Buckets:
| <meta charset="utf-8" /><meta http-equiv="content-security-policy" content=""><meta name="hf:doc:metadata" content="{"local":"diffusion","sections":[{"local":"","title":"학습 구성"},{"local":"","title":"데이터셋 불러오기"},{"local":"unet2dmodel","title":"UNet2DModel 생성하기"},{"local":"","title":"스케줄러 생성하기"},{"local":"","title":"모델 학습하기"},{"local":"","title":"다음 단계"}],"title":"Diffusion 모델을 학습하기"}" data-svelte="svelte-1phssyn"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/assets/pages/__layout.svelte-hf-doc-builder.css"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/start-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/chunks/vendor-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/chunks/paths-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/pages/__layout.svelte-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/pages/tutorials/basic_training.mdx-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/chunks/Tip-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/chunks/IconCopyLink-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/chunks/CodeBlock-hf-doc-builder.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/v0.19.2/ko/_app/chunks/DocNotebookDropdown-hf-doc-builder.js"> | |
| <div class="flex space-x-1 absolute z-10 right-0 top-0"> | |
| <div class="relative colab-dropdown "> | |
| <button class=" " type="button"> | |
| <img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"> | |
| </button> | |
| </div> | |
| <div class="relative colab-dropdown "> | |
| <button class=" " type="button"> | |
| <img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"> | |
| </button> | |
| </div></div> | |
| <h1 class="relative group"><a id="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="#diffusion"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> | |
| <span>Diffusion 모델을 학습하기 | |
| </span></h1> | |
| <p>Unconditional 이미지 생성은 학습에 사용된 데이터셋과 유사한 이미지를 생성하는 diffusion 모델에서 인기 있는 어플리케이션입니다. 일반적으로, 가장 좋은 결과는 특정 데이터셋에 사전 훈련된 모델을 파인튜닝하는 것으로 얻을 수 있습니다. 이 <a href="https://huggingface.co/search/full-text?q=unconditional-image-generation&type=model" rel="nofollow">허브</a>에서 이러한 많은 체크포인트를 찾을 수 있지만, 만약 마음에 드는 체크포인트를 찾지 못했다면, 언제든지 스스로 학습할 수 있습니다!</p> | |
| <p>이 튜토리얼은 나만의 🦋 나비 🦋를 생성하기 위해 <a href="https://huggingface.co/datasets/huggan/smithsonian_butterflies_subset" rel="nofollow">Smithsonian Butterflies</a> 데이터셋의 하위 집합에서 <code>UNet2DModel</code> 모델을 학습하는 방법을 가르쳐줄 것입니다.</p> | |
| <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><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> 노트북 기반으로 합니다. Diffusion 모델의 작동 방식 및 자세한 내용은 노트북을 확인하세요!</p></div> | |
| <p>시작 전에, 🤗 Datasets을 불러오고 전처리하기 위해 데이터셋이 설치되어 있는지 다수 GPU에서 학습을 간소화하기 위해 🤗 Accelerate 가 설치되어 있는지 확인하세요. 그 후 학습 메트릭을 시각화하기 위해 <a href="https://www.tensorflow.org/tensorboard" rel="nofollow">TensorBoard</a>를 또한 설치하세요. (또한 학습 추적을 위해 <a href="https://docs.wandb.ai/" rel="nofollow">Weights & Biases</a>를 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 class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START -->!pip install diffusers[training]<!-- HTML_TAG_END --></pre></div> | |
| <p>커뮤니티에 모델을 공유할 것을 권장하며, 이를 위해서 Hugging Face 계정에 로그인을 해야 합니다. (계정이 없다면 <a href="https://hf.co/join" rel="nofollow">여기</a>에서 만들 수 있습니다.) 노트북에서 로그인할 수 있으며 메시지가 표시되면 토큰을 입력할 수 있습니다.</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> notebook_login | |
| <span class="hljs-meta">>>> </span>notebook_login()<!-- HTML_TAG_END --></pre></div> | |
| <p>또는 터미널로 로그인할 수 있습니다:</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START -->huggingface-cli login<!-- HTML_TAG_END --></pre></div> | |
| <p>모델 체크포인트가 상당히 크기 때문에 <a href="https://git-lfs.com/" rel="nofollow">Git-LFS</a>에서 대용량 파일의 버전 관리를 할 수 있습니다.</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START -->!sudo apt -qq install git-lfs | |
| !git config --global credential.helper store<!-- HTML_TAG_END --></pre></div> | |
