diffusers/pr_13881/zh/_app/immutable/nodes/44.B7239Shi.js

import"../chunks/DsnmJJEf.js";import{i as B,h as C,C as G,H as s,a,E as V,s as R}from"../chunks/DdZvggmf.js";import{p as N,o as F,s as l,f as Y,a as f,b as k,c as W,n as v}from"../chunks/BbekZcyp.js";const A='{"title":"InstructPix2Pix","local":"instructpix2pix","sections":[{"title":"脚本参数","local":"脚本参数","sections":[],"depth":2},{"title":"训练脚本","local":"训练脚本","sections":[],"depth":2},{"title":"启动脚本","local":"启动脚本","sections":[],"depth":2},{"title":"Stable Diffusion XL","local":"stable-diffusion-xl","sections":[],"depth":2},{"title":"后续步骤","local":"后续步骤","sections":[],"depth":2}],"depth":1}';var Q=W('<meta name="hf:doc:metadata"/>'),S=W(`<p></p> <!> <!> <p><a href="https://hf.co/papers/2211.09800" rel="nofollow">InstructPix2Pix</a> 是一个基于 Stable Diffusion 训练的模型，用于根据人类提供的指令编辑图像。例如，您的提示可以是“将云变成雨天”，模型将相应编辑输入图像。该模型以文本提示（或编辑指令）和输入图像为条件。</p> <p>本指南将探索 <a href="https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py" rel="nofollow">train_instruct_pix2pix.py</a> 训练脚本，帮助您熟悉它，以及如何将其适应您自己的用例。</p> <p>在运行脚本之前，请确保从源代码安装库：</p> <!> <p>然后导航到包含训练脚本的示例文件夹，并安装脚本所需的依赖项：</p> <!> <blockquote class="tip"><p>🤗 Accelerate 是一个库，用于帮助您在多个 GPU/TPU 上或使用混合精度进行训练。它将根据您的硬件和环境自动配置训练设置。查看 🤗 Accelerate <a href="https://huggingface.co/docs/accelerate/quicktour" rel="nofollow">快速导览</a> 以了解更多信息。</p></blockquote> <p>初始化一个 🤗 Accelerate 环境：</p> <!> <p>要设置一个默认的 🤗 Accelerate 环境，无需选择任何配置：</p> <!> <p>或者，如果您的环境不支持交互式 shell，例如笔记本，您可以使用：</p> <!> <p>最后，如果您想在自己的数据集上训练模型，请查看 <a href="create_dataset">创建用于训练的数据集</a> 指南，了解如何创建与训练脚本兼容的数据集。</p> <blockquote class="tip"><p>以下部分重点介绍了训练脚本中对于理解如何修改它很重要的部分，但并未详细涵盖脚本的每个方面。如果您有兴趣了解更多，请随时阅读 <a href="https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py" rel="nofollow">脚本</a>，并告诉我们如果您有任何问题或疑虑。</p></blockquote> <!> <p>训练脚本有许多参数可帮助您自定义训练运行。所有
参数及其描述可在 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L65" rel="nofollow"><code>parse_args()</code></a> 函数中找到。大多数参数都提供了默认值，这些值效果相当不错，但如果您愿意，也可以在训练命令中设置自己的值。</p> <p>例如，要增加输入图像的分辨率：</p> <!> <p>许多基本和重要的参数在 <a href="text2image#script-parameters">文本到图像</a> 训练指南中已有描述，因此本指南仅关注与 InstructPix2Pix 相关的参数：</p> <ul><li><code>--original_image_column</code>：编辑前的原始图像</li> <li><code>--edited_image_column</code>：编辑后的图像</li> <li><code>--edit_prompt_column</code>：编辑图像的指令</li> <li><code>--conditioning_dropout_prob</code>：训练期间编辑图像和编辑提示的 dropout 概率，这为一种或两种条件输入启用了无分类器引导（CFG）</li></ul> <!