Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /ko /training /custom_diffusion.html
mishig's picture
mishig
download
raw
59.6 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Custom Diffusion&quot;,&quot;local&quot;:&quot;custom-diffusion&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;스크립트 매개변수&quot;,&quot;local&quot;:&quot;script-parameters&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;사전 보존 손실&quot;,&quot;local&quot;:&quot;prior-preservation-loss&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;정규화&quot;,&quot;local&quot;:&quot;regularization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;학습 스크립트&quot;,&quot;local&quot;:&quot;training-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;스크립트 실행&quot;,&quot;local&quot;:&quot;launch-the-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;다음 단계&quot;,&quot;local&quot;:&quot;next-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
<link href="/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/I9eooN77.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bupkjr9f.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CyvZ3FS5.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/l0QnnOED.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DnVIJfC5.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DteTvmQZ.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CoSrdQax.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/366.V3_Yhunl.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/D5eijdnr.js" rel="modulepreload">
<!--1gvvfk7--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Custom Diffusion&quot;,&quot;local&quot;:&quot;custom-diffusion&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Script parameters&quot;,&quot;local&quot;:&quot;script-parameters&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Prior preservation loss&quot;,&quot;local&quot;:&quot;prior-preservation-loss&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Regularization&quot;,&quot;local&quot;:&quot;regularization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Training script&quot;,&quot;local&quot;:&quot;training-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Launch the script&quot;,&quot;local&quot;:&quot;launch-the-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Next steps&quot;,&quot;local&quot;:&quot;next-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
<link href="/docs/diffusers/main/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;"><div dir="ltr" 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 dir="auto">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> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="custom-diffusion" 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="#custom-diffusion"><span><svg 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><!---->Custom Diffusion<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/papers/2212.04488" rel="nofollow">Custom Diffusion</a>은 이미지 생성 모델을 개인화하기 위한 학습 기법입니다. Textual Inversion, DreamBooth, LoRA와 마찬가지로 Custom Diffusion은 소수의 (~4-5개) 예시 이미지만 필요합니다. 이 기법은 cross-attention 레이어의 가중치만 학습하는 방식으로 동작하며, 새로 학습된 개념을 나타내기 위해 특수한 단어를 사용합니다. Custom Diffusion은 여러 개념을 동시에 학습할 수도 있다는 점에서 독특합니다.</p> <p>vRAM이 제한된 GPU에서 학습하는 경우, <code>--enable_xformers_memory_efficient_attention</code>을 사용하여 xFormers를 활성화하면 vRAM 요구 사항(16GB)을 낮추고 더 빠르게 학습할 수 있습니다. 메모리를 더 절약하려면 학습 인자에 <code>--set_grads_to_none</code>을 추가하여 그라디언트를 0 대신 <code>None</code>으로 설정하세요(이 옵션은 일부 문제를 일으킬 수 있으므로 문제가 발생하면 이 매개변수를 제거해 보세요).</p> <p>이 가이드에서는 <a href="https://github.com/huggingface/diffusers/blob/main/examples/custom_diffusion/train_custom_diffusion.py" rel="nofollow">train_custom_diffusion.py</a> 스크립트를 살펴보고, 이를 더 잘 이해하며 자신의 사용 사례에 맞게 수정하는 방법을 안내합니다.</p> <p>…</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->git <span class="hljs-built_in">clone</span> https://github.com/huggingface/diffusers
<span class="hljs-built_in">cd</span> diffusers