| <h2 id="">학습 구성</h2> | |
| <p>편의를 위해 학습 파라미터들을 포함한 <code>TrainingConfig</code> 클래스를 생성합니다 (자유롭게 조정 가능):</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> dataclasses <span class="hljs-keyword">import</span> dataclass | |
| <span class="hljs-meta">>>> </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"># 생성되는 이미지 해상도</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"># 평가 동안에 샘플링할 이미지 수</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">"fp16"</span> <span class="hljs-comment"># `no`는 float32, 자동 혼합 정밀도를 위한 `fp16`</span> | |
| <span class="hljs-meta">... </span> output_dir = <span class="hljs-string">"ddpm-butterflies-128"</span> <span class="hljs-comment"># 로컬 및 HF Hub에 저장되는 모델명</span> | |
| <span class="hljs-meta">... </span> push_to_hub = <span class="hljs-literal">True</span> <span class="hljs-comment"># 저장된 모델을 HF Hub에 업로드할지 여부</span> | |
| <span class="hljs-meta">... </span> hub_private_repo = <span class="hljs-literal">False</span> | |
| <span class="hljs-meta">... </span> overwrite_output_dir = <span class="hljs-literal">True</span> <span class="hljs-comment"># 노트북을 다시 실행할 때 이전 모델에 덮어씌울지</span> | |
| <span class="hljs-meta">... </span> seed = <span class="hljs-number">0</span> | |
| <span class="hljs-meta">>>> </span>config = TrainingConfig()<!-- HTML_TAG_END --></pre></div> | |
| <h2 id="">데이터셋 불러오기</h2> | |
| <p>🤗 Datasets 라이브러리와 <a href="https://huggingface.co/datasets/huggan/smithsonian_butterflies_subset" rel="nofollow">Smithsonian Butterflies</a> 데이터셋을 쉽게 불러올 수 있습니다.</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset | |
| <span class="hljs-meta">>>> </span>config.dataset_name = <span class="hljs-string">"huggan/smithsonian_butterflies_subset"</span> | |
| <span class="hljs-meta">>>> </span>dataset = load_dataset(config.dataset_name, split=<span class="hljs-string">"train"</span>)<!-- HTML_TAG_END --></pre></div> | |
| <p>💡<a href="https://huggingface.co/huggan" rel="nofollow">HugGan Community Event</a> 에서 추가의 데이터셋을 찾거나 로컬의 <a href="https://huggingface.co/docs/datasets/image_dataset#imagefolder" rel="nofollow"><code>ImageFolder</code></a>를 만듦으로써 나만의 데이터셋을 사용할 수 있습니다. HugGan Community Event 에 가져온 데이터셋의 경우 레포지토리의 id로 <code>config.dataset_name</code> 을 설정하고, 나만의 이미지를 사용하는 경우 <code>imagefolder</code> 를 설정합니다.</p> | |
| <p>🤗 Datasets은 <code>Image</code> 기능을 사용해 자동으로 이미지 데이터를 디코딩하고 <a href="https://pillow.readthedocs.io/en/stable/reference/Image.html" rel="nofollow"><code>PIL.Image</code></a>로 불러옵니다. 이를 시각화 해보면:</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> matplotlib.pyplot <span class="hljs-keyword">as</span> plt | |
| <span class="hljs-meta">>>> </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">>>> </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">"image"</span>]): | |
| <span class="hljs-meta">... </span> axs[i].imshow(image) | |
| <span class="hljs-meta">... </span> axs[i].set_axis_off() | |
| <span class="hljs-meta">>>> </span>fig.show()<!-- HTML_TAG_END --></pre></div> | |
| <p><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/butterflies_ds.png"></p> | |
| <p>이미지는 모두 다른 사이즈이기 때문에, 우선 전처리가 필요합니다:</p> | |
| <ul><li><code>Resize</code> 는 <code>config.image_size</code> 에 정의된 이미지 사이즈로 변경합니다.</li> | |
| <li><code>RandomHorizontalFlip</code> 은 랜덤적으로 이미지를 미러링하여 데이터셋을 보강합니다.</li> | |
| <li><code>Normalize</code> 는 모델이 예상하는 [-1, 1] 범위로 픽셀 값을 재조정 하는데 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 class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> torchvision <span class="hljs-keyword">import</span> transforms | |
| <span class="hljs-meta">>>> </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>)<!-- HTML_TAG_END --></pre></div> | |
