Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /main /ko /using-diffusers /reproducibility.html
rtrm's picture
rtrm
download
raw
31 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;재현 가능한 파이프라인 생성하기&quot;,&quot;local&quot;:&quot;재현-가능한-파이프라인-생성하기&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;무작위성 제어하기&quot;,&quot;local&quot;:&quot;무작위성-제어하기&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CPU&quot;,&quot;local&quot;:&quot;cpu&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;GPU&quot;,&quot;local&quot;:&quot;gpu&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;결정론적 알고리즘&quot;,&quot;local&quot;:&quot;결정론적-알고리즘&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
 <link href="/docs/diffusers/main/ko/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/entry/start.0574fe93.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/scheduler.e3739aa0.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/singletons.6c595ea8.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/index.02f4400c.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/paths.b2b36d06.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/entry/app.f5608ddb.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/index.13f5b837.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/nodes/0.a357fdeb.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/each.e59479a4.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/nodes/49.4e1d7105.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/Tip.51dedfeb.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/CodeBlock.de02009a.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/DocNotebookDropdown.5cbf7847.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/globals.7f7f1b26.js">
 <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/EditOnGithub.72bac8d8.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;재현 가능한 파이프라인 생성하기&quot;,&quot;local&quot;:&quot;재현-가능한-파이프라인-생성하기&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;무작위성 제어하기&quot;,&quot;local&quot;:&quot;무작위성-제어하기&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CPU&quot;,&quot;local&quot;:&quot;cpu&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;GPU&quot;,&quot;local&quot;:&quot;gpu&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;결정론적 알고리즘&quot;,&quot;local&quot;:&quot;결정론적-알고리즘&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="재현-가능한-파이프라인-생성하기" 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="#재현-가능한-파이프라인-생성하기"><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>재현 가능한 파이프라인 생성하기</span></h1> <div class="flex space-x-1 absolute z-10 right-0 top-0"> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"> </button> </div> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"> </button> </div></div> <p data-svelte-h="svelte-1jbr77x">재현성은 테스트, 결과 재현, 그리고 <a href="resuing_seeds">이미지 퀄리티 높이기</a>에서 중요합니다.
그러나 diffusion 모델의 무작위성은 매번 모델이 돌아갈 때마다 파이프라인이 다른 이미지를 생성할 수 있도록 하는 이유로 필요합니다.
플랫폼 간에 정확하게 동일한 결과를 얻을 수는 없지만, 특정 허용 범위 내에서 릴리스 및 플랫폼 간에 결과를 재현할 수는 있습니다.
그럼에도 diffusion 파이프라인과 체크포인트에 따라 허용 오차가 달라집니다.</p> <p data-svelte-h="svelte-hcwx4f">diffusion 모델에서 무작위성의 원천을 제어하거나 결정론적 알고리즘을 사용하는 방법을 이해하는 것이 중요한 이유입니다.</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 data-svelte-h="svelte-3yn5i4">💡 Pytorch의 <a href="https://pytorch.org/docs/stable/notes/randomness.html" rel="nofollow">재현성에 대한 선언</a>를 꼭 읽어보길 추천합니다:</p> <blockquote data-svelte-h="svelte-va9th1"><p>완전하게 재현가능한 결과는 Pytorch 배포, 개별적인 커밋, 혹은 다른 플랫폼들에서 보장되지 않습니다.