pip install .<!----></pre></div><!----> <p>학습 스크립트가 포함된 예제 폴더로 이동하고 필요한 의존성을 설치합니다:</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">cd</span> examples/custom_diffusion
pip install -r requirements.txt
pip install clip-retrieval<!----></pre></div><!----> <blockquote class="tip"><p>🤗 Accelerate는 여러 GPU/TPU 또는 혼합 정밀도로 학습하는 것을 돕는 라이브러리입니다. 하드웨어와 환경에 따라 학습 설정을 자동으로 구성합니다. 자세한 내용은 🤗 Accelerate <a href="https://huggingface.co/docs/accelerate/quicktour" rel="nofollow">Quick tour</a>를 참고하세요.</p></blockquote> <p>🤗 Accelerate 환경을 초기화합니다:</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate config<!----></pre></div><!----> <p>어떤 구성도 선택하지 않고 기본 🤗 Accelerate 환경을 설정하려면:</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate config default<!----></pre></div><!----> <p>또는 노트북과 같이 인터랙티브 셸을 지원하지 않는 환경에서는 다음을 사용할 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 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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> accelerate.utils <span class="hljs-keyword">import</span> write_basic_config
write_basic_config()<!----></pre></div><!----> <p>마지막으로, 자체 데이터셋으로 모델을 학습하려면 <a href="create_dataset">학습용 데이터셋 만들기</a> 가이드를 참고하여 학습 스크립트와 호환되는 데이터셋을 만드는 방법을 배워보세요.</p> <blockquote class="tip"><p>다음 섹션에서는 스크립트를 수정하는 방법을 이해하는 데 중요한 학습 스크립트의 일부 부분을 강조합니다. 그러나 스크립트의 모든 측면을 자세히 다루지는 않습니다. 더 자세히 알고 싶으시다면 <a href="https://github.com/huggingface/diffusers/blob/main/examples/custom_diffusion/train_custom_diffusion.py" rel="nofollow">스크립트</a>를 읽어보시고, 질문이나 우려 사항이 있으시면 알려주시기 바랍니다.</p></blockquote> <!--[1--><h2 class="relative group"><a id="script-parameters" 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="#script-parameters"><span><svg 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><!---->스크립트 매개변수<!----></span></h2><!--]--><!----> <p>학습 스크립트에는 학습 실행을 사용자 정의하는 데 도움이 되는 모든 매개변수가 포함되어 있습니다. 이들은 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/custom_diffusion/train_custom_diffusion.py#L319" rel="nofollow"><code>parse_args()</code></a> 함수에서 찾을 수 있습니다. 이 함수에는 기본값이 제공되지만, 원하시는 경우 학습 명령에서 자체 값을 설정할 수도 있습니다.</p> <p>예를 들어, 입력 이미지의 해상도를 변경하려면:</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_custom_diffusion.py \
--resolution=256<!----></pre></div><!----> <p>기본 매개변수 중 상당수는 <a href="dreambooth#script-parameters">DreamBooth</a> 학습 가이드에 설명되어 있으므로, 이 가이드에서는 Custom Diffusion에 고유한 매개변수에 초점을 맞춥니다:</p> <ul><li><code>--freeze_model</code>: cross-attention 레이어의 key와 value 파라미터를 동결합니다. 기본값은 <code>crossattn_kv</code>이며, cross-attention 레이어의 모든 파라미터를 학습하려면 <code>crossattn</code>으로 설정할 수 있습니다.</li> <li><code>--concepts_list</code>: 여러 개념을 학습하려면 개념을 포함하는 JSON 파일의 경로를 제공합니다</li> <li><code>--modifier_token</code>: 학습된 개념을 나타내는 데 사용되는 특수 단어</li> <li><code>--initializer_token</code>: <code>modifier_token</code>의 임베딩을 초기화하는 데 사용되는 특수 단어</li></ul> <!--[2--><h3 class="relative group"><a id="prior-preservation-loss" 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="#prior-preservation-loss"><span><svg 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 보존 손실<!----></span></h3><!--]--><!----> <p>Prior preservation loss는 모델이 스스로 생성한 샘플을 활용하여 더 다양한 이미지를 생성하는 방법을 학습하도록 돕는 기법입니다. 이러한 생성된 샘플 이미지는 사용자가 제공한 이미지와 동일한 클래스에 속하므로, 모델이 해당 클래스에 대해 학습한 내용을 유지하고 이미 알고 있는 클래스 지식을 활용하여 새로운 구성을 만들어내는 데 도움이 됩니다.</p> <p>prior preservation loss의 많은 매개변수는 <a href="dreambooth#prior-preservation-loss">DreamBooth</a> 학습 가이드에 설명되어 있습니다.</p> <!--[2--><h3 class="relative group"><a id="regularization" 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="#regularization"><span><svg 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><!---->정규화<!----></span></h3><!--]--><!----> <p>Custom Diffusion은 과적합을 방지하기 위해 소수의 실제 이미지로 대상 이미지를 학습하는 것을 포함합니다. 단 몇 장의 이미지로만 학습하는 경우이므로, 이는 쉽게 수행할 수 있습니다. <code>clip_retrieval</code>을 사용하여 실제 이미지 200장을 다운로드하세요. <code>class_prompt</code>는 대상 이미지와 동일한 카테고리여야 합니다. 이 이미지들은 <code>class_data_dir</code>에 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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->python retrieve.py --class_prompt <span class="hljs-built_in">cat</span> --class_data_dir real_reg/samples_cat --num_class_images 200<!----></pre></div><!----> <p>정규화를 활성화하려면 다음 매개변수를 추가하세요:</p> <ul><li><code>--with_prior_preservation</code>: prior preservation loss를 사용할지 여부</li> <li><code>--prior_loss_weight</code>: 모델에 대한 사전 보존 손실의 영향을 제어합니다</li> <li><code>--real_prior</code>: 과적합을 방지하기 위해 실제 이미지의 작은 집합을 사용할지 여부</li></ul> <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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_custom_diffusion.py \