| <p>학습 도중에 <code>preprocess</code> 함수를 적용하려면 🤗 Datasets의 <code>set_transform</code> 방법이 사용됩니다.</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </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">"RGB"</span>)) <span class="hljs-keyword">for</span> image <span class="hljs-keyword">in</span> examples[<span class="hljs-string">"image"</span>]] | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> {<span class="hljs-string">"images"</span>: images} | |
| <span class="hljs-meta">>>> </span>dataset.set_transform(transform)<!-- HTML_TAG_END --></pre></div> | |
| <p>이미지의 크기가 조정되었는지 확인하기 위해 이미지를 다시 시각화해보세요. 이제 <a href="https://pytorch.org/docs/stable/data#torch.utils.data.DataLoader" rel="nofollow">DataLoader</a>에 데이터셋을 포함해 학습할 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 class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-meta">>>> </span>train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=config.train_batch_size, shuffle=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> | |
| <h2 class="relative group"><a id="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="#unet2dmodel"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> | |
| <span>UNet2DModel 생성하기 | |
| </span></h2> | |
| <p>🧨 Diffusers에 사전학습된 모델들은 모델 클래스에서 원하는 파라미터로 쉽게 생성할 수 있습니다. 예를 들어, <code>UNet2DModel</code>를 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 class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> UNet2DModel | |
| <span class="hljs-meta">>>> </span>model = UNet2DModel( | |
| <span class="hljs-meta">... </span> sample_size=config.image_size, <span class="hljs-comment"># 타겟 이미지 해상도</span> | |
| <span class="hljs-meta">... </span> in_channels=<span class="hljs-number">3</span>, <span class="hljs-comment"># 입력 채널 수, RGB 이미지에서 3</span> | |
| <span class="hljs-meta">... </span> out_channels=<span class="hljs-number">3</span>, <span class="hljs-comment"># 출력 채널 수</span> | |
| <span class="hljs-meta">... </span> layers_per_block=<span class="hljs-number">2</span>, <span class="hljs-comment"># UNet 블럭당 몇 개의 ResNet 레이어가 사용되는지</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"># 각 UNet 블럭을 위한 출력 채널 수</span> | |
| <span class="hljs-meta">... </span> down_block_types=( | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"DownBlock2D"</span>, <span class="hljs-comment"># 일반적인 ResNet 다운샘플링 블럭</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"DownBlock2D"</span>, | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"DownBlock2D"</span>, | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"DownBlock2D"</span>, | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"AttnDownBlock2D"</span>, <span class="hljs-comment"># spatial self-attention이 포함된 일반적인 ResNet 다운샘플링 블럭</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"DownBlock2D"</span>, | |
| <span class="hljs-meta">... </span> ), | |
| <span class="hljs-meta">... </span> up_block_types=( | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"UpBlock2D"</span>, <span class="hljs-comment"># 일반적인 ResNet 업샘플링 블럭</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"AttnUpBlock2D"</span>, <span class="hljs-comment"># spatial self-attention이 포함된 일반적인 ResNet 업샘플링 블럭</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"UpBlock2D"</span>, | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"UpBlock2D"</span>, | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"UpBlock2D"</span>, | |
| <span class="hljs-meta">... </span> <span class="hljs-string">"UpBlock2D"</span>, | |
| <span class="hljs-meta">... </span> ), | |
| <span class="hljs-meta">... </span>)<!-- HTML_TAG_END --></pre></div> | |
| <p>샘플의 이미지 크기와 모델 출력 크기가 맞는지 빠르게 확인하기 위한 좋은 아이디어가 있습니다:</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>sample_image = dataset[<span class="hljs-number">0</span>][<span class="hljs-string">"images"</span>].unsqueeze(<span class="hljs-number">0</span>) | |