> <p>数据集预处理代码和训练循环可在 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L374" rel="nofollow"><code>main()</code></a> 函数中找到。这是您将修改训练脚本以适应自己用例的地方。</p> <p>与脚本参数类似，<a href="text2image#training-script">文本到图像</a> 训练指南提供了训练脚本的逐步说明。相反，本指南将查看脚本中与 InstructPix2Pix 相关的部分。</p> <p>脚本首先修改 UNet 的第一个卷积层中的 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L445" rel="nofollow">输入通道数</a>，以适应 InstructPix2Pix 的额外条件图像：</p> <!> <p>这些 UNet 参数由优化器 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L545C1-L551C6" rel="nofollow">更新</a>：</p> <!> <p>接下来，编辑后的图像和编辑指令被 <a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L624" rel="nofollow">预处理</a>并被<a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L610C24-L610C24" rel="nofollow">tokenized</a>。重要的是，对原始图像和编辑后的图像应用相同的图像变换。</p> <!> <p>最后，在<a href="https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L730" rel="nofollow">训练循环</a>中，它首先将编辑后的图像编码到潜在空间：</p> <!> <p>然后，脚本对原始图像和编辑指令嵌入应用 dropout 以支持 CFG（Classifier-Free Guidance）。这使得模型能够调节编辑指令和原始图像对编辑后图像的影响。</p> <!> <p>差不多就是这样了！除了这里描述的不同之处，脚本的其余部分与<a href="text2image#training-script">文本到图像</a>训练脚本非常相似，所以请随意查看以获取更多细节。如果您想了解更多关于训练循环如何工作的信息，请查看<a href="../using-diffusers/write_own_pipeline">理解管道、模型和调度器</a>教程，该教程分解了去噪过程的基本模式。</p> <!> <p>一旦您对脚本的更改感到满意，或者如果您对默认配置没问题，您
准备好启动训练脚本！🚀</p> <p>本指南使用 <a href="https://huggingface.co/datasets/fusing/instructpix2pix-1000-samples" rel="nofollow">fusing/instructpix2pix-1000-samples</a> 数据集，这是 <a href="https://huggingface.co/datasets/timbrooks/instructpix2pix-clip-filtered" rel="nofollow">原始数据集</a> 的一个较小版本。您也可以创建并使用自己的数据集（请参阅 <a href="create_dataset">创建用于训练的数据集</a> 指南）。</p> <p>将 <code>MODEL_NAME</code> 环境变量设置为模型名称（可以是 Hub 上的模型 ID 或本地模型的路径），并将 <code>DATASET_ID</code> 设置为 Hub 上数据集的名称。脚本会创建并保存所有组件（特征提取器、调度器、文本编码器、UNet 等）到您的仓库中的一个子文件夹。</p> <blockquote class="tip"><p>为了获得更好的结果，尝试使用更大的数据集进行更长时间的训练。我们只在较小规模的数据集上测试过此训练脚本。</p> <br/> <p>要使用 Weights and Biases 监控训练进度，请将 <code>--report_to=wandb</code> 参数添加到训练命令中，并使用 <code>--val_image_url</code> 指定验证图像，使用 <code>--validation_prompt</code> 指定验证提示。这对于调试模型非常有用。</p></blockquote> <p>如果您在多个 GPU 上训练，请将 <code>--multi_gpu</code> 参数添加到 <code>accelerate launch</code> 命令中。</p> <!> <p>训练完成后，您可以使用您的新 InstructPix2Pix 进行推理：</p> <!> <p>您应该尝试不同的 <code>num_inference_steps</code>、<code>image_guidance_scale</code> 和 <code>guidance_scale</code> 值，以查看它们如何影响推理速度和质量。指导比例参数