--with_prior_preservation \
--prior_loss_weight=1.0 \
--class_data_dir=<span class="hljs-string">&quot;./real_reg/samples_cat&quot;</span> \
--class_prompt=<span class="hljs-string">&quot;cat&quot;</span> \
--real_prior=True \<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="training-script" 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="#training-script"><span><svg 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><!---->학습 스크립트<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>Custom Diffusion 학습 스크립트의 많은 코드는 <a href="dreambooth#training-script">DreamBooth</a> 스크립트와 유사합니다. 이 가이드는 대신 Custom Diffusion와 관련된 코드에 초점을 맞춥니다.</p></blockquote> <p>Custom Diffusion 학습 스크립트에는 두 가지 데이터셋 클래스가 있습니다:</p> <ul><li><a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/custom_diffusion/train_custom_diffusion.py#L165" rel="nofollow"><code>CustomDiffusionDataset</code></a>: 학습을 위해 이미지, 클래스 이미지, 프롬프트를 전처리합니다</li> <li><a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/custom_diffusion/train_custom_diffusion.py#L148" rel="nofollow"><code>PromptDataset</code></a>: 클래스 이미지 생성을 위한 프롬프트를 준비합니다</li></ul> <p>다음으로, <code>modifier_token</code>은 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/custom_diffusion/train_custom_diffusion.py#L811" rel="nofollow">토크나이저에 추가</a>되고 토큰 ID로 변환되며, 새로운 <code>modifier_token</code>을 반영하도록 토큰 임베딩의 크기가 조정됩니다. 그런 다음 <code>modifier_token</code>의 임베딩은 <code>initializer_token</code>의 임베딩으로 초기화됩니다. 텍스트 인코더의 모든 매개변수는 동결되지만, 모델이 개념과 연관 지어 학습하려는 대상이 토큰 임베딩이므로 토큰 임베딩만 예외입니다.</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!---->params_to_freeze = itertools.chain(
text_encoder.text_model.encoder.parameters(),
text_encoder.text_model.final_layer_norm.parameters(),
text_encoder.text_model.embeddings.position_embedding.parameters(),
)
freeze_params(params_to_freeze)<!----></pre></div><!----> <p>이제 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/custom_diffusion/train_custom_diffusion.py#L911C3-L911C3" rel="nofollow">Custom Diffusion weights</a>를 attention 레이어에 추가해야 합니다. 이는 attention weights의 shape와 size를 올바르게 설정하고 각 UNet 블록에 적절한 수의 attention processors를 설정하는 데 매우 중요한 단계입니다.</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!---->st = unet.state_dict()
<span class="hljs-keyword">for</span> name, _ <span class="hljs-keyword">in</span> unet.attn_processors.items():
cross_attention_dim = <span class="hljs-literal">None</span> <span class="hljs-keyword">if</span> name.endswith(<span class="hljs-string">&quot;attn1.processor&quot;</span>) <span class="hljs-keyword">else</span> unet.config.cross_attention_dim
<span class="hljs-keyword">if</span> name.startswith(<span class="hljs-string">&quot;mid_block&quot;</span>):
hidden_size = unet.config.block_out_channels[-<span class="hljs-number">1</span>]
<span class="hljs-keyword">elif</span> name.startswith(<span class="hljs-string">&quot;up_blocks&quot;</span>):
block_id = <span class="hljs-built_in">int</span>(name[<span class="hljs-built_in">len</span>(<span class="hljs-string">&quot;up_blocks.&quot;</span>)])
hidden_size = <span class="hljs-built_in">list</span>(<span class="hljs-built_in">reversed</span>(unet.config.block_out_channels))[block_id]
<span class="hljs-keyword">elif</span> name.startswith(<span class="hljs-string">&quot;down_blocks&quot;</span>):
block_id = <span class="hljs-built_in">int</span>(name[<span class="hljs-built_in">len</span>(<span class="hljs-string">&quot;down_blocks.&quot;</span>)])