또한, 결과는 CPU와 GPU 실행간에 심지어 같은 seed를 사용할 때도 재현 가능하지 않을 수 있습니다.</p></blockquote></div> <h2 class="relative group"><a id="무작위성-제어하기" 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="#무작위성-제어하기"><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>무작위성 제어하기</span></h2> <p data-svelte-h="svelte-rovd75">추론에서, 파이프라인은 노이즈를 줄이기 위해 가우시안 노이즈를 생성하거나 스케줄링 단계에 노이즈를 더하는 등의 랜덤 샘플링 실행에 크게 의존합니다,</p> <p data-svelte-h="svelte-1oq8xcu"><a href="https://huggingface.co/docs/diffusers/v0.18.0/en/api/pipelines/ddim#diffusers.DDIMPipeline" rel="nofollow">DDIMPipeline</a>에서 두 추론 단계 이후의 텐서 값을 살펴보세요:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDIMPipeline
<span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np
model_id = <span class="hljs-string">&quot;google/ddpm-cifar10-32&quot;</span>
<span class="hljs-comment"># 모델과 스케줄러를 불러오기</span>
ddim = DDIMPipeline.from_pretrained(model_id)
<span class="hljs-comment"># 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기</span>
image = ddim(num_inference_steps=<span class="hljs-number">2</span>, output_type=<span class="hljs-string">&quot;np&quot;</span>).images
<span class="hljs-built_in">print</span>(np.<span class="hljs-built_in">abs</span>(image).<span class="hljs-built_in">sum</span>())<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-zj7eo2">위의 코드를 실행하면 하나의 값이 나오지만, 다시 실행하면 다른 값이 나옵니다. 무슨 일이 일어나고 있는 걸까요?</p> <p data-svelte-h="svelte-znzen0">파이프라인이 실행될 때마다, <a href="https://pytorch.org/docs/stable/generated/torch.randn.html" rel="nofollow">torch.randn</a>은
단계적으로 노이즈 제거되는 가우시안 노이즈가 생성하기 위한 다른 랜덤 seed를 사용합니다.</p> <p data-svelte-h="svelte-tun6p5">그러나 동일한 이미지를 안정적으로 생성해야 하는 경우에는 CPU에서 파이프라인을 실행하는지 GPU에서 실행하는지에 따라 달라집니다.</p> <h3 class="relative group"><a id="cpu" 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="#cpu"><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>CPU</span></h3> <p data-svelte-h="svelte-1ozlo3o">CPU에서 재현 가능한 결과를 생성하려면, PyTorch <a href="https://pytorch.org/docs/stable/generated/torch.randn.html" rel="nofollow">Generator</a>로 seed를 고정합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDIMPipeline
<span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np
model_id = <span class="hljs-string">&quot;google/ddpm-cifar10-32&quot;</span>
<span class="hljs-comment"># 모델과 스케줄러 불러오기</span>
ddim = DDIMPipeline.from_pretrained(model_id)
<span class="hljs-comment"># 재현성을 위해 generator 만들기</span>
generator = torch.Generator(device=<span class="hljs-string">&quot;cpu&quot;</span>).manual_seed(<span class="hljs-number">0</span>)
<span class="hljs-comment"># 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기</span>
image = ddim(num_inference_steps=<span class="hljs-number">2</span>, output_type=<span class="hljs-string">&quot;np&quot;</span>, generator=generator).images
<span class="hljs-built_in">print</span>(np.<span class="hljs-built_in">abs</span>(image).<span class="hljs-built_in">sum</span>())<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1fjfua9">이제 위의 코드를 실행하면 seed를 가진 <code>Generator</code> 객체가 파이프라인의 모든 랜덤 함수에 전달되므로 항상 <code>1491.1711</code> 값이 출력됩니다.</p> <p data-svelte-h="svelte-j3qgnm">특정 하드웨어 및 PyTorch 버전에서 이 코드 예제를 실행하면 동일하지는 않더라도 유사한 결과를 얻을 수 있습니다.</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 data-svelte-h="svelte-8pt0v2">💡 처음에는 시드를 나타내는 정수값 대신에 <code>Generator</code> 개체를 파이프라인에 전달하는 것이 약간 비직관적일 수 있지만,