| <span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(<span class="hljs-string">"Input shape:"</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">>>> </span><span class="hljs-built_in">print</span>(<span class="hljs-string">"Output shape:"</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>])<!-- HTML_TAG_END --></pre></div> | |
| <p>훌륭해요! 다음, 이미지에 약간의 노이즈를 더하기 위해 스케줄러가 필요합니다.</p> | |
| <h2 id="">스케줄러 생성하기</h2> | |
| <p>스케줄러는 모델을 학습 또는 추론에 사용하는지에 따라 다르게 작동합니다. 추론시에, 스케줄러는 노이즈로부터 이미지를 생성합니다. 학습시 스케줄러는 diffusion 과정에서의 특정 포인트로부터 모델의 출력 또는 샘플을 가져와 <em>노이즈 스케줄</em> 과 <em>업데이트 규칙</em>에 따라 이미지에 노이즈를 적용합니다.</p> | |
| <p><code>DDPMScheduler</code>를 보면 이전으로부터 <code>sample_image</code>에 랜덤한 노이즈를 더하는 <code>add_noise</code> 메서드를 사용합니다:</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDPMScheduler | |
| <span class="hljs-meta">>>> </span>noise_scheduler = DDPMScheduler(num_train_timesteps=<span class="hljs-number">1000</span>) | |
| <span class="hljs-meta">>>> </span>noise = torch.randn(sample_image.shape) | |
| <span class="hljs-meta">>>> </span>timesteps = torch.LongTensor([<span class="hljs-number">50</span>]) | |
| <span class="hljs-meta">>>> </span>noisy_image = noise_scheduler.add_noise(sample_image, noise, timesteps) | |
| <span class="hljs-meta">>>> </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>])<!-- HTML_TAG_END --></pre></div> | |
| <p><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/noisy_butterfly.png"></p> | |
| <p>모델의 학습 목적은 이미지에 더해진 노이즈를 예측하는 것입니다. 이 단계에서 손실은 다음과 같이 계산될 수 있습니다:</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> torch.nn.functional <span class="hljs-keyword">as</span> F | |
| <span class="hljs-meta">>>> </span>noise_pred = model(noisy_image, timesteps).sample | |
| <span class="hljs-meta">>>> </span>loss = F.mse_loss(noise_pred, noise)<!-- HTML_TAG_END --></pre></div> | |
| <h2 id="">모델 학습하기</h2> | |
| <p>지금까지, 모델 학습을 시작하기 위해 많은 부분을 갖추었으며 이제 남은 것은 모든 것을 조합하는 것입니다.</p> | |
| <p>우선 옵티마이저(optimizer)와 학습률 스케줄러(learning rate scheduler)가 필요할 것입니다:</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers.optimization <span class="hljs-keyword">import</span> get_cosine_schedule_with_warmup | |
| <span class="hljs-meta">>>> </span>optimizer = torch.optim.AdamW(model.parameters(), lr=config.learning_rate) | |
| <span class="hljs-meta">>>> </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>)<!-- HTML_TAG_END --></pre></div> | |
| <p>그 후, 모델을 평가하는 방법이 필요합니다. 평가를 위해, <code>DDPMPipeline</code>을 사용해 배치의 이미지 샘플들을 생성하고 그리드 형태로 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 class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDPMPipeline | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> math | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> os | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">def</span> <span class="hljs-title function_">make_grid</span>(<span class="hljs-params">images, rows, cols</span>): | |
| <span class="hljs-meta">... </span> w, h = images[<span class="hljs-number">0</span>].size | |
| <span class="hljs-meta">... </span> grid = Image.new(<span class="hljs-string">"RGB"</span>, size=(cols * w, rows * h)) | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">for</span> i, image <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(images): | |
| <span class="hljs-meta">... </span> grid.paste(image, box=(i % cols * w, i // cols * h)) | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> grid | |
| <span class="hljs-meta">>>> </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"># 랜덤한 노이즈로 부터 이미지를 추출합니다.(이는 역전파 diffusion 과정입니다.)</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># 기본 파이프라인 출력 형태는 `List[PIL.Image]` 입니다.</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.manual_seed(config.seed), | |
| <span class="hljs-meta">... </span> ).images | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># 이미지들을 그리드로 만들어줍니다.</span> | |
| <span class="hljs-meta">... </span> image_grid = make_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"># 이미지들을 저장합니다.</span> | |
