Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / transformers /main /ko /tasks /visual_question_answering.html

rtrm

28 days ago

download

raw

74 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"시각적 질의응답 (Visual Question Answering)","local":"시각적-질의응답-visual-question-answering","sections":[{"title":"ViLT 미세 조정","local":"finetuning-vilt","sections":[],"depth":2},{"title":"데이터 가져오기","local":"load-the-data","sections":[],"depth":2},{"title":"데이터 전처리","local":"preprocessing-data","sections":[],"depth":2},{"title":"모델 훈련","local":"train-the-model","sections":[],"depth":2},{"title":"추론","local":"inference","sections":[],"depth":2},{"title":"제로샷 VQA","local":"zeroshot-vqa","sections":[],"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.9aa88961.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/scheduler.9bc65507.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/singletons.9eec45c3.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.3b203c72.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/paths.566078f7.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/entry/app.84fb67c3.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.707bf1b6.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/nodes/0.1c99376b.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/84.ab636240.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/CodeBlock.54a9f38d.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/DocNotebookDropdown.41f65cb5.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/globals.7f7f1b26.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/EditOnGithub.922df6ba.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"시각적 질의응답 (Visual Question Answering)","local":"시각적-질의응답-visual-question-answering","sections":[{"title":"ViLT 미세 조정","local":"finetuning-vilt","sections":[],"depth":2},{"title":"데이터 가져오기","local":"load-the-data","sections":[],"depth":2},{"title":"데이터 전처리","local":"preprocessing-data","sections":[],"depth":2},{"title":"모델 훈련","local":"train-the-model","sections":[],"depth":2},{"title":"추론","local":"inference","sections":[],"depth":2},{"title":"제로샷 VQA","local":"zeroshot-vqa","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="시각적-질의응답-visual-question-answering" 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="#시각적-질의응답-visual-question-answering"><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>시각적 질의응답 (Visual Question Answering)</span></h1> <div class="flex space-x-1 absolute z-10 right-0 top-0"> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"> </button> </div> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"> </button> </div></div> <p data-svelte-h="svelte-lsf3d3">시각적 질의응답(VQA)은 이미지를 기반으로 개방형 질문에 대응하는 작업입니다. 이 작업을 지원하는 모델의 입력은 대부분 이미지와 질문의 조합이며, 출력은 자연어로 된 답변입니다.</p> <p data-svelte-h="svelte-1m5x0xt">VQA의 주요 사용 사례는 다음과 같습니다:</p> <ul data-svelte-h="svelte-19c3jz6"><li>시각 장애인을 위한 접근성 애플리케이션을 구축할 수 있습니다.</li> <li>교육: 강의나 교과서에 나온 시각 자료에 대한 질문에 답할 수 있습니다. 또한 체험형 전시와 유적 등에서도 VQA를 활용할 수 있습니다.</li> <li>고객 서비스 및 전자상거래: VQA는 사용자가 제품에 대해 질문할 수 있게 함으로써 사용자 경험을 향상시킬 수 있습니다.</li> <li>이미지 검색: VQA 모델을 사용하여 원하는 특성을 가진 이미지를 검색할 수 있습니다. 예를 들어 사용자는 “강아지가 있어?”라고 물어봐서 주어진 이미지 묶음에서 강아지가 있는 모든 이미지를 받아볼 수 있습니다.</li></ul> <p data-svelte-h="svelte-27ww4z">이 가이드에서 학습할 내용은 다음과 같습니다:</p> <ul data-svelte-h="svelte-4zovkh"><li>VQA 모델 중 하나인 <a href="../../en/model_doc/vilt">ViLT</a>를 <a href="https://huggingface.co/datasets/Graphcore/vqa" rel="nofollow"><code>Graphcore/vqa</code> 데이터셋</a> 에서 미세조정하는 방법</li> <li>미세조정된 ViLT 모델로 추론하는 방법</li> <li>BLIP-2 같은 생성 모델로 제로샷 VQA 추론을 실행하는 방법</li></ul> <h2 class="relative group"><a id="finetuning-vilt" 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="#finetuning-vilt"><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>ViLT 미세 조정</span></h2> <p data-svelte-h="svelte-l5s5vs">ViLT는 Vision Transformer (ViT) 내에 텍스트 임베딩을 포함하여 비전/자연어 사전훈련(VLP; Vision-and-Language Pretraining)을 위한 기본 디자인을 제공합니다.
	ViLT 모델은 비전 트랜스포머(ViT)에 텍스트 임베딩을 넣어 비전/언어 사전훈련(VLP; Vision-and-Language Pre-training)을 위한 기본적인 디자인을 갖췄습니다. 이 모델은 여러 다운스트림 작업에 사용할 수 있습니다. VQA 태스크에서는 (<code>[CLS]</code> 토큰의 최종 은닉 상태 위에 선형 레이어인) 분류 헤더가 있으며 무작위로 초기화됩니다.
	따라서 여기에서 시각적 질의응답은 <strong>분류 문제</strong>로 취급됩니다.</p> <p data-svelte-h="svelte-1gnol8d">최근의 BLIP, BLIP-2, InstructBLIP와 같은 모델들은 VQA를 생성형 작업으로 간주합니다. 가이드의 후반부에서는 이런 모델들을 사용하여 제로샷 VQA 추론을 하는 방법에 대해 설명하겠습니다.</p> <p data-svelte-h="svelte-16n8ks6">시작하기 전 필요한 모든 라이브러리를 설치했는지 확인하세요.</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=""><!-- HTML_TAG_START -->pip install -q transformers datasets<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1tfz9ll">커뮤니티에 모델을 공유하는 것을 권장 드립니다. Hugging Face 계정에 로그인하여 🤗 Hub에 업로드할 수 있습니다.
	메시지가 나타나면 로그인할 토큰을 입력하세요:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> notebook_login