这些参数尤其重要，因为它们控制原始图像和编辑指令对编辑后图像的影响程度。</p> <!> <p>Stable Diffusion XL (SDXL) 是一个强大的文本到图像模型，能够生成高分辨率图像，并在其架构中添加了第二个文本编码器。使用 <a href="https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py" rel="nofollow"><code>train_instruct_pix2pix_sdxl.py</code></a> 脚本来训练 SDXL 模型以遵循图像编辑指令。</p> <p>SDXL 训练脚本在 <a href="sdxl">SDXL 训练</a> 指南中有更详细的讨论。</p> <!> <p>恭喜您训练了自己的 InstructPix2Pix 模型！🥳 要了解更多关于该模型的信息，可能有助于：</p> <ul><li>阅读 <a href="https://huggingface.co/blog/instruction-tuning-sd" rel="nofollow">Instruction-tuning Stable Diffusion with InstructPix2Pix</a> 博客文章，了解更多我们使用 InstructPix2Pix 进行的一些实验、数据集准备以及不同指令的结果。</li></ul> <!> <p></p>`,1);function L(X,I){N(I,!1),F(()=>{new URLSearchParams(window.location.search).get("fw")}),B();var e=S();C("bny2f6",j=>{var Z=Q();R(Z,"content",A),f(j,Z)});var n=l(Y(e),2);G(n,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var i=l(n,2);s(i,{title:"InstructPix2Pix",local:"instructpix2pix",headingTag:"h1"});var t=l(i,8);a(t,{code:"Z2l0JTIwY2xvbmUlMjBodHRwcyUzQSUyRiUyRmdpdGh1Yi5jb20lMkZodWdnaW5nZmFjZSUyRmRpZmZ1c2VycyUwQWNkJTIwZGlmZnVzZXJzJTBBcGlwJTIwaW5zdGFsbCUyMC4=",highlighted:`git <span class="hljs-built_in">clone</span> https://github.com/huggingface/diffusers
<span class="hljs-built_in">cd</span> diffusers
pip install .`,lang:"bash",wrap:!1});var p=l(t,4);a(p,{code:"Y2QlMjBleGFtcGxlcyUyRmluc3RydWN0X3BpeDJwaXglMEFwaXAlMjBpbnN0YWxsJTIwLXIlMjByZXF1aXJlbWVudHMudHh0",highlighted:`<span class="hljs-built_in">cd</span> examples/instruct_pix2pix
pip install -r requirements.txt`,lang:"bash",wrap:!1});var c=l(p,6);a(c,{code:"YWNjZWxlcmF0ZSUyMGNvbmZpZw==",highlighted:"accelerate config",lang:"bash",wrap:!1});var o=l(c,4);a(o,{code:"YWNjZWxlcmF0ZSUyMGNvbmZpZyUyMGRlZmF1bHQ=",highlighted:"accelerate config default",lang:"bash",wrap:!1});var d=l(o,4);a(d,{code:"ZnJvbSUyMGFjY2VsZXJhdGUudXRpbHMlMjBpbXBvcnQlMjB3cml0ZV9iYXNpY19jb25maWclMEElMEF3cml0ZV9iYXNpY19jb25maWcoKQ==",highlighted:`<span class="hljs-keyword">from</span> accelerate.utils <span class="hljs-keyword">import</span> write_basic_config

