Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /main /ko /tutorials /basic_training.html
mishig's picture
mishig
download
raw
73 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;확산 모델 학습&quot;,&quot;local&quot;:&quot;train-a-diffusion-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;학습 구성&quot;,&quot;local&quot;:&quot;training-configuration&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;데이터셋 로드&quot;,&quot;local&quot;:&quot;load-the-dataset&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;UNet2DModel 생성&quot;,&quot;local&quot;:&quot;create-a-unet2dmodel&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;스케줄러 생성&quot;,&quot;local&quot;:&quot;create-a-scheduler&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;모델 학습&quot;,&quot;local&quot;:&quot;train-the-model&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.D3LnKZ4R.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/BF9UNEM9.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DPnyVmXs.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/entry/app.C0rBoMxR.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSGUT_OW.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/C045MWWx.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.DTZsP_3B.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/IaDvMxLQ.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/319.OvCVDJ2Q.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/bKdwVYpU.js" rel="modulepreload">
<!--hfx23e--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Train a diffusion model&quot;,&quot;local&quot;:&quot;train-a-diffusion-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Training configuration&quot;,&quot;local&quot;:&quot;training-configuration&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Load the dataset&quot;,&quot;local&quot;:&quot;load-the-dataset&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Create a UNet2DModel&quot;,&quot;local&quot;:&quot;create-a-unet2dmodel&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Create a scheduler&quot;,&quot;local&quot;:&quot;create-a-scheduler&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Train the model&quot;,&quot;local&quot;:&quot;train-the-model&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><!----> <div class="flex space-x-1 " style="float: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;" dir="ltr"><!----> <!--[1--><div class="relative colab-dropdown "><button class=" " type="button"><!--[0--><!--[-1--><img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"/><!--]--><!----><!--]--></button> <!--[-1--><!--]--></div><!--]--> <!--[1--><div class="relative colab-dropdown "><button class=" " type="button"><!--[0--><!--[-1--><img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"/><!--]--><!----><!--]--></button> <!--[-1--><!--]--></div><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="train-a-diffusion-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-a-diffusion-model"><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><!---->확산 모델 학습<!----></span></h1><!--]--><!----> <p>무조건적 이미지 생성은 학습에 사용된 데이터셋의 이미지와 유사한 이미지를 생성하는 diffusion models의 인기 있는 응용 분야입니다. 일반적으로 특정 데이터셋에서 사전 학습된 모델을 fine-tuning할 때 가장 좋은 결과를 얻을 수 있습니다. 이러한 checkpoint 중 많은 것을 <a href="https://huggingface.co/search/full-text?q=unconditional-image-generation&amp;type=model" rel="nofollow">Hub</a>에서 찾을 수 있지만, 마음에 드는 checkpoint를 찾지 못한다면 직접 학습할 수도 있습니다!</p> <p>이 튜토리얼에서는 <a href="https://huggingface.co/datasets/huggan/smithsonian_butterflies_subset" rel="nofollow">Smithsonian Butterflies</a> 데이터셋의 하위 집합을 사용하여 <a href="/docs/diffusers/main/ko/api/models/unet2d#diffusers.UNet2DModel">UNet2DModel</a>을 처음부터 학습하여 나만의 🦋 나비 🦋를 생성하는 방법을 배웁니다.</p> <blockquote class="tip"><p>💡 이 학습 튜토리얼은 <a href="https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb" rel="nofollow">Training with 🧨 Diffusers</a> 노트북을 기반으로 합니다. Diffusion 모델이 어떻게 작동하는지 등 추가적인 세부 사항과 배경을 확인하려면 노트북을 참고하세요!</p></blockquote> <p>시작하기 전에 이미지 데이터셋을 로드하고 전처리하는 데 필요한 🤗 Datasets와, GPU 개수에 관계없이 학습을 단순화하는 🤗 Accelerate가 설치되어 있는지 확인하세요. 다음 명령을 실행하면 학습 메트릭을 시각화하는 데 사용할 <a href="https://www.tensorflow.org/tensorboard" rel="nofollow">TensorBoard</a>도 함께 설치됩니다(학습을 추적하는 데 <a href="https://docs.wandb.ai/" rel="nofollow">Weights &amp; Biases</a>를 사용할 수도 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-comment"># uncomment to install the necessary libraries in Colab</span>