	<span class="hljs-meta">>>> </span>notebook_login()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-8hbytf">모델 체크포인트를 전역 변수로 선언하세요.</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>model_checkpoint = <span class="hljs-string">"dandelin/vilt-b32-mlm"</span><!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="load-the-data" 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="#load-the-data"><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 가져오기</span></h2> <p data-svelte-h="svelte-ogma3f">이 가이드에서는 <code>Graphcore/vqa</code> 데이터세트의 작은 샘플을 사용합니다. 전체 데이터세트는 <a href="https://huggingface.co/datasets/Graphcore/vqa" rel="nofollow">🤗 Hub</a> 에서 확인할 수 있습니다.</p> <p data-svelte-h="svelte-1yqcape"><a href="https://huggingface.co/datasets/Graphcore/vqa" rel="nofollow"><code>Graphcore/vqa</code> 데이터세트</a> 의 대안으로 공식 <a href="https://visualqa.org/download.html" rel="nofollow">VQA 데이터세트 페이지</a> 에서 동일한 데이터를 수동으로 다운로드할 수 있습니다. 직접 공수한 데이터로 튜토리얼을 따르고 싶다면 <a href="https://huggingface.co/docs/datasets/image_dataset#loading-script" rel="nofollow">이미지 데이터세트 만들기</a> 라는
	🤗 Datasets 문서를 참조하세요.</p> <p data-svelte-h="svelte-1e6q69w">검증 데이터의 첫 200개 항목을 불러와 데이터세트의 특성을 확인해 보겠습니다:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset

	<span class="hljs-meta">>>> </span>dataset = load_dataset(<span class="hljs-string">"Graphcore/vqa"</span>, split=<span class="hljs-string">"validation[:200]"</span>)
	<span class="hljs-meta">>>> </span>dataset
	Dataset({
	features: [<span class="hljs-string">'question'</span>, <span class="hljs-string">'question_type'</span>, <span class="hljs-string">'question_id'</span>, <span class="hljs-string">'image_id'</span>, <span class="hljs-string">'answer_type'</span>, <span class="hljs-string">'label'</span>],
	num_rows: <span class="hljs-number">200</span>
	})<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-k3mzt8">예제를 하나 뽑아 데이터세트의 특성을 이해해 보겠습니다.</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>dataset[<span class="hljs-number">0</span>]
	{<span class="hljs-string">'question'</span>: <span class="hljs-string">'Where is he looking?'</span>,
	<span class="hljs-string">'question_type'</span>: <span class="hljs-string">'none of the above'</span>,
	<span class="hljs-string">'question_id'</span>: <span class="hljs-number">262148000</span>,
	<span class="hljs-string">'image_id'</span>: <span class="hljs-string">'/root/.cache/huggingface/datasets/downloads/extracted/ca733e0e000fb2d7a09fbcc94dbfe7b5a30750681d0e965f8e0a23b1c2f98c75/val2014/COCO_val2014_000000262148.jpg'</span>,
	<span class="hljs-string">'answer_type'</span>: <span class="hljs-string">'other'</span>,
	<span class="hljs-string">'label'</span>: {<span class="hljs-string">'ids'</span>: [<span class="hljs-string">'at table'</span>, <span class="hljs-string">'down'</span>, <span class="hljs-string">'skateboard'</span>, <span class="hljs-string">'table'</span>],
	<span class="hljs-string">'weights'</span>: [<span class="hljs-number">0.30000001192092896</span>,
	<span class="hljs-number">1.0</span>,
	<span class="hljs-number">0.30000001192092896</span>,
	<span class="hljs-number">0.30000001192092896</span>]}}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1vwvj02">데이터세트에는 다음과 같은 특성이 포함되어 있습니다:</p> <ul data-svelte-h="svelte-1bz53k7"><li><code>question</code>: 이미지에 대한 질문</li> <li><code>image_id</code>: 질문과 관련된 이미지의 경로</li> <li><code>label</code>: 데이터의 레이블 (annotations)</li></ul> <p data-svelte-h="svelte-1hynqfk">나머지 특성들은 필요하지 않기 때문에 삭제해도 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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>dataset = dataset.remove_columns([<span class="hljs-string">'question_type'</span>, <span class="hljs-string">'question_id'</span>, <span class="hljs-string">'answer_type'</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-6vh6f7">보시다시피 <code>label</code> 특성은 같은 질문마다 답변이 여러 개 있을 수 있습니다. 모두 다른 데이터 라벨러들로부터 수집되었기 때문인데요. 질문의 답변은 주관적일 수 있습니다. 이 경우 질문은 “그는 어디를 보고 있나요?” 였지만, 어떤 사람들은 “아래”로 레이블을 달았고, 다른 사람들은 “테이블” 또는 “스케이트보드” 등으로 주석을 달았습니다.</p> <p data-svelte-h="svelte-1o0kwl9">아래의 이미지를 보고 어떤 답변을 선택할 것인지 생각해 보세요:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image

	<span class="hljs-meta">>>> </span>image = Image.<span class="hljs-built_in">open</span>(dataset[<span class="hljs-number">0</span>][<span class="hljs-string">'image_id'</span>])
	<span class="hljs-meta">>>> </span>image<!-- HTML_TAG_END --></pre></div> <div class="flex justify-center" data-svelte-h="svelte-1tjg4st"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/vqa-example.png" alt="VQA Image Example"></div> <p data-svelte-h="svelte-bmlc3i">질문과 답변의 모호성으로 인해 이러한 데이터세트는 여러 개의 답변이 가능하므로 다중 레이블 분류 문제로 처리됩니다. 게다가, 원핫(one-hot) 인코딩 벡터를 생성하기보다는 레이블에서 특정 답변이 나타나는 횟수를 기반으로 소프트 인코딩을 생성합니다.</p> <p data-svelte-h="svelte-1isdfaj">위의 예시에서 “아래”라는 답변이 다른 답변보다 훨씬 더 자주 선택되었기 때문에 데이터세트에서 <code>weight</code>라고 불리는 점수로 1.0을 가지며, 나머지 답변들은 1.0 미만의 점수를 가집니다.</p> <p data-svelte-h="svelte-1hwga8k">적절한 분류 헤더로 모델을 나중에 인스턴스화하기 위해 레이블을 정수로 매핑한 딕셔너리 하나, 반대로 정수를 레이블로 매핑한 딕셔너리 하나 총 2개의 딕셔너리를 생성하세요:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> itertools