<code>Generator</code>는 순차적으로 여러 파이프라인에 전달될 수 있는 \랜덤상태\이기 때문에 PyTorch에서 확률론적 모델을 다룰 때 권장되는 설계입니다.</p></div> <h3 class="relative group"><a id="gpu" 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="#gpu"><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>GPU</span></h3> <p data-svelte-h="svelte-1cafmef">예를 들면, GPU 상에서 같은 코드 예시를 실행하면:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDIMPipeline
<span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np
model_id = <span class="hljs-string">&quot;google/ddpm-cifar10-32&quot;</span>
<span class="hljs-comment"># 모델과 스케줄러 불러오기</span>
ddim = DDIMPipeline.from_pretrained(model_id)
ddim.to(<span class="hljs-string">&quot;cuda&quot;</span>)
<span class="hljs-comment"># 재현성을 위한 generator 만들기</span>
generator = torch.Generator(device=<span class="hljs-string">&quot;cuda&quot;</span>).manual_seed(<span class="hljs-number">0</span>)
<span class="hljs-comment"># 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기</span>
image = ddim(num_inference_steps=<span class="hljs-number">2</span>, output_type=<span class="hljs-string">&quot;np&quot;</span>, generator=generator).images
<span class="hljs-built_in">print</span>(np.<span class="hljs-built_in">abs</span>(image).<span class="hljs-built_in">sum</span>())<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-141n7wf">GPU가 CPU와 다른 난수 생성기를 사용하기 때문에 동일한 시드를 사용하더라도 결과가 같지 않습니다.</p> <p data-svelte-h="svelte-7f5ryc">이 문제를 피하기 위해 🧨 Diffusers는 CPU에 임의의 노이즈를 생성한 다음 필요에 따라 텐서를 GPU로 이동시키는
<a href="https://huggingface.co/docs/diffusers/v0.18.0/en/api/utilities#diffusers.utils.randn_tensor" rel="nofollow">randn_tensor()</a>기능을 가지고 있습니다.
<code>randn_tensor</code> 기능은 파이프라인 내부 어디에서나 사용되므로 파이프라인이 GPU에서 실행되더라도 <strong>항상</strong> CPU <code>Generator</code>를 통과할 수 있습니다.</p> <p data-svelte-h="svelte-1xwk1mu">이제 결과에 훨씬 더 다가왔습니다!</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDIMPipeline
<span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np
model_id = <span class="hljs-string">&quot;google/ddpm-cifar10-32&quot;</span>
<span class="hljs-comment"># 모델과 스케줄러 불러오기</span>
ddim = DDIMPipeline.from_pretrained(model_id)
ddim.to(<span class="hljs-string">&quot;cuda&quot;</span>)
<span class="hljs-comment">#재현성을 위한 generator 만들기 (GPU에 올리지 않도록 조심한다!)</span>
generator = torch.manual_seed(<span class="hljs-number">0</span>)
<span class="hljs-comment"># 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기</span>
image = ddim(num_inference_steps=<span class="hljs-number">2</span>, output_type=<span class="hljs-string">&quot;np&quot;</span>, generator=generator).images
<span class="hljs-built_in">print</span>(np.<span class="hljs-built_in">abs</span>(image).<span class="hljs-built_in">sum</span>())<!-- HTML_TAG_END --></pre></div> <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 data-svelte-h="svelte-784xbp">💡 재현성이 중요한 경우에는 항상 CPU generator를 전달하는 것이 좋습니다.
성능 손실은 무시할 수 없는 경우가 많으며 파이프라인이 GPU에서 실행되었을 때보다 훨씬 더 비슷한 값을 생성할 수 있습니다.</p></div> <p data-svelte-h="svelte-1evdu50">마지막으로 <a href="https://huggingface.co/docs/diffusers/v0.18.0/en/api/pipelines/unclip#diffusers.UnCLIPPipeline" rel="nofollow">UnCLIPPipeline</a>과 같은
더 복잡한 파이프라인의 경우, 이들은 종종 정밀 오차 전파에 극도로 취약합니다.
다른 GPU 하드웨어 또는 PyTorch 버전에서 유사한 결과를 기대하지 마세요.
이 경우 완전한 재현성을 위해 완전히 동일한 하드웨어 및 PyTorch 버전을 실행해야 합니다.</p> <h2 class="relative group"><a id="결정론적-알고리즘" 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="#결정론적-알고리즘"><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>결정론적 알고리즘</span></h2> <p data-svelte-h="svelte-124dlbu">결정론적 알고리즘을 사용하여 재현 가능한 파이프라인을 생성하도록 PyTorch를 구성할 수도 있습니다.