<span class="hljs-comment">#!pip install diffusers[training]</span><!----></pre></div><!----> <p>모델을 커뮤니티와 공유하는 것을 권장하며, 이를 위해 Hugging Face 계정에 로그인해야 합니다(계정이 없다면 <a href="https://hf.co/join" rel="nofollow">여기</a>에서 생성하세요!). 노트북에서 로그인하고 프롬프트가 표시되면 토큰을 입력할 수 있습니다. 토큰에 write 역할이 있는지 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 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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> notebook_login
<span class="hljs-meta">&gt;&gt;&gt; </span>notebook_login()<!----></pre></div><!----> <p>또는 터미널에서 로그인하세요:</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 "><!---->hf auth login<!----></pre></div><!----> <p>모델 체크포인트가 상당히 크기 때문에 이러한 대용량 파일의 버전 관리를 위해 <a href="https://git-lfs.com/" rel="nofollow">Git-LFS</a>를 설치하세요:</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">sudo</span> apt -qq install git-lfs
!git config --global credential.helper store<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="training-configuration" 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-configuration"><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 구성<!----></span></h2><!--]--><!----> <p>편의상 학습 하이퍼파라미터를 포함하는 <code>TrainingConfig</code> 클래스를 생성합니다(값은 자유롭게 조정할 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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> dataclasses <span class="hljs-keyword">import</span> dataclass
<span class="hljs-meta">&gt;&gt;&gt; </span>@dataclass
<span class="hljs-meta">... </span><span class="hljs-keyword">class</span> <span class="hljs-title class_">TrainingConfig</span>:
<span class="hljs-meta">... </span> image_size = <span class="hljs-number">128</span> <span class="hljs-comment"># the generated image resolution</span>
<span class="hljs-meta">... </span> train_batch_size = <span class="hljs-number">16</span>
<span class="hljs-meta">... </span> eval_batch_size = <span class="hljs-number">16</span> <span class="hljs-comment"># how many images to sample during evaluation</span>
<span class="hljs-meta">... </span> num_epochs = <span class="hljs-number">50</span>
<span class="hljs-meta">... </span> gradient_accumulation_steps = <span class="hljs-number">1</span>
<span class="hljs-meta">... </span> learning_rate = <span class="hljs-number">1e-4</span>
<span class="hljs-meta">... </span> lr_warmup_steps = <span class="hljs-number">500</span>
<span class="hljs-meta">... </span> save_image_epochs = <span class="hljs-number">10</span>
<span class="hljs-meta">... </span> save_model_epochs = <span class="hljs-number">30</span>
<span class="hljs-meta">... </span> mixed_precision = <span class="hljs-string">&quot;fp16&quot;</span> <span class="hljs-comment"># `no` for float32, `fp16` for automatic mixed precision</span>
<span class="hljs-meta">... </span> output_dir = <span class="hljs-string">&quot;ddpm-butterflies-128&quot;</span> <span class="hljs-comment"># the model name locally and on the HF Hub</span>
<span class="hljs-meta">... </span> push_to_hub = <span class="hljs-literal">True</span> <span class="hljs-comment"># whether to upload the saved model to the HF Hub</span>
<span class="hljs-meta">... </span> hub_model_id = <span class="hljs-string">&quot;&lt;your-username&gt;/&lt;my-awesome-model&gt;&quot;</span> <span class="hljs-comment"># the name of the repository to create on the HF Hub</span>
<span class="hljs-meta">... </span> hub_private_repo = <span class="hljs-literal">None</span>
<span class="hljs-meta">... </span> overwrite_output_dir = <span class="hljs-literal">True</span> <span class="hljs-comment"># overwrite the old model when re-running the notebook</span>
<span class="hljs-meta">... </span> seed = <span class="hljs-number">0</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>config = TrainingConfig()<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="load-the-dataset" 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-dataset"><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><!---->데이터셋 로드<!----></span></h2><!--]--><!----> <p>🤗 Datasets 라이브러리를 사용하면 <a href="https://huggingface.co/datasets/huggan/smithsonian_butterflies_subset" rel="nofollow">Smithsonian Butterflies</a> 데이터셋을 쉽게 로드할 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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset
<span class="hljs-meta">&gt;&gt;&gt; </span>config.dataset_name = <span class="hljs-string">&quot;huggan/smithsonian_butterflies_subset&quot;</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>dataset = load_dataset(config.dataset_name, split=<span class="hljs-string">&quot;train&quot;</span>)<!----></pre></div><!----> <blockquote class="tip"><p>💡 <a href="https://huggingface.co/huggan" rel="nofollow">HugGan Community Event</a>에서 추가 데이터셋을 찾을 수 있거나, 로컬 <a href="https://huggingface.co/docs/datasets/image_dataset#imagefolder" rel="nofollow"><code>ImageFolder</code></a>를 만들어서 자체 데이터셋을 사용할 수도 있습니다. HugGan Community Event에서 가져온 데이터셋인 경우 <code>config.dataset_name</code>을 해당 데이터셋의 저장소 id로 설정하고, 자체 이미지를 사용하는 경우 <code>imagefolder</code>로 설정하세요.</p></blockquote> <p>🤗 Datasets는 <a href="https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Image" rel="nofollow">Image</a> 기능을 사용하여 이미지 데이터를 자동으로 디코딩하고 시각화할 수 있는 <a href="https://pillow.readthedocs.io/en/stable/reference/Image.html" rel="nofollow"><code>PIL.Image</code></a>로 로드합니다:</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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> matplotlib.pyplot <span class="hljs-keyword">as</span> plt
<span class="hljs-meta">&gt;&gt;&gt; </span>fig, axs = plt.subplots(<span class="hljs-number">1</span>, <span class="hljs-number">4</span>, figsize=(<span class="hljs-number">16</span>, <span class="hljs-number">4</span>))
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">for</span> i, image <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(dataset[:<span class="hljs-number">4</span>][<span class="hljs-string">&quot;image&quot;</span>]):
<span class="hljs-meta">... </span> axs[i].imshow(image)
<span class="hljs-meta">... </span> axs[i].set_axis_off()
<span class="hljs-meta">&gt;&gt;&gt; </span>fig.show()<!----></pre></div><!----> <div class="flex justify-center"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/butterflies_ds.png"/></div> <p>이미지 크기가 모두 다르므로 먼저 전처리해야 합니다:</p> <ul><li><code>Resize</code>는 이미지의 크기를 <code>config.image_size</code>에 정의된 값으로 변경합니다.</li> <li><code>RandomHorizontalFlip</code>는 이미지를 무작위로 좌우 반전시켜 데이터셋을 증강합니다.</li> <li><code>Normalize</code>는 모델이 기대하는 [-1, 1] 범위로 픽셀 값을 다시 스케일링하는 데 중요합니다.</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-py "><!----><span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> torchvision <span class="hljs-keyword">import</span> transforms
<span class="hljs-meta">&gt;&gt;&gt; </span>preprocess = transforms.Compose(
<span class="hljs-meta">... </span> [
<span class="hljs-meta">... </span> transforms.Resize((config.image_size, config.image_size)),
<span class="hljs-meta">... </span> transforms.RandomHorizontalFlip(),
<span class="hljs-meta">... </span> transforms.ToTensor(),
<span class="hljs-meta">... </span> transforms.Normalize([<span class="hljs-number">0.5</span>], [<span class="hljs-number">0.5</span>]),
<span class="hljs-meta">... </span> ]
<span class="hljs-meta">... </span>)<!----></pre></div><!----> <p>🤗 Datasets의 <a href="https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.set_transform" rel="nofollow">set_transform</a> 메서드를 사용하여 학습 중에 <code>preprocess</code> 함수를 실시간으로 적용합니다:</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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">def</span> <span class="hljs-title function_">transform</span>(<span class="hljs-params">examples</span>):
<span class="hljs-meta">... </span> images = [preprocess(image.convert(<span class="hljs-string">&quot;RGB&quot;</span>)) <span class="hljs-keyword">for</span> image <span class="hljs-keyword">in</span> examples[<span class="hljs-string">&quot;image&quot;</span>]]
<span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> {<span class="hljs-string">&quot;images&quot;</span>: images}