	<span class="hljs-meta">>>> </span>labels = [item[<span class="hljs-string">'ids'</span>] <span class="hljs-keyword">for</span> item <span class="hljs-keyword">in</span> dataset[<span class="hljs-string">'label'</span>]]
	<span class="hljs-meta">>>> </span>flattened_labels = <span class="hljs-built_in">list</span>(itertools.chain(*labels))
	<span class="hljs-meta">>>> </span>unique_labels = <span class="hljs-built_in">list</span>(<span class="hljs-built_in">set</span>(flattened_labels))

	<span class="hljs-meta">>>> </span>label2id = {label: idx <span class="hljs-keyword">for</span> idx, label <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(unique_labels)}
	<span class="hljs-meta">>>> </span>id2label = {idx: label <span class="hljs-keyword">for</span> label, idx <span class="hljs-keyword">in</span> label2id.items()} <!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1f1gsw8">이제 매핑이 완료되었으므로 문자열 답변을 해당 id로 교체하고, 데이터세트의 더 편리한 후처리를 위해 편평화 할 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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">def</span> <span class="hljs-title function_">replace_ids</span>(<span class="hljs-params">inputs</span>):
	<span class="hljs-meta">... </span> inputs[<span class="hljs-string">"label"</span>][<span class="hljs-string">"ids"</span>] = [label2id[x] <span class="hljs-keyword">for</span> x <span class="hljs-keyword">in</span> inputs[<span class="hljs-string">"label"</span>][<span class="hljs-string">"ids"</span>]]
	<span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> inputs


	<span class="hljs-meta">>>> </span>dataset = dataset.<span class="hljs-built_in">map</span>(replace_ids)
	<span class="hljs-meta">>>> </span>flat_dataset = dataset.flatten()
	<span class="hljs-meta">>>> </span>flat_dataset.features
	{<span class="hljs-string">'question'</span>: Value(dtype=<span class="hljs-string">'string'</span>, <span class="hljs-built_in">id</span>=<span class="hljs-literal">None</span>),
	<span class="hljs-string">'image_id'</span>: Value(dtype=<span class="hljs-string">'string'</span>, <span class="hljs-built_in">id</span>=<span class="hljs-literal">None</span>),
	<span class="hljs-string">'label.ids'</span>: <span class="hljs-type">Sequence</span>(feature=Value(dtype=<span class="hljs-string">'int64'</span>, <span class="hljs-built_in">id</span>=<span class="hljs-literal">None</span>), length=-<span class="hljs-number">1</span>, <span class="hljs-built_in">id</span>=<span class="hljs-literal">None</span>),
	<span class="hljs-string">'label.weights'</span>: <span class="hljs-type">Sequence</span>(feature=Value(dtype=<span class="hljs-string">'float64'</span>, <span class="hljs-built_in">id</span>=<span class="hljs-literal">None</span>), length=-<span class="hljs-number">1</span>, <span class="hljs-built_in">id</span>=<span class="hljs-literal">None</span>)}<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="preprocessing-data" 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="#preprocessing-data"><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 전처리</span></h2> <p data-svelte-h="svelte-u0lfur">다음 단계는 모델을 위해 이미지와 텍스트 데이터를 준비하기 위해 ViLT 프로세서를 가져오는 것입니다.
	<code>ViltProcessor</code>는 BERT 토크나이저와 ViLT 이미지 프로세서를 편리하게 하나의 프로세서로 묶습니다:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> ViltProcessor