| <span class="hljs-meta">... </span> test_dir = os.path.join(config.output_dir, <span class="hljs-string">"samples"</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"<span class="hljs-subst">{test_dir}</span>/<span class="hljs-subst">{epoch:04d}</span>.png"</span>)<!-- HTML_TAG_END --></pre></div> | |
| <p>TensorBoard에 로깅, 그래디언트 누적 및 혼합 정밀도 학습을 쉽게 수행하기 위해 🤗 Accelerate를 학습 루프에 함께 앞서 말한 모든 구성 정보들을 묶어 진행할 수 있습니다. 허브에 모델을 업로드 하기 위해 레포지토리 이름 및 정보를 가져오기 위한 함수를 작성하고 허브에 업로드할 수 있습니다.</p> | |
| <p>💡아래의 학습 루프는 어렵고 길어 보일 수 있지만, 나중에 한 줄의 코드로 학습을 한다면 그만한 가치가 있을 것입니다! 만약 기다리지 못하고 이미지를 생성하고 싶다면, 아래 코드를 자유롭게 붙여넣고 작동시키면 됩니다. 🤗</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> Accelerator | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> HfFolder, Repository, whoami | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> tqdm.auto <span class="hljs-keyword">import</span> tqdm | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> pathlib <span class="hljs-keyword">import</span> Path | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> os | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">def</span> <span class="hljs-title function_">get_full_repo_name</span>(<span class="hljs-params">model_id: <span class="hljs-built_in">str</span>, organization: <span class="hljs-built_in">str</span> = <span class="hljs-literal">None</span>, token: <span class="hljs-built_in">str</span> = <span class="hljs-literal">None</span></span>): | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> token <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span>: | |
| <span class="hljs-meta">... </span> token = HfFolder.get_token() | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">if</span> organization <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span>: | |
| <span class="hljs-meta">... </span> username = whoami(token)[<span class="hljs-string">"name"</span>] | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> <span class="hljs-string">f"<span class="hljs-subst">{username}</span>/<span class="hljs-subst">{model_id}</span>"</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">else</span>: | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> <span class="hljs-string">f"<span class="hljs-subst">{organization}</span>/<span class="hljs-subst">{model_id}</span>"</span> | |
| <span class="hljs-meta">>>> </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"># accelerator와 tensorboard 로깅 초기화</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">"tensorboard"</span>, | |
| <span class="hljs-meta">... </span> logging_dir=os.path.join(config.output_dir, <span class="hljs-string">"logs"</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.push_to_hub: | |
| <span class="hljs-meta">... </span> repo_name = get_full_repo_name(Path(config.output_dir).name) | |
| <span class="hljs-meta">... </span> repo = Repository(config.output_dir, clone_from=repo_name) | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">elif</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> accelerator.init_trackers(<span class="hljs-string">"train_example"</span>) | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># 모든 것이 준비되었습니다.</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># 기억해야 할 특정한 순서는 없으며 준비한 방법에 제공한 것과 동일한 순서로 객체의 압축을 풀면 됩니다.</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"># 이제 모델을 학습합니다.</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"Epoch <span class="hljs-subst">{epoch}</span>"</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">"images"</span>] | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># 이미지에 더할 노이즈를 샘플링합니다.</span> | |