<span class="hljs-meta">&gt;&gt;&gt; </span>dataset.set_transform(transform)<!----></pre></div><!----> <p>이미지가 리사이즈되었는지 확인하기 위해 다시 시각화해 보셔도 좋습니다. 이제 학습을 위해 데이터셋을 <a href="https://pytorch.org/docs/stable/data#torch.utils.data.DataLoader" rel="nofollow">DataLoader</a>로 감싸기 준비가 되었습니다!</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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch
<span class="hljs-meta">&gt;&gt;&gt; </span>train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=config.train_batch_size, shuffle=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="create-a-unet2dmodel" 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="#create-a-unet2dmodel"><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><!---->UNet2DModel 생성<!----></span></h2><!--]--><!----> <p>🧨 Diffusers의 사전 학습 모델은 원하는 매개변수를 사용하여 해당 모델 클래스로 쉽게 생성할 수 있습니다. 예를 들어, <a href="/docs/diffusers/main/ko/api/models/unet2d#diffusers.UNet2DModel">UNet2DModel</a>을 생성하려면:</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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> UNet2DModel
<span class="hljs-meta">&gt;&gt;&gt; </span>model = UNet2DModel(
<span class="hljs-meta">... </span> sample_size=config.image_size, <span class="hljs-comment"># the target image resolution</span>
<span class="hljs-meta">... </span> in_channels=<span class="hljs-number">3</span>, <span class="hljs-comment"># the number of input channels, 3 for RGB images</span>
<span class="hljs-meta">... </span> out_channels=<span class="hljs-number">3</span>, <span class="hljs-comment"># the number of output channels</span>
<span class="hljs-meta">... </span> layers_per_block=<span class="hljs-number">2</span>, <span class="hljs-comment"># how many ResNet layers to use per UNet block</span>
<span class="hljs-meta">... </span> block_out_channels=(<span class="hljs-number">128</span>, <span class="hljs-number">128</span>, <span class="hljs-number">256</span>, <span class="hljs-number">256</span>, <span class="hljs-number">512</span>, <span class="hljs-number">512</span>), <span class="hljs-comment"># the number of output channels for each UNet block</span>
<span class="hljs-meta">... </span> down_block_types=(
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;DownBlock2D&quot;</span>, <span class="hljs-comment"># a regular ResNet downsampling block</span>
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;DownBlock2D&quot;</span>,
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;DownBlock2D&quot;</span>,
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;DownBlock2D&quot;</span>,
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;AttnDownBlock2D&quot;</span>, <span class="hljs-comment"># a ResNet downsampling block with spatial self-attention</span>
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;DownBlock2D&quot;</span>,
<span class="hljs-meta">... </span> ),
<span class="hljs-meta">... </span> up_block_types=(
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;UpBlock2D&quot;</span>, <span class="hljs-comment"># a regular ResNet upsampling block</span>
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;AttnUpBlock2D&quot;</span>, <span class="hljs-comment"># a ResNet upsampling block with spatial self-attention</span>
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;UpBlock2D&quot;</span>,
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;UpBlock2D&quot;</span>,
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;UpBlock2D&quot;</span>,
<span class="hljs-meta">... </span> <span class="hljs-string">&quot;UpBlock2D&quot;</span>,
<span class="hljs-meta">... </span> ),
<span class="hljs-meta">... </span>)<!----></pre></div><!----> <p>샘플 이미지의 형상이 모델 출력 형상과 일치하는지 빠르게 확인하는 것이 좋은 경우가 많습니다:</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-meta">&gt;&gt;&gt; </span>sample_image = dataset[<span class="hljs-number">0</span>][<span class="hljs-string">&quot;images&quot;</span>].unsqueeze(<span class="hljs-number">0</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Input shape:&quot;</span>, sample_image.shape)
Input shape: torch.Size([<span class="hljs-number">1</span>, <span class="hljs-number">3</span>, <span class="hljs-number">128</span>, <span class="hljs-number">128</span>])