	<span class="hljs-meta">>>> </span>processor = ViltProcessor.from_pretrained(model_checkpoint)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ihxpp7">데이터를 전처리하려면 이미지와 질문을 <code>ViltProcessor</code>로 인코딩해야 합니다. 프로세서는 <code>BertTokenizerFast</code>로 텍스트를 토크나이즈하고 텍스트 데이터를 위해 <code>input_ids</code>, <code>attention_mask</code> 및 <code>token_type_ids</code>를 생성합니다.
	이미지는 <code>ViltImageProcessor</code>로 이미지를 크기 조정하고 정규화하며, <code>pixel_values</code>와 <code>pixel_mask</code>를 생성합니다.</p> <p data-svelte-h="svelte-8a14gg">이런 전처리 단계는 모두 내부에서 이루어지므로, <code>processor</code>를 호출하기만 하면 됩니다. 하지만 아직 타겟 레이블이 완성되지 않았습니다. 타겟의 표현에서 각 요소는 가능한 답변(레이블)에 해당합니다. 정확한 답변의 요소는 해당 점수(weight)를 유지시키고 나머지 요소는 0으로 설정해야 합니다.</p> <p data-svelte-h="svelte-x4dc7">아래 함수가 위에서 설명한대로 이미지와 질문에 <code>processor</code>를 적용하고 레이블을 형식에 맞춥니다:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> torch

	<span class="hljs-meta">>>> </span><span class="hljs-keyword">def</span> <span class="hljs-title function_">preprocess_data</span>(<span class="hljs-params">examples</span>):
	<span class="hljs-meta">... </span> image_paths = examples[<span class="hljs-string">'image_id'</span>]
	<span class="hljs-meta">... </span> images = [Image.<span class="hljs-built_in">open</span>(image_path) <span class="hljs-keyword">for</span> image_path <span class="hljs-keyword">in</span> image_paths]
	<span class="hljs-meta">... </span> texts = examples[<span class="hljs-string">'question'</span>]

	<span class="hljs-meta">... </span> encoding = processor(images, texts, padding=<span class="hljs-string">"max_length"</span>, truncation=<span class="hljs-literal">True</span>, return_tensors=<span class="hljs-string">"pt"</span>)

	<span class="hljs-meta">... </span> <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> encoding.items():
	<span class="hljs-meta">... </span> encoding[k] = v.squeeze()

	<span class="hljs-meta">... </span> targets = []

	<span class="hljs-meta">... </span> <span class="hljs-keyword">for</span> labels, scores <span class="hljs-keyword">in</span> <span class="hljs-built_in">zip</span>(examples[<span class="hljs-string">'label.ids'</span>], examples[<span class="hljs-string">'label.weights'</span>]):
	<span class="hljs-meta">... </span> target = torch.zeros(<span class="hljs-built_in">len</span>(id2label))

	<span class="hljs-meta">... </span> <span class="hljs-keyword">for</span> label, score <span class="hljs-keyword">in</span> <span class="hljs-built_in">zip</span>(labels, scores):
	<span class="hljs-meta">... </span> target[label] = score

	<span class="hljs-meta">... </span> targets.append(target)

	<span class="hljs-meta">... </span> encoding[<span class="hljs-string">"labels"</span>] = targets

	<span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> encoding<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-y8sr6z">전체 데이터세트에 전처리 함수를 적용하려면 🤗 Datasets의 <code>map</code> 함수를 사용하십시오. <code>batched=True</code>를 설정하여 데이터세트의 여러 요소를 한 번에 처리함으로써 <code>map</code>을 더 빠르게 할 수 있습니다. 이 시점에서 필요하지 않은 열은 제거하세요.</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>processed_dataset = flat_dataset.<span class="hljs-built_in">map</span>(preprocess_data, batched=<span class="hljs-literal">True</span>, remove_columns=[<span class="hljs-string">'question'</span>,<span class="hljs-string">'question_type'</span>, <span class="hljs-string">'question_id'</span>, <span class="hljs-string">'image_id'</span>, <span class="hljs-string">'answer_type'</span>, <span class="hljs-string">'label.ids'</span>, <span class="hljs-string">'label.weights'</span>])
	<span class="hljs-meta">>>> </span>processed_dataset
	Dataset({
	features: [<span class="hljs-string">'input_ids'</span>, <span class="hljs-string">'token_type_ids'</span>, <span class="hljs-string">'attention_mask'</span>, <span class="hljs-string">'pixel_values'</span>, <span class="hljs-string">'pixel_mask'</span>, <span class="hljs-string">'labels'</span>],
	num_rows: <span class="hljs-number">200</span>
	})<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-12etfja">마지막 단계로, <code>DefaultDataCollator</code>를 사용하여 예제로 쓸 배치를 생성하세요:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> DefaultDataCollator