hidden_size = unet.config.block_out_channels[block_id]
layer_name = name.split(<span class="hljs-string">&quot;.processor&quot;</span>)[<span class="hljs-number">0</span>]
weights = {
<span class="hljs-string">&quot;to_k_custom_diffusion.weight&quot;</span>: st[layer_name + <span class="hljs-string">&quot;.to_k.weight&quot;</span>],
<span class="hljs-string">&quot;to_v_custom_diffusion.weight&quot;</span>: st[layer_name + <span class="hljs-string">&quot;.to_v.weight&quot;</span>],
}
<span class="hljs-keyword">if</span> train_q_out:
weights[<span class="hljs-string">&quot;to_q_custom_diffusion.weight&quot;</span>] = st[layer_name + <span class="hljs-string">&quot;.to_q.weight&quot;</span>]
weights[<span class="hljs-string">&quot;to_out_custom_diffusion.0.weight&quot;</span>] = st[layer_name + <span class="hljs-string">&quot;.to_out.0.weight&quot;</span>]
weights[<span class="hljs-string">&quot;to_out_custom_diffusion.0.bias&quot;</span>] = st[layer_name + <span class="hljs-string">&quot;.to_out.0.bias&quot;</span>]
<span class="hljs-keyword">if</span> cross_attention_dim <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>:
custom_diffusion_attn_procs[name] = attention_class(
train_kv=train_kv,
train_q_out=train_q_out,
hidden_size=hidden_size,
cross_attention_dim=cross_attention_dim,
).to(unet.device)
custom_diffusion_attn_procs[name].load_state_dict(weights)
<span class="hljs-keyword">else</span>:
custom_diffusion_attn_procs[name] = attention_class(
train_kv=<span class="hljs-literal">False</span>,
train_q_out=<span class="hljs-literal">False</span>,
hidden_size=hidden_size,
cross_attention_dim=cross_attention_dim,
)
<span class="hljs-keyword">del</span> st
unet.set_attn_processor(custom_diffusion_attn_procs)
custom_diffusion_layers = AttnProcsLayers(unet.attn_processors)<!----></pre></div><!----> <p><a href="https://github.com/huggingface/diffusers/blob/84cd9e8d01adb47f046b1ee449fc76a0c32dc4e2/examples/custom_diffusion/train_custom_diffusion.py#L982" rel="nofollow">optimizer</a>는 크로스 어텐션 레이어의 매개변수를 업데이트하도록 초기화됩니다:</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!---->optimizer = optimizer_class(
itertools.chain(text_encoder.get_input_embeddings().parameters(), custom_diffusion_layers.parameters())
<span class="hljs-keyword">if</span> args.modifier_token <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>
<span class="hljs-keyword">else</span> custom_diffusion_layers.parameters(),
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)<!----></pre></div><!----> <p><a href="https://github.com/huggingface/diffusers/blob/84cd9e8d01adb47f046b1ee449fc76a0c32dc4e2/examples/custom_diffusion/train_custom_diffusion.py#L1048" rel="nofollow">training loop</a>에서는 학습하려는 개념에 해당하는 임베딩만 업데이트하는 것이 중요합니다. 즉, 나머지 모든 토큰 임베딩의 그래디언트를 0으로 설정해야 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 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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">if</span> args.modifier_token <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>:
<span class="hljs-keyword">if</span> accelerator.num_processes &gt; <span class="hljs-number">1</span>:
grads_text_encoder = text_encoder.module.get_input_embeddings().weight.grad
<span class="hljs-keyword">else</span>:
grads_text_encoder = text_encoder.get_input_embeddings().weight.grad
index_grads_to_zero = torch.arange(<span class="hljs-built_in">len</span>(tokenizer)) != modifier_token_id[<span class="hljs-number">0</span>]
<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(modifier_token_id[<span class="hljs-number">1</span>:])):
index_grads_to_zero = index_grads_to_zero &amp; (
torch.arange(<span class="hljs-built_in">len</span>(tokenizer)) != modifier_token_id[i]
)
grads_text_encoder.data[index_grads_to_zero, :] = grads_text_encoder.data[
index_grads_to_zero, :