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Output shape:&quot;</span>, model(sample_image, timestep=<span class="hljs-number">0</span>).sample.shape)
Output shape: torch.Size([<span class="hljs-number">1</span>, <span class="hljs-number">3</span>, <span class="hljs-number">128</span>, <span class="hljs-number">128</span>])<!----></pre></div><!----> <p>좋습니다! 다음으로, 이미지에 노이즈를 추가하려면 스케줄러가 필요합니다.</p> <!--[1--><h2 class="relative group"><a id="create-a-scheduler" 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="#create-a-scheduler"><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><!---->스케줄러 생성<!----></span></h2><!--]--><!----> <p>스케줄러는 모델을 학습에 사용하는지 추론에 사용하는지에 따라 동작이 다릅니다. 추론 중에는 스케줄러가 노이즈로부터 이미지를 생성합니다. 학습 중에는 스케줄러가 확산 과정의 특정 시점에서의 모델 출력, 즉 샘플을 받아 <em>노이즈 스케줄</em>과 <em>업데이트 규칙</em>에 따라 이미지에 노이즈를 적용합니다.</p> <p><a href="/docs/diffusers/main/ko/api/schedulers/ddpm#diffusers.DDPMScheduler">DDPMScheduler</a>를 살펴보고 <code>add_noise</code> 메서드를 사용해 이전의 <code>sample_image</code>에 무작위 노이즈를 추가해 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 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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDPMScheduler
<span class="hljs-meta">&gt;&gt;&gt; </span>noise_scheduler = DDPMScheduler(num_train_timesteps=<span class="hljs-number">1000</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>noise = torch.randn(sample_image.shape)
<span class="hljs-meta">&gt;&gt;&gt; </span>timesteps = torch.LongTensor([<span class="hljs-number">50</span>])
<span class="hljs-meta">&gt;&gt;&gt; </span>noisy_image = noise_scheduler.add_noise(sample_image, noise, timesteps)
<span class="hljs-meta">&gt;&gt;&gt; </span>Image.fromarray(((noisy_image.permute(<span class="hljs-number">0</span>, <span class="hljs-number">2</span>, <span class="hljs-number">3</span>, <span class="hljs-number">1</span>) + <span class="hljs-number">1.0</span>) * <span class="hljs-number">127.5</span>).<span class="hljs-built_in">type</span>(torch.uint8).numpy()[<span class="hljs-number">0</span>])<!----></pre></div><!----> <div class="flex justify-center"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/noisy_butterfly.png"/></div> <p>모델의 학습 목표는 이미지에 추가된 노이즈를 예측하는 것입니다. 이 단계의 손실은 다음 식으로 계산할 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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch.nn.functional <span class="hljs-keyword">as</span> F
<span class="hljs-meta">&gt;&gt;&gt; </span>noise_pred = model(noisy_image, timesteps).sample
<span class="hljs-meta">&gt;&gt;&gt; </span>loss = F.mse_loss(noise_pred, noise)<!----></pre></div><!----> <!--[1--><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 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 학습<!----></span></h2><!--]--><!----> <p>지금까지 모델 학습을 시작하는 데 필요한 대부분의 요소를 갖추셨으며, 이제 모든 것을 하나로 통합하는 것만 남았습니다.</p> <p>먼저, 옵티마이저와 학습률 스케줄러가 필요합니다:</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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> diffusers.optimization <span class="hljs-keyword">import</span> get_cosine_schedule_with_warmup
<span class="hljs-meta">&gt;&gt;&gt; </span>optimizer = torch.optim.AdamW(model.parameters(), lr=config.learning_rate)
<span class="hljs-meta">&gt;&gt;&gt; </span>lr_scheduler = get_cosine_schedule_with_warmup(
<span class="hljs-meta">... </span> optimizer=optimizer,
<span class="hljs-meta">... </span> num_warmup_steps=config.lr_warmup_steps,
<span class="hljs-meta">... </span> num_training_steps=(<span class="hljs-built_in">len</span>(train_dataloader) * config.num_epochs),
<span class="hljs-meta">... </span>)<!----></pre></div><!----> <p>그 다음으로, 모델을 평가하는 방법이 필요합니다. 평가에는 <a href="/docs/diffusers/main/ko/api/pipelines/ddpm#diffusers.DDPMPipeline">DDPMPipeline</a>을 사용해 샘플 이미지 배치 생성 후 그리드로 저장할 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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDPMPipeline
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> make_image_grid
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> os
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">def</span> <span class="hljs-title function_">evaluate</span>(<span class="hljs-params">config, epoch, pipeline</span>):