	<span class="hljs-meta">>>> </span>data_collator = DefaultDataCollator()<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="train-the-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="#train-the-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>모델 훈련</span></h2> <p data-svelte-h="svelte-117iw8t">이제 모델을 훈련하기 위해 준비되었습니다! <code>ViltForQuestionAnswering</code>으로 ViLT를 가져올 차례입니다. 레이블의 수와 레이블 매핑을 지정하세요:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> ViltForQuestionAnswering

	<span class="hljs-meta">>>> </span>model = ViltForQuestionAnswering.from_pretrained(model_checkpoint, num_labels=<span class="hljs-built_in">len</span>(id2label), id2label=id2label, label2id=label2id)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-ahfk2n">이 시점에서는 다음 세 단계만 남았습니다:</p> <ol data-svelte-h="svelte-9y82ex"><li><code>TrainingArguments</code>에서 훈련 하이퍼파라미터를 정의하세요:</li></ol> <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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> TrainingArguments

	<span class="hljs-meta">>>> </span>repo_id = <span class="hljs-string">"MariaK/vilt_finetuned_200"</span>

	<span class="hljs-meta">>>> </span>training_args = TrainingArguments(
	<span class="hljs-meta">... </span> output_dir=repo_id,
	<span class="hljs-meta">... </span> per_device_train_batch_size=<span class="hljs-number">4</span>,
	<span class="hljs-meta">... </span> num_train_epochs=<span class="hljs-number">20</span>,
	<span class="hljs-meta">... </span> save_steps=<span class="hljs-number">200</span>,
	<span class="hljs-meta">... </span> logging_steps=<span class="hljs-number">50</span>,
	<span class="hljs-meta">... </span> learning_rate=<span class="hljs-number">5e-5</span>,
	<span class="hljs-meta">... </span> save_total_limit=<span class="hljs-number">2</span>,
	<span class="hljs-meta">... </span> remove_unused_columns=<span class="hljs-literal">False</span>,
	<span class="hljs-meta">... </span> push_to_hub=<span class="hljs-literal">True</span>,
	<span class="hljs-meta">... </span>)<!-- HTML_TAG_END --></pre></div> <ol start="2" data-svelte-h="svelte-18s4qm6"><li>모델, 데이터세트, 프로세서, 데이터 콜레이터와 함께 훈련 인수를 <code>Trainer</code>에 전달하세요:</li></ol> <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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> Trainer

	<span class="hljs-meta">>>> </span>trainer = Trainer(
	<span class="hljs-meta">... </span> model=model,
	<span class="hljs-meta">... </span> args=training_args,
	<span class="hljs-meta">... </span> data_collator=data_collator,
	<span class="hljs-meta">... </span> train_dataset=processed_dataset,
	<span class="hljs-meta">... </span> tokenizer=processor,
	<span class="hljs-meta">... </span>)<!-- HTML_TAG_END --></pre></div> <ol start="3" data-svelte-h="svelte-1jabqg6"><li><code>train()</code>을 호출하여 모델을 미세 조정하세요:</li></ol> <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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>trainer.train() <!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-njd4j0">훈련이 완료되면, <code>push_to_hub()</code> 메소드를 사용하여 🤗 Hub에 모델을 공유하세요:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>trainer.push_to_hub()<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="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="#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>추론</span></h2> <p data-svelte-h="svelte-1uxaimc">ViLT 모델을 미세 조정하고 🤗 Hub에 업로드했다면 추론에 사용할 수 있습니다. 미세 조정된 모델을 추론에 사용해보는 가장 간단한 방법은 <code>Pipeline</code>에서 사용하는 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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> pipeline

