative group"><a id="토큰-타입-id-token-type-ids" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#토큰-타입-id-token-type-ids"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>토큰 타입 ID (token type IDs)</span></h3> <p data-svelte-h="svelte-u39i49">일부 모델은 문장 쌍 분류나 질의 응답 작업을 수행하는 데 사용됩니다.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/0u3ioSwev3s" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-1bxrv6l">이러한 작업에서는 두 개의 서로 다른 시퀀스를 하나의 “input_ids” 항목으로 결합해야 하며, 일반적으로 <code>[CLS]</code> 분류용 및 <code>[SEP]</code> 구분용과 같은 특수 토큰을 사용하여 처리합니다. 예를 들어, BERT 모델은 두 개의 시퀀스를 다음과 같은 방식으로 구성합니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-comment"># [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-4ix337">두 개의 시퀀스를 <code>tokenizer</code>에 리스트가 아닌 개별 인자로 전달하면, 토크나이저가 자동으로 이러한 문장을 생성해 줍니다. 예시는 다음과 같습니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BertTokenizer

	<span class="hljs-meta">>>> </span>tokenizer = BertTokenizer.from_pretrained(<span class="hljs-string">"google-bert/bert-base-cased"</span>)
	<span class="hljs-meta">>>> </span>sequence_a = <span class="hljs-string">"HuggingFace is based in NYC"</span>
	<span class="hljs-meta">>>> </span>sequence_b = <span class="hljs-string">"Where is HuggingFace based?"</span>

	<span class="hljs-meta">>>> </span>encoded_dict = tokenizer(sequence_a, sequence_b)
	<span class="hljs-meta">>>> </span>decoded = tokenizer.decode(encoded_dict[<span class="hljs-string">"input_ids"</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-16lu7ds">결과는 아래와 같습니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(decoded)
	[CLS] HuggingFace <span class="hljs-keyword">is</span> based <span class="hljs-keyword">in</span> NYC [SEP] Where <span class="hljs-keyword">is</span> HuggingFace based? [SEP]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-12fbvt8">이 코드는 일부 모델이 두 개의 시퀀스를 어떻게 구분하는지 이해하는 데 충분합니다. 그러나 BERT와 같은 다른 모델은 토큰 타입 ID(또는 세그먼트 ID)를 추가로 사용합니다. 이 ID는 0과 1로 구성된 이진 마스크로, 두 시퀀스를 구분하는 역할을 합니다.</p> <p data-svelte-h="svelte-1pj3vk">토크나이저는 이 마스크를 “token_type_id” 항목으로 반환합니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-python "><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>encoded_dict[<span class="hljs-string">"token_type_ids"</span>]
	[<span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-q2c612">질문에 사용되는 첫 번째 시퀀스인 “context”는 모든 토큰이 <code>0</code>으로 표시됩니다. 반면 두 번째 시퀀스인 “question”은 모든 토큰이 <code>1</code>로 표시됩니다.</p> <p data-svelte-h="svelte-160d5tn">일부 모델(예: <code>XLNetModel</code>)은 <code>2</code>로 표시되는 추가 토큰을 사용하기도 합니다.</p> <h3 class="relative group"><a id="전이학습-transfer-learning" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#전이학습-transfer-learning"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>전이학습 (transfer learning)</span></h3> <p data-svelte-h="svelte-usqo96">사전 학습된(pretrained) 모델을 가져와 특정 작업에 맞는 데이터셋에 대해 추가 학습하는 기술입니다. 모델을 처음부터 학습시키는 대신, 기존 모델이 학습한 지식을 출발점으로 삼아 더욱 빠르게 학습할 수 있습니다. 이를 통해 학습 속도를 높이고 필요한 데이터양도 줄일 수 있습니다.</p> <h3 class="relative group"><a id="트랜스포머-transformer" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#트랜스포머-transformer"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>트랜스포머 (transformer)</span></h3> <p data-svelte-h="svelte-8sw1pe">셀프 어텐션 메커니즘을 기반으로 한 딥러닝 모델 아키텍처입니다.</p> <h2 class="relative group"><a id="u" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#u"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>U</span></h2> <h3 class="relative group"><a id="비지도-학습-unsupervised-learning" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#비지도-학습-unsupervised-learning"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>비지도 학습 (unsupervised learning)</span></h3> <p data-svelte-h="svelte-11urs62">정답(레이블)이 포함되지 않은 데이터를 이용해 모델을 학습시키는 방식입니다. 비지도 학습은 데이터 분포의 통계적 특성을 활용해 유용한 패턴을 찾아냅니다.</p> <h2 class="relative group"><a id="z" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#z"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Z</span></h2> <h3 class="relative group"><a id="zero-redundancy-optimizer-zero" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#zero-redundancy-optimizer-zero"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Zero Redundancy Optimizer (ZeRO)</span></h3> <p data-svelte-h="svelte-aw7h73"><a href="#tensor-parallelism-tp">TensorParallel</a>과 유사하게 텐서를 샤딩(sharding)하는 병렬 처리 기법이지만, 순전파(forward)나 역전파(backward) 계산 시점에 전체 텐서를 다시 복원한다는 점에서 차이가 있습니다. 따라서 모델 자체를 수정할 필요가 없습니다. 이 방법은 GPU 메모리가 부족할 경우 이를 보완하기 위한 다양한 오프로딩 (offloading) 기법도 지원합니다.
	ZeRO에 대해 더 알아보려면 <a href="perf_train_gpu_many#zero-data-parallelism">이 문서</a>를 참고하세요.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/transformers/blob/main/docs/source/ko/glossary.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg> <span data-svelte-h="svelte-zjs2n5"><span class="underline">Update</span> on GitHub</span></a> <p></p>

	<script>
	{
	__sveltekit_1hot10d = {
	assets: "/docs/transformers/main/ko",
	base: "/docs/transformers/main/ko",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/transformers/main/ko/_app/immutable/entry/start.fac1c405.js"),
	import("/docs/transformers/main/ko/_app/immutable/entry/app.fb40223f.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 20],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 162 kB
Xet hash:: b1bae858d238b80e39e5f1ba0512a879c5f302e175f73d8de557525e63792a9f

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.