].fill_(<span class="hljs-number">0</span>)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="launch-the-script" 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="#launch-the-script"><span><svg 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><!---->스크립트 실행<!----></span></h2><!--]--><!----> <p>모든 변경 사항을 완료했거나 기본 구성으로 진행하는 데 문제가 없다면, 학습 스크립트를 실행할 준비가 된 것입니다! 🚀</p> <p>이 가이드에서는 <a href="https://www.cs.cmu.edu/~custom-diffusion/assets/data.zip" rel="nofollow">고양이 이미지</a> 예제를 다운로드하여 사용합니다. 원하시는 경우, 직접 데이터셋을 생성하여 사용할 수도 있습니다(<a href="create_dataset">학습용 데이터셋 만들기</a> 가이드 참고).</p> <p>환경 변수 <code>MODEL_NAME</code>을 Hub의 모델 ID 또는 로컬 모델 경로로, <code>INSTANCE_DIR</code>을 방금 고양이 이미지를 다운로드한 경로로, <code>OUTPUT_DIR</code>을 모델을 저장할 위치로 설정합니다. 새로 학습된 임베딩을 연결할 특수 단어로 <code>&lt;new1></code>를 사용합니다. 스크립트는 모델 체크포인트와 pytorch_custom_diffusion_weights.bin 파일을 저장소에 생성하고 저장합니다.</p> <p>Weights and Biases로 학습 진행 상황을 모니터링하려면 학습 명령에 <code>--report_to=wandb</code> 매개변수를 추가하고 <code>--validation_prompt</code>로 검증 프롬프트를 지정하세요. 이는 디버깅 및 중간 결과 저장에 유용합니다.</p> <blockquote class="tip"><p>인간 얼굴로 학습하는 경우, Custom Diffusion 팀은 다음 매개변수가 잘 작동한다고 발견했습니다:</p> <ul><li><code>--learning_rate=5e-6</code></li> <li><code>--max_train_steps</code>은 1000에서 2000 사이의 값으로 설정할 수 있습니다</li> <li><code>--freeze_model=crossattn</code></li> <li>학습에 최소 15~20개의 이미지를 사용하세요</li></ul></blockquote> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><!--[--><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">단일 개념</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">여러 개념</div><!--]--></div> <div class="language-select"><div><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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;path-to-save-model&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;./data/cat&quot;</span>
accelerate launch train_custom_diffusion.py \
--pretrained_model_name_or_path=<span class="hljs-variable">$MODEL_NAME</span> \
--instance_data_dir=<span class="hljs-variable">$INSTANCE_DIR</span> \
--output_dir=<span class="hljs-variable">$OUTPUT_DIR</span> \
--class_data_dir=./real_reg/samples_cat/ \
--with_prior_preservation \
--real_prior \
--prior_loss_weight=1.0 \
--class_prompt=<span class="hljs-string">&quot;cat&quot;</span> \
--num_class_images=200 \
--instance_prompt=<span class="hljs-string">&quot;photo of a &lt;new1&gt; cat&quot;</span> \
--resolution=512 \
--train_batch_size=2 \
--learning_rate=1e-5 \
--lr_warmup_steps=0 \
--max_train_steps=250 \
--scale_lr \
--hflip \
--modifier_token <span class="hljs-string">&quot;&lt;new1&gt;&quot;</span> \
--validation_prompt=<span class="hljs-string">&quot;&lt;new1&gt; cat sitting in a bucket&quot;</span> \
--report_to=<span class="hljs-string">&quot;wandb&quot;</span> \
--push_to_hub<!----></pre></div><!----></div><!----> <div hidden=""><p>Custom Diffusion은 학습해야 할 각 개념에 대한 세부 정보가 포함된 <a href="https://github.com/adobe-research/custom-diffusion/blob/main/assets/concept_list.json" rel="nofollow">JSON</a> 파일을 제공하면 여러 개념을 학습할 수도 있습니다.</p> <p>정규화에 사용할 실제 이미지를 수집하기 위해 clip-retrieval을 실행합니다:</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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->pip install clip-retrieval
python retrieve.py --class_prompt {} --class_data_dir {} --num_class_images 200<!----></pre></div><!----> <p>그 후 스크립트를 실행할 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 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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;path-to-save-model&quot;</span>
accelerate launch train_custom_diffusion.py \
--pretrained_model_name_or_path=<span class="hljs-variable">$MODEL_NAME</span> \
--output_dir=<span class="hljs-variable">$OUTPUT_DIR</span> \
--concepts_list=./concept_list.json \
--with_prior_preservation \
--real_prior \
--prior_loss_weight=1.0 \
--resolution=512 \
--train_batch_size=2 \
--learning_rate=1e-5 \
--lr_warmup_steps=0 \
--max_train_steps=500 \
--num_class_images=200 \
--scale_lr \
--hflip \
--modifier_token <span class="hljs-string">&quot;&lt;new1&gt;+&lt;new2&gt;&quot;</span> \