	<span class="hljs-meta">>>> </span>pipe = pipeline(<span class="hljs-string">"visual-question-answering"</span>, model=<span class="hljs-string">"MariaK/vilt_finetuned_200"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-2v1cek">이 가이드의 모델은 200개의 예제에서만 훈련되었으므로 그다지 많은 것을 기대할 수는 없습니다. 데이터세트의 첫 번째 예제를 사용하여 추론 결과를 설명해보겠습니다:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>example = dataset[<span class="hljs-number">0</span>]
	<span class="hljs-meta">>>> </span>image = Image.<span class="hljs-built_in">open</span>(example[<span class="hljs-string">'image_id'</span>])
	<span class="hljs-meta">>>> </span>question = example[<span class="hljs-string">'question'</span>]
	<span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(question)
	<span class="hljs-meta">>>> </span>pipe(image, question, top_k=<span class="hljs-number">1</span>)
	<span class="hljs-string">"Where is he looking?"</span>
	[{<span class="hljs-string">'score'</span>: <span class="hljs-number">0.5498199462890625</span>, <span class="hljs-string">'answer'</span>: <span class="hljs-string">'down'</span>}]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-tjyuxs">비록 확신은 별로 없지만, 모델은 실제로 무언가를 배웠습니다. 더 많은 예제와 더 긴 훈련 기간이 주어진다면 분명 더 나은 결과를 얻을 수 있을 것입니다!</p> <p data-svelte-h="svelte-4epvs7">원한다면 파이프라인의 결과를 수동으로 복제할 수도 있습니다:</p> <ol data-svelte-h="svelte-gtxums"><li>이미지와 질문을 가져와서 프로세서를 사용하여 모델에 준비합니다.</li> <li>전처리된 결과를 모델에 전달합니다.</li> <li>로짓에서 가장 가능성 있는 답변의 id를 가져와서 <code>id2label</code>에서 실제 답변을 찾습니다.</li></ol> <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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>processor = ViltProcessor.from_pretrained(<span class="hljs-string">"MariaK/vilt_finetuned_200"</span>)

	<span class="hljs-meta">>>> </span>image = Image.<span class="hljs-built_in">open</span>(example[<span class="hljs-string">'image_id'</span>])
	<span class="hljs-meta">>>> </span>question = example[<span class="hljs-string">'question'</span>]

	<span class="hljs-meta">>>> </span><span class="hljs-comment"># prepare inputs</span>
	<span class="hljs-meta">>>> </span>inputs = processor(image, question, return_tensors=<span class="hljs-string">"pt"</span>)

	<span class="hljs-meta">>>> </span>model = ViltForQuestionAnswering.from_pretrained(<span class="hljs-string">"MariaK/vilt_finetuned_200"</span>)

	<span class="hljs-meta">>>> </span><span class="hljs-comment"># forward pass</span>
	<span class="hljs-meta">>>> </span><span class="hljs-keyword">with</span> torch.no_grad():
	<span class="hljs-meta">... </span> outputs = model(**inputs)

	<span class="hljs-meta">>>> </span>logits = outputs.logits
	<span class="hljs-meta">>>> </span>idx = logits.argmax(-<span class="hljs-number">1</span>).item()
	<span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(<span class="hljs-string">"Predicted answer:"</span>, model.config.id2label[idx])
	Predicted answer: down<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="zeroshot-vqa" 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="#zeroshot-vqa"><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>제로샷 VQA</span></h2> <p data-svelte-h="svelte-y6okgx">이전 모델은 VQA를 분류 문제로 처리했습니다. BLIP, BLIP-2 및 InstructBLIP와 같은 최근의 모델은 VQA를 생성 작업으로 접근합니다. <a href="../../en/model_doc/blip-2">BLIP-2</a>를 예로 들어 보겠습니다. 이 모델은 사전훈련된 비전 인코더와 LLM의 모든 조합을 사용할 수 있는 새로운 비전-자연어 사전 학습 패러다임을 도입했습니다. (<a href="https://huggingface.co/blog/blip-2" rel="nofollow">BLIP-2 블로그 포스트</a>를 통해 더 자세히 알아볼 수 있어요)
	이를 통해 시각적 질의응답을 포함한 여러 비전-자연어 작업에서 SOTA를 달성할 수 있었습니다.</p> <p data-svelte-h="svelte-17wlrgr">이 모델을 어떻게 VQA에 사용할 수 있는지 설명해 보겠습니다. 먼저 모델을 가져와 보겠습니다. 여기서 GPU가 사용 가능한 경우 모델을 명시적으로 GPU로 전송할 것입니다. 이전에는 훈련할 때 쓰지 않은 이유는 <code>Trainer</code>가 이 부분을 자동으로 처리하기 때문입니다:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoProcessor, Blip2ForConditionalGeneration
	<span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> torch