write_basic_config()`,lang:"py",wrap:!1});var r=l(d,6);s(r,{title:"脚本参数",local:"脚本参数",headingTag:"h2"});var M=l(r,6);a(M,{code:"YWNjZWxlcmF0ZSUyMGxhdW5jaCUyMHRyYWluX2luc3RydWN0X3BpeDJwaXgucHklMjAlNUMlMEElMjAlMjAtLXJlc29sdXRpb24lM0Q1MTIlMjAlNUM=",highlighted:`accelerate launch train_instruct_pix2pix.py \\
  --resolution=512 \\`,lang:"bash",wrap:!1});var u=l(M,6);s(u,{title:"训练脚本",local:"训练脚本",headingTag:"h2"});var h=l(u,8);a(h,{code:"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",highlighted:`in_channels = <span class="hljs-number">8</span>
out_channels = unet.conv_in.out_channels
unet.register_to_config(in_channels=in_channels)

<span class="hljs-keyword">with</span> torch.no_grad():
    new_conv_in = nn.Conv2d(
        in_channels, out_channels, unet.conv_in.kernel_size, unet.conv_in.stride, unet.conv_in.padding
    )
    new_conv_in.weight.zero_()
    new_conv_in.weight[:, :<span class="hljs-number">4</span>, :, :].copy_(unet.conv_in.weight)
    unet.conv_in = new_conv_in`,lang:"py",wrap:!1});var m=l(h,4);a(m,{code:"b3B0aW1pemVyJTIwJTNEJTIwb3B0aW1pemVyX2NscyglMEElMjAlMjAlMjAlMjB1bmV0LnBhcmFtZXRlcnMoKSUyQyUwQSUyMCUyMCUyMCUyMGxyJTNEYXJncy5sZWFybmluZ19yYXRlJTJDJTBBJTIwJTIwJTIwJTIwYmV0YXMlM0QoYXJncy5hZGFtX2JldGExJTJDJTIwYXJncy5hZGFtX2JldGEyKSUyQyUwQSUyMCUyMCUyMCUyMHdlaWdodF9kZWNheSUzRGFyZ3MuYWRhbV93ZWlnaHRfZGVjYXklMkMlMEElMjAlMjAlMjAlMjBlcHMlM0RhcmdzLmFkYW1fZXBzaWxvbiUyQyUwQSk=",highlighted:`optimizer = optimizer_cls(
    unet.parameters(),
    lr=args.learning_rate,
    betas=(args.adam_beta1, args.adam_beta2),
    weight_decay=args.adam_weight_decay,
    eps=args.adam_epsilon,
)`,lang:"py",wrap:!1});var J=l(m,4);a(J,{code:"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",highlighted:`<span class="hljs-keyword">def</span> <span class="hljs-title function_">preprocess_train</span>(<span class="hljs-params">examples</span>):
    preprocessed_images = preprocess_images(examples)

    original_images, edited_images = preprocessed_images.chunk(<span class="hljs-number">2</span>)
    original_images = original_images.reshape(-<span class="hljs-number">1</span>, <span class="hljs-number">3</span>, args.resolution, args.resolution)
    edited_images = edited_images.reshape(-<span class="hljs-number">1</span>, <span class="hljs-number">3</span>, args.resolution, args.resolution)

    examples[<span class="hljs-string">&quot;original_pixel_values&quot;</span>] = original_images
    examples[<span class="hljs-string">&quot;edited_pixel_values&quot;</span>] = edited_images

    captions = <span class="hljs-built_in">list</span>(examples[edit_prompt_column])
    examples[<span class="hljs-string">&quot;input_ids&quot;</span>] = tokenize_captions(captions)
    <span class="hljs-keyword">return</span> examples`,lang:"py",wrap:!1});var y=l(J,4);a(y,{code:"bGF0ZW50cyUyMCUzRCUyMHZhZS5lbmNvZGUoYmF0Y2glNUIlMjJlZGl0ZWRfcGl4ZWxfdmFsdWVzJTIyJTVELnRvKHdlaWdodF9kdHlwZSkpLmxhdGVudF9kaXN0LnNhbXBsZSgpJTBBbGF0ZW50cyUyMCUzRCUyMGxhdGVudHMlMjAqJTIwdmFlLmNvbmZpZy5zY2FsaW5nX2ZhY3Rvcg==",highlighted:`latents = vae.encode(batch[<span class="hljs-string">&quot;edited_pixel_values&quot;</span>].to(weight_dtype)).latent_dist.sample()
latents = latents * vae.config.scaling_factor`,lang:"py",wrap:!1});var b=l(y,4);a(b,{code:"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",highlighted:`encoder_hidden_states = text_encoder(batch[<span class="hljs-string">&quot;input_ids&quot;</span>])[<span class="hljs-number">0</span>]
original_image_embeds = vae.encode(batch[<span class="hljs-string">&quot;original_pixel_values&quot;</span>].to(weight_dtype)).latent_dist.mode()