<span class="hljs-meta">... </span> <span class="hljs-comment"># Sample some images from random noise (this is the backward diffusion process).</span>
<span class="hljs-meta">... </span> <span class="hljs-comment"># The default pipeline output type is `List[PIL.Image]`</span>
<span class="hljs-meta">... </span> images = pipeline(
<span class="hljs-meta">... </span> batch_size=config.eval_batch_size,
<span class="hljs-meta">... </span> generator=torch.Generator(device=<span class="hljs-string">&#x27;cpu&#x27;</span>).manual_seed(config.seed), <span class="hljs-comment"># Use a separate torch generator to avoid rewinding the random state of the main training loop</span>
<span class="hljs-meta">... </span> ).images
<span class="hljs-meta">... </span> <span class="hljs-comment"># Make a grid out of the images</span>
<span class="hljs-meta">... </span> image_grid = make_image_grid(images, rows=<span class="hljs-number">4</span>, cols=<span class="hljs-number">4</span>)
<span class="hljs-meta">... </span> <span class="hljs-comment"># Save the images</span>
<span class="hljs-meta">... </span> test_dir = os.path.join(config.output_dir, <span class="hljs-string">&quot;samples&quot;</span>)
<span class="hljs-meta">... </span> os.makedirs(test_dir, exist_ok=<span class="hljs-literal">True</span>)
<span class="hljs-meta">... </span> image_grid.save(<span class="hljs-string">f&quot;<span class="hljs-subst">{test_dir}</span>/<span class="hljs-subst">{epoch:04d}</span>.png&quot;</span>)<!----></pre></div><!----> <p>이제 🤗 Accelerate를 사용하여 TensorBoard 로깅, 그라디언트 축적, 혼합 정밀도 학습을 쉽게 수행할 수 있는 학습 루프에 이 모든 구성 요소를 묶을 수 있습니다. 모델을 Hub에 업로드하려면 저장소 이름과 정보를 가져오는 함수를 작성한 후 Hub에 push하세요.</p> <blockquote class="tip"><p>💡 아래 학습 루프는 길고 복잡해 보일 수 있지만, 나중에 한 줄의 코드로 학습을 시작할 때 그 가치가 드러날 것입니다. 이미 생성하기를 기다릴 수 없다면, 아래 코드를 복사하여 실행해 보세요. 나중에, 예를 들어 모델 학습이 완료될 때까지 기다리는 동안 언제든 돌아와 학습 루프를 더 자세히 살펴볼 수 있습니다. 🤗</p></blockquote> <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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> Accelerator
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> create_repo, upload_folder
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> tqdm.auto <span class="hljs-keyword">import</span> tqdm
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> pathlib <span class="hljs-keyword">import</span> Path
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> os
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">def</span> <span class="hljs-title function_">train_loop</span>(<span class="hljs-params">config, model, noise_scheduler, optimizer, train_dataloader, lr_scheduler</span>):
<span class="hljs-meta">... </span> <span class="hljs-comment"># Initialize accelerator and tensorboard logging</span>
<span class="hljs-meta">... </span> accelerator = Accelerator(
<span class="hljs-meta">... </span> mixed_precision=config.mixed_precision,
<span class="hljs-meta">... </span> gradient_accumulation_steps=config.gradient_accumulation_steps,
<span class="hljs-meta">... </span> log_with=<span class="hljs-string">&quot;tensorboard&quot;</span>,
<span class="hljs-meta">... </span> project_dir=os.path.join(config.output_dir, <span class="hljs-string">&quot;logs&quot;</span>),
<span class="hljs-meta">... </span> )
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> accelerator.is_main_process:
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> config.output_dir <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>:
<span class="hljs-meta">... </span> os.makedirs(config.output_dir, exist_ok=<span class="hljs-literal">True</span>)
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> config.push_to_hub:
<span class="hljs-meta">... </span> repo_id = create_repo(
<span class="hljs-meta">... </span> repo_id=config.hub_model_id <span class="hljs-keyword">or</span> Path(config.output_dir).name, exist_ok=<span class="hljs-literal">True</span>
<span class="hljs-meta">... </span> ).repo_id
<span class="hljs-meta">... </span> accelerator.init_trackers(<span class="hljs-string">&quot;train_example&quot;</span>)
<span class="hljs-meta">... </span> <span class="hljs-comment"># Prepare everything</span>