| <span class="hljs-meta">... </span> noise = torch.randn(clean_images.shape).to(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"># 각 이미지를 위한 랜덤한 타임스텝(timestep)을 샘플링합니다.</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> ).long() | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># 각 타임스텝의 노이즈 크기에 따라 깨끗한 이미지에 노이즈를 추가합니다.</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># (이는 foward diffusion 과정입니다.)</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"># 노이즈를 반복적으로 예측합니다.</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> 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">"loss"</span>: loss.detach().item(), <span class="hljs-string">"lr"</span>: lr_scheduler.get_last_lr()[<span class="hljs-number">0</span>], <span class="hljs-string">"step"</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"># 각 에포크가 끝난 후 evaluate()와 몇 가지 데모 이미지를 선택적으로 샘플링하고 모델을 저장합니다.</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> repo.push_to_hub(commit_message=<span class="hljs-string">f"Epoch <span class="hljs-subst">{epoch}</span>"</span>, blocking=<span class="hljs-literal">True</span>) | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">else</span>: | |
| <span class="hljs-meta">... </span> pipeline.save_pretrained(config.output_dir)<!-- HTML_TAG_END --></pre></div> | |
| <p>휴, 코드가 꽤 많았네요! 하지만 🤗 Accelerate의 <code>notebook_launcher</code> 함수와 학습을 시작할 준비가 되었습니다. 함수에 학습 루프, 모든 학습 인수, 학습에 사용할 프로세스 수(사용 가능한 GPU의 수를 변경할 수 있음)를 전달합니다:</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> notebook_launcher | |
| <span class="hljs-meta">>>> </span>args = (config, model, noise_scheduler, optimizer, train_dataloader, lr_scheduler) | |
| <span class="hljs-meta">>>> </span>notebook_launcher(train_loop, args, num_processes=<span class="hljs-number">1</span>)<!-- HTML_TAG_END --></pre></div> | |
| <p>한번 학습이 완료되면, diffusion 모델로 생성된 최종 🦋이미지🦋를 확인해보길 바랍니다!</p> | |
| <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> | |
| <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> | |
| Copied</div></button></div> | |
| <pre><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> glob | |
| <span class="hljs-meta">>>> </span>sample_images = <span class="hljs-built_in">sorted</span>(glob.glob(<span class="hljs-string">f"<span class="hljs-subst">{config.output_dir}</span>/samples/*.png"</span>)) | |
| <span class="hljs-meta">>>> </span>Image.<span class="hljs-built_in">open</span>(sample_images[-<span class="hljs-number">1</span>])<!-- HTML_TAG_END --></pre></div> | |
| <p><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/butterflies_final.png"></p> | |
| <h2 id="">다음 단계</h2> | |
| <p>Unconditional 이미지 생성은 학습될 수 있는 작업 중 하나의 예시입니다. 다른 작업과 학습 방법은 <a href="../training/overview">🧨 Diffusers 학습 예시</a> 페이지에서 확인할 수 있습니다. 다음은 학습할 수 있는 몇 가지 예시입니다:</p> | |
| <ul><li><a href="../training/text_inversion">Textual Inversion</a>, 특정 시각적 개념을 학습시켜 생성된 이미지에 통합시키는 알고리즘입니다.</li> | |
| <li><a href="../training/dreambooth">DreamBooth</a>, 주제에 대한 몇 가지 입력 이미지들이 주어지면 주제에 대한 개인화된 이미지를 생성하기 위한 기술입니다.</li> | |
| <li><a href="../training/text2image">Guide</a> 데이터셋에 Stable Diffusion 모델을 파인튜닝하는 방법입니다.</li> | |
| <li><a href="../training/lora">Guide</a> LoRA를 사용해 매우 큰 모델을 빠르게 파인튜닝하기 위한 메모리 효율적인 기술입니다.</li></ul> | |
| <script type="module" data-hydrate="hs2q6g"> | |
| import { start } from "/docs/diffusers/v0.19.2/ko/_app/start-hf-doc-builder.js"; | |
| start({ | |
| target: document.querySelector('[data-hydrate="hs2q6g"]').parentNode, | |
| paths: {"base":"/docs/diffusers/v0.19.2/ko","assets":"/docs/diffusers/v0.19.2/ko"}, | |
| session: {}, | |
| route: false, | |
| spa: false, | |
| trailing_slash: "never", | |
| hydrate: { | |
| status: 200, | |
| error: null, | |
| nodes: [ | |
| import("/docs/diffusers/v0.19.2/ko/_app/pages/__layout.svelte-hf-doc-builder.js"), | |
| import("/docs/diffusers/v0.19.2/ko/_app/pages/tutorials/basic_training.mdx-hf-doc-builder.js") | |
| ], | |
| params: {} | |
| } | |
| }); | |
| </script> | |
Xet Storage Details
- Size:
- 63.3 kB
- Xet hash:
- 098577e9456ee2e9eb235b41fb7e7cb0035c52b64a895f949fd55d4a6c9218a2
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.