	<span class="hljs-meta">>>> </span>processor = AutoProcessor.from_pretrained(<span class="hljs-string">"Salesforce/blip2-opt-2.7b"</span>)
	<span class="hljs-meta">>>> </span>model = Blip2ForConditionalGeneration.from_pretrained(<span class="hljs-string">"Salesforce/blip2-opt-2.7b"</span>, torch_dtype=torch.float16)
	<span class="hljs-meta">>>> </span>device = <span class="hljs-string">"cuda"</span> <span class="hljs-keyword">if</span> torch.cuda.is_available() <span class="hljs-keyword">else</span> <span class="hljs-string">"cpu"</span>
	<span class="hljs-meta">>>> </span>model.to(device)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-bl55fy">모델은 이미지와 텍스트를 입력으로 받으므로, VQA 데이터세트의 첫 번째 예제에서와 동일한 이미지/질문 쌍을 사용해 보겠습니다:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>example = dataset[<span class="hljs-number">0</span>]
	<span class="hljs-meta">>>> </span>image = Image.<span class="hljs-built_in">open</span>(example[<span class="hljs-string">'image_id'</span>])
	<span class="hljs-meta">>>> </span>question = example[<span class="hljs-string">'question'</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-179kzpa">BLIP-2를 시각적 질의응답 작업에 사용하려면 텍스트 프롬프트가 <code>Question: {} Answer:</code> 형식을 따라야 합니다.</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>prompt = <span class="hljs-string">f"Question: <span class="hljs-subst">{question}</span> Answer:"</span> <!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-les5jw">이제 모델의 프로세서로 이미지/프롬프트를 전처리하고, 처리된 입력을 모델을 통해 전달하고, 출력을 디코드해야 합니다:</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=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>inputs = processor(image, text=prompt, return_tensors=<span class="hljs-string">"pt"</span>).to(device, torch.float16)

	<span class="hljs-meta">>>> </span>generated_ids = model.generate(**inputs, max_new_tokens=<span class="hljs-number">10</span>)
	<span class="hljs-meta">>>> </span>generated_text = processor.batch_decode(generated_ids, skip_special_tokens=<span class="hljs-literal">True</span>)[<span class="hljs-number">0</span>].strip()
	<span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(generated_text)
	<span class="hljs-string">"He is looking at the crowd"</span> <!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1wddc64">보시다시피 모델은 군중을 인식하고, 얼굴의 방향(아래쪽을 보고 있음)을 인식했지만, 군중이 스케이터 뒤에 있다는 사실을 놓쳤습니다. 그러나 사람이 직접 라벨링한 데이터셋을 얻을 수 없는 경우에, 이 접근법은 빠르게 유용한 결과를 생성할 수 있습니다.</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/tasks/visual_question_answering.md" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span data-svelte-h="svelte-x0xyl0">></span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>

	<script>
	{
	__sveltekit_1hrx8 = {
	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.9aa88961.js"),
	import("/docs/transformers/main/ko/_app/immutable/entry/app.84fb67c3.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 84],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 74 kB
Xet hash:: f0179be094ec772bf1d614149c3f19d4b8e03a92b54faea91110a063fbfa1fa8

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