--push_to_hub<!----></pre></div><!----><!----></div><!----><!----></div><!----> <p>학습이 완료되면, 새로 만든 Custom Diffusion 모델을 추론에 사용할 수 있습니다.</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><!--[--><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">단일 개념</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">여러 개념</div><!--]--></div> <div class="language-select"><div><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 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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline
pipeline = DiffusionPipeline.from_pretrained(
<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>, dtype=torch.float16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
pipeline.unet.load_attn_procs(<span class="hljs-string">&quot;path-to-save-model&quot;</span>, weight_name=<span class="hljs-string">&quot;pytorch_custom_diffusion_weights.bin&quot;</span>)
pipeline.load_textual_inversion(<span class="hljs-string">&quot;path-to-save-model&quot;</span>, weight_name=<span class="hljs-string">&quot;&lt;new1&gt;.bin&quot;</span>)
image = pipeline(
<span class="hljs-string">&quot;&lt;new1&gt; cat sitting in a bucket&quot;</span>,
num_inference_steps=<span class="hljs-number">100</span>,
guidance_scale=<span class="hljs-number">6.0</span>,
eta=<span class="hljs-number">1.0</span>,
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;cat.png&quot;</span>)<!----></pre></div><!----></div><!----> <div hidden=""><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 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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> huggingface_hub.repocard <span class="hljs-keyword">import</span> RepoCard
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline
pipeline = DiffusionPipeline.from_pretrained(
<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>, dtype=torch.float16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
model_id = <span class="hljs-string">&quot;sayakpaul/custom-diffusion-cat-wooden-pot&quot;</span>
pipeline.unet.load_attn_procs(model_id, weight_name=<span class="hljs-string">&quot;pytorch_custom_diffusion_weights.bin&quot;</span>)
pipeline.load_textual_inversion(model_id, weight_name=<span class="hljs-string">&quot;&lt;new1&gt;.bin&quot;</span>)
pipeline.load_textual_inversion(model_id, weight_name=<span class="hljs-string">&quot;&lt;new2&gt;.bin&quot;</span>)
image = pipeline(
<span class="hljs-string">&quot;the &lt;new1&gt; cat sculpture in the style of a &lt;new2&gt; wooden pot&quot;</span>,
num_inference_steps=<span class="hljs-number">100</span>,
guidance_scale=<span class="hljs-number">6.0</span>,
eta=<span class="hljs-number">1.0</span>,
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;multi-subject.png&quot;</span>)<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="next-steps" 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="#next-steps"><span><svg 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><!---->다음 단계<!----></span></h2><!--]--><!----> <p>Custom Diffusion으로 모델을 학습시키신 것을 축하드립니다! 🎉 더 자세히 알아보려면:</p> <ul><li>Custom Diffusion 팀의 실험 결과에 대한 자세한 내용은 <a href="https://www.cs.cmu.edu/~custom-diffusion/" rel="nofollow">텍스트-이미지 확산의 다중 개념 커스터마이징</a> 블로그 포스트를 참고하세요.</li></ul> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/training/custom_diffusion.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><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]-->
<script>
{
__sveltekit_1qsdbwf = {
base: "/docs/diffusers/main/en",
assets: "/docs/diffusers/main/en"
};
const element = document.currentScript.parentElement;
Promise.all([
import("/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js"),
import("/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 366],
data: [null,null],
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
59.6 kB
·
Xet hash:
cff194dcbb680a60807f1f4ac90626facd29d0fcd4630e62422adedafb795583

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