그러나 결정론적 알고리즘은 비결정론적 알고리즘보다 느리고 성능이 저하될 수 있습니다.
하지만 재현성이 중요하다면, 이것이 최선의 방법입니다!</p> <p data-svelte-h="svelte-1pjtl3n">둘 이상의 CUDA 스트림에서 작업이 시작될 때 비결정론적 동작이 발생합니다.
이 문제를 방지하려면 환경 변수 <a href="https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility" rel="nofollow">CUBLAS_WORKSPACE_CONFIG</a>를 <code>:16:8</code>로 설정해서
런타임 중에 오직 하나의 버퍼 크리만 사용하도록 설정합니다.</p> <p data-svelte-h="svelte-wj31jr">PyTorch는 일반적으로 가장 빠른 알고리즘을 선택하기 위해 여러 알고리즘을 벤치마킹합니다.
하지만 재현성을 원하는 경우, 벤치마크가 매 순간 다른 알고리즘을 선택할 수 있기 때문에 이 기능을 사용하지 않도록 설정해야 합니다.
마지막으로, <a href="https://pytorch.org/docs/stable/generated/torch.use_deterministic_algorithms.html" rel="nofollow">torch.use_deterministic_algorithms</a>에
<code>True</code>를 통과시켜 결정론적 알고리즘이 활성화 되도록 합니다.</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> os
os.environ[<span class="hljs-string">&quot;CUBLAS_WORKSPACE_CONFIG&quot;</span>] = <span class="hljs-string">&quot;:16:8&quot;</span>
torch.backends.cudnn.benchmark = <span class="hljs-literal">False</span>
torch.use_deterministic_algorithms(<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-uyrq4d">이제 동일한 파이프라인을 두번 실행하면 동일한 결과를 얻을 수 있습니다.</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDIMScheduler, StableDiffusionPipeline
<span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np
model_id = <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>
pipe = StableDiffusionPipeline.from_pretrained(model_id).to(<span class="hljs-string">&quot;cuda&quot;</span>)
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
g = torch.Generator(device=<span class="hljs-string">&quot;cuda&quot;</span>)
prompt = <span class="hljs-string">&quot;A bear is playing a guitar on Times Square&quot;</span>
g.manual_seed(<span class="hljs-number">0</span>)
result1 = pipe(prompt=prompt, num_inference_steps=<span class="hljs-number">50</span>, generator=g, output_type=<span class="hljs-string">&quot;latent&quot;</span>).images
g.manual_seed(<span class="hljs-number">0</span>)
result2 = pipe(prompt=prompt, num_inference_steps=<span class="hljs-number">50</span>, generator=g, output_type=<span class="hljs-string">&quot;latent&quot;</span>).images
<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;L_inf dist = &quot;</span>, <span class="hljs-built_in">abs</span>(result1 - result2).<span class="hljs-built_in">max</span>())
<span class="hljs-string">&quot;L_inf dist = tensor(0., device=&#x27;cuda:0&#x27;)&quot;</span><!-- HTML_TAG_END --></pre></div> <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/ko/using-diffusers/reproducibility.md" target="_blank"><span data-svelte-h="svelte-1kd6by1">&lt;</span> <span data-svelte-h="svelte-x0xyl0">&gt;</span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>
<script>
 {
 __sveltekit_198thv9 = {
 assets: "/docs/diffusers/main/ko",
 base: "/docs/diffusers/main/ko",
 env: {}
 };
 const element = document.currentScript.parentElement;
 const data = [null,null];
 Promise.all([
 import("/docs/diffusers/main/ko/_app/immutable/entry/start.0574fe93.js"),
 import("/docs/diffusers/main/ko/_app/immutable/entry/app.f5608ddb.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 49],
 data,
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
31 kB
·
Xet hash:
9d15983f1ea2b1dbc6b3c95b79cf83ab22b26c3fb8760fb9d0ff7a5c235aa174

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