<span class="hljs-keyword">if</span> args.conditioning_dropout_prob <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>:
    random_p = torch.rand(bsz, device=latents.device, generator=generator)
    prompt_mask = random_p &lt; <span class="hljs-number">2</span> * args.conditioning_dropout_prob
    prompt_mask = prompt_mask.reshape(bsz, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>)
    null_conditioning = text_encoder(tokenize_captions([<span class="hljs-string">&quot;&quot;</span>]).to(accelerator.device))[<span class="hljs-number">0</span>]
    encoder_hidden_states = torch.where(prompt_mask, null_conditioning, encoder_hidden_states)

    image_mask_dtype = original_image_embeds.dtype
    image_mask = <span class="hljs-number">1</span> - (
        (random_p &gt;= args.conditioning_dropout_prob).to(image_mask_dtype)
        * (random_p &lt; <span class="hljs-number">3</span> * args.conditioning_dropout_prob).to(image_mask_dtype)
    )
    image_mask = image_mask.reshape(bsz, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>)
    original_image_embeds = image_mask * original_image_embeds`,lang:"py",wrap:!1});var w=l(b,4);s(w,{title:"启动脚本",local:"启动脚本",headingTag:"h2"});var g=l(w,12);a(g,{code:"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",highlighted:`accelerate launch --mixed_precision=<span class="hljs-string">&quot;fp16&quot;</span> train_instruct_pix2pix.py \\
    --pretrained_model_name_or_path=<span class="hljs-variable">$MODEL_NAME</span> \\
    --dataset_name=<span class="hljs-variable">$DATASET_ID</span> \\
    --enable_xformers_memory_efficient_attention \\
    --resolution=256 \\
    --random_flip \\
    --train_batch_size=4 \\
    --gradient_accumulation_steps=4 \\
    --gradient_checkpointing \\
    --max_train_steps=15000 \\
    --checkpointing_steps=5000 \\
    --checkpoints_total_limit=1 \\
    --learning_rate=5e-05 \\
    --max_grad_norm=1 \\
    --lr_warmup_steps=0 \\
    --conditioning_dropout_prob=0.05 \\
    --mixed_precision=fp16 \\
    --seed=42 \\
    --push_to_hub`,lang:"bash",wrap:!1});var _=l(g,4);a(_,{code:"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",highlighted:`<span class="hljs-keyword">import</span> PIL
<span class="hljs-keyword">import</span> requests
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionInstructPix2PixPipeline
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> load_image

pipeline = StableDiffusionInstructPix2PixPipeline.from_pretrained(<span class="hljs-string">&quot;your_cool_model&quot;</span>, torch_dtype=torch.float16).to(<span class="hljs-string">&quot;cuda&quot;</span>)
generator = torch.Generator(<span class="hljs-string">&quot;cuda&quot;</span>).manual_seed(<span class="hljs-number">0</span>)

image = load_image(<span class="hljs-string">&quot;https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/test_pix2pix_4.png&quot;</span>)
prompt = <span class="hljs-string">&quot;add some ducks to the lake&quot;</span>
num_inference_steps = <span class="hljs-number">20</span>
image_guidance_scale = <span class="hljs-number">1.5</span>
guidance_scale = <span class="hljs-number">10</span>

edited_image = pipeline(
   prompt,
   image=image,
   num_inference_steps=num_inference_steps,
   image_guidance_scale=image_guidance_scale,
   guidance_scale=guidance_scale,
   generator=generator,
).images[<span class="hljs-number">0</span>]
edited_image.save(<span class="hljs-string">&quot;edited_image.png&quot;</span>)`,lang:"py",wrap:!1});var T=l(_,4);s(T,{title:"Stable Diffusion XL",local:"stable-diffusion-xl",headingTag:"h2"});var U=l(T,6);s(U,{title:"后续步骤",local:"后续步骤",headingTag:"h2"});var x=l(U,6);V(x,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/zh/training/instructpix2pix.md"}),v(2),f(X,e),k()}export{L as component};