<span class="hljs-meta">... </span> <span class="hljs-comment"># There is no specific order to remember, you just need to unpack the</span>
<span class="hljs-meta">... </span> <span class="hljs-comment"># objects in the same order you gave them to the prepare method.</span>
<span class="hljs-meta">... </span> model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
<span class="hljs-meta">... </span> model, optimizer, train_dataloader, lr_scheduler
<span class="hljs-meta">... </span> )
<span class="hljs-meta">... </span> global_step = <span class="hljs-number">0</span>
<span class="hljs-meta">... </span> <span class="hljs-comment"># Now you train the model</span>
<span class="hljs-meta">... </span> <span class="hljs-keyword">for</span> epoch <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(config.num_epochs):
<span class="hljs-meta">... </span> progress_bar = tqdm(total=<span class="hljs-built_in">len</span>(train_dataloader), disable=<span class="hljs-keyword">not</span> accelerator.is_local_main_process)
<span class="hljs-meta">... </span> progress_bar.set_description(<span class="hljs-string">f&quot;Epoch <span class="hljs-subst">{epoch}</span>&quot;</span>)
<span class="hljs-meta">... </span> <span class="hljs-keyword">for</span> step, batch <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(train_dataloader):
<span class="hljs-meta">... </span> clean_images = batch[<span class="hljs-string">&quot;images&quot;</span>]
<span class="hljs-meta">... </span> <span class="hljs-comment"># Sample noise to add to the images</span>
<span class="hljs-meta">... </span> noise = torch.randn(clean_images.shape, device=clean_images.device)
<span class="hljs-meta">... </span> bs = clean_images.shape[<span class="hljs-number">0</span>]
<span class="hljs-meta">... </span> <span class="hljs-comment"># Sample a random timestep for each image</span>
<span class="hljs-meta">... </span> timesteps = torch.randint(
<span class="hljs-meta">... </span> <span class="hljs-number">0</span>, noise_scheduler.config.num_train_timesteps, (bs,), device=clean_images.device,
<span class="hljs-meta">... </span> dtype=torch.int64
<span class="hljs-meta">... </span> )
<span class="hljs-meta">... </span> <span class="hljs-comment"># Add noise to the clean images according to the noise magnitude at each timestep</span>
<span class="hljs-meta">... </span> <span class="hljs-comment"># (this is the forward diffusion process)</span>
<span class="hljs-meta">... </span> noisy_images = noise_scheduler.add_noise(clean_images, noise, timesteps)
<span class="hljs-meta">... </span> <span class="hljs-keyword">with</span> accelerator.accumulate(model):
<span class="hljs-meta">... </span> <span class="hljs-comment"># Predict the noise residual</span>
<span class="hljs-meta">... </span> noise_pred = model(noisy_images, timesteps, return_dict=<span class="hljs-literal">False</span>)[<span class="hljs-number">0</span>]
<span class="hljs-meta">... </span> loss = F.mse_loss(noise_pred, noise)
<span class="hljs-meta">... </span> accelerator.backward(loss)
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> accelerator.sync_gradients:
<span class="hljs-meta">... </span> accelerator.clip_grad_norm_(model.parameters(), <span class="hljs-number">1.0</span>)
<span class="hljs-meta">... </span> optimizer.step()
<span class="hljs-meta">... </span> lr_scheduler.step()
<span class="hljs-meta">... </span> optimizer.zero_grad()
<span class="hljs-meta">... </span> progress_bar.update(<span class="hljs-number">1</span>)
<span class="hljs-meta">... </span> logs = {<span class="hljs-string">&quot;loss&quot;</span>: loss.detach().item(), <span class="hljs-string">&quot;lr&quot;</span>: lr_scheduler.get_last_lr()[<span class="hljs-number">0</span>], <span class="hljs-string">&quot;step&quot;</span>: global_step}
<span class="hljs-meta">... </span> progress_bar.set_postfix(**logs)
<span class="hljs-meta">... </span> accelerator.log(logs, step=global_step)
<span class="hljs-meta">... </span> global_step += <span class="hljs-number">1</span>
<span class="hljs-meta">... </span> <span class="hljs-comment"># After each epoch you optionally sample some demo images with evaluate() and save the model</span>
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> accelerator.is_main_process:
<span class="hljs-meta">... </span> pipeline = DDPMPipeline(unet=accelerator.unwrap_model(model), scheduler=noise_scheduler)
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> (epoch + <span class="hljs-number">1</span>) % config.save_image_epochs == <span class="hljs-number">0</span> <span class="hljs-keyword">or</span> epoch == config.num_epochs - <span class="hljs-number">1</span>:
<span class="hljs-meta">... </span> evaluate(config, epoch, pipeline)
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> (epoch + <span class="hljs-number">1</span>) % config.save_model_epochs == <span class="hljs-number">0</span> <span class="hljs-keyword">or</span> epoch == config.num_epochs - <span class="hljs-number">1</span>:
<span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> config.push_to_hub:
<span class="hljs-meta">... </span> upload_folder(
<span class="hljs-meta">... </span> repo_id=repo_id,
<span class="hljs-meta">... </span> folder_path=config.output_dir,
<span class="hljs-meta">... </span> commit_message=<span class="hljs-string">f&quot;Epoch <span class="hljs-subst">{epoch}</span>&quot;</span>,
<span class="hljs-meta">... </span> ignore_patterns=[<span class="hljs-string">&quot;step_*&quot;</span>, <span class="hljs-string">&quot;epoch_*&quot;</span>],
<span class="hljs-meta">... </span> )
<span class="hljs-meta">... </span> <span class="hljs-keyword">else</span>:
<span class="hljs-meta">... </span> pipeline.save_pretrained(config.output_dir)<!----></pre></div><!----> <p>후, 꽤 많은 코드였네요! 하지만 이제 🤗 Accelerate의 <a href="https://huggingface.co/docs/accelerate/main/en/package_reference/launchers#accelerate.notebook_launcher" rel="nofollow">notebook_launcher</a> 함수로 학습을 시작할 준비가 끝났습니다. 학습 루프, 모든 학습 인자, 그리고 학습에 사용할 프로세스 수(이 값은 사용 가능한 GPU 수로 변경할 수 있습니다)를 함수에 전달하세요:</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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> notebook_launcher
<span class="hljs-meta">&gt;&gt;&gt; </span>args = (config, model, noise_scheduler, optimizer, train_dataloader, lr_scheduler)
<span class="hljs-meta">&gt;&gt;&gt; </span>notebook_launcher(train_loop, args, num_processes=<span class="hljs-number">1</span>)<!----></pre></div><!----> <p>학습이 완료되면, diffusion 모델이 생성한 최종 🦋 이미지 🦋를 확인해 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 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-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> glob
<span class="hljs-meta">&gt;&gt;&gt; </span>sample_images = <span class="hljs-built_in">sorted</span>(glob.glob(<span class="hljs-string">f&quot;<span class="hljs-subst">{config.output_dir}</span>/samples/*.png&quot;</span>))
<span class="hljs-meta">&gt;&gt;&gt; </span>Image.<span class="hljs-built_in">open</span>(sample_images[-<span class="hljs-number">1</span>])<!----></pre></div><!----> <div class="flex justify-center"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/butterflies_final.png"/></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>무조건적 이미지 생성은 학습할 수 있는 작업의 한 예시입니다. <a href="../training/overview">🧨 Diffusers Training Examples</a> 페이지를 방문하여 다른 작업과 학습 기법을 탐색해 보세요. 아래는 배울 수 있는 내용들의 예시입니다:</p> <ul><li><a href="../training/text_inversion">Textual Inversion</a>는 모델이 특정 시각 개념을 학습하도록 하고 이를 생성된 이미지에 통합하는 알고리즘입니다.</li> <li>주제에 대한 여러 입력 이미지를 제공하여 해당 주제의 개인화된 이미지를 생성하는 기법인 <a href="../training/dreambooth">DreamBooth</a>.</li> <li><a href="../training/text2image">가이드</a>: 자체 데이터셋으로 Stable Diffusion 모델을 파인튜닝하는 방법</li> <li>매우 큰 모델을 더 빠르게 미세 조정하기 위한 메모리 효율적인 기법인 LoRA 사용 <a href="../training/lora">가이드</a>.</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/tutorials/basic_training.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_g37pnb = {
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.D3LnKZ4R.js"),
import("/docs/diffusers/main/en/_app/immutable/entry/app.C0rBoMxR.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 319],
data: [null,null],
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
73 kB
·
Xet hash:
25199198c092b9bd311e4e5bb9171365ef3a927a08820adf1f5f104a3ced3c23

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