diffusers/pr_13881/zh/_app/immutable/nodes/50.BmPwp3a6.js

import"../chunks/DsnmJJEf.js";import{i as R,h as G,C as I,H as a,a as l,E as V,s as B}from"../chunks/DdZvggmf.js";import{p as x,o as C,s as e,f as N,a as j,b as E,c as T,n as F}from"../chunks/BbekZcyp.js";const Q='{"title":"Wuerstchen","local":"wuerstchen","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":"下一步","local":"下一步","sections":[],"depth":2}],"depth":1}';var Y=T('<meta name="hf:doc:metadata"/>'),A=T(`<p></p> <!> <!> <p><a href="https://hf.co/papers/2306.00637" rel="nofollow">Wuerstchen</a> 模型通过将潜在空间压缩 42 倍，在不影响图像质量的情况下大幅降低计算成本并加速推理。在训练过程中，Wuerstchen 使用两个模型（VQGAN + 自动编码器）来压缩潜在表示，然后第三个模型（文本条件潜在扩散模型）在这个高度压缩的空间上进行条件化以生成图像。</p> <p>为了将先验模型放入 GPU 内存并加速训练，尝试分别启用 <code>gradient_accumulation_steps</code>、<code>gradient_checkpointing</code> 和 <code>mixed_precision</code>。</p> <p>本指南探讨 <a href="https://github.com/huggingface/diffusers/blob/main/examples/wuerstchen/text_to_image/train_text_to_image_prior.py" rel="nofollow">train_text_to_image_prior.py</a> 脚本，帮助您更熟悉它，以及如何根据您的用例进行适配。</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 是一个帮助您在多个 GPU/TPU 上或使用混合精度进行训练的库。它会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">快速入门</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 环境而不选择任何配置：</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">创建训练数据集</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，但并未涵盖 <a href="https://github.com/huggingface/diffusers/blob/main/examples/wuerstchen/text_to_image/train_text_to_image_prior.py" rel="nofollow">脚本</a> 的详细信息。如果您有兴趣了解更多，请随时阅读脚本，并告诉我们您是否有任何问题或疑虑。</p></blockquote> <!> <p>训练脚本提供了许多参数来帮助您自定义训练运行。所有参数及其描述都可以在 <a href="https://github.com/huggingface/diffusers/blob/6e68c71503682c8693cb5b06a4da4911dfd655ee/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L192" rel="nofollow"><code>parse_args()</code></a> 函数中找到。它为每个参数提供了默认值，例如训练批次大小和学习率，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，要使用 fp16 格式的混合精度加速训练，请在训练命令中添加 <code>--mixed_precision</code> 参数：</p> <!> <p>大多数参数与 <a href="text2image#script-parameters">文本到图像</a> 训练指南中的参数相同，因此让我们直接深入 Wuerstchen 训练脚本！</p> <!> <p>训练脚本也与 <a href="text2image#training-script">文本到图像</a> 训练指南类似，但已修改以支持 Wuerstchen。本指南重点介绍 Wuerstchen 训练脚本中独特的代码。</p> <p><a href="https://github.com/huggingface/diffusers/blob/6e68c71503682c8693cb5b06a4da4911dfd655ee/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L441" rel="nofollow"><code>main()</code></a> 函数首先初始化图像编码器 - 一个 <a href="https://github.com/huggingface/diffusers/blob/main/examples/wuerstchen/text_to_image/modeling_efficient_net_encoder.py" rel="nofollow">EfficientNet</a> - 以及通常的调度器和分词器。</p> <!> <p>您还将加载 <code>WuerstchenPrior</code> 模型以进行优化。</p> <!> <p>接下来，您将对图像应用一些 <a href="https://github.com/huggingface/diffusers/blob/65ef7a0c5c594b4f84092e328fbdd73183613b30/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L656" rel="nofollow">transforms</a> 并对标题进行 <a href="https://github.com/huggingface/diffusers/blob/65ef7a0c5c594b4f84092e328fbdd73183613b30/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L637" rel="nofollow">tokenize</a>：</p> <!> <p>最后，<a href="https://github.com/huggingface/diffusers/blob/65ef7a0c5c594b4f84092e328fbdd73183613b30/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L656" rel="nofollow">训练循环</a>处理使用<code>EfficientNetEncoder</code>将图像压缩到潜在空间，向潜在表示添加噪声，并使用<code>WuerstchenPrior</code>模型预测噪声残差。</p> <!> <p>如果您想了解更多关于训练循环的工作原理，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>一旦您完成了所有更改或对默认配置满意，就可以启动训练脚本了！🚀</p> <p>设置<code>DATASET_NAME</code>环境变量为Hub中的数据集名称。本指南使用<a href="https://huggingface.co/datasets/lambdalabs/naruto-blip-captions" rel="nofollow">Naruto BLIP captions</a>数据集，但您也可以创建和训练自己的数据集（参见<a href="create_dataset">创建用于训练的数据集</a>指南）。</p> <blockquote class="tip"><p>要使用Weights & Biases监控训练进度，请在训练命令中添加<code>--report_to=wandb</code>参数。您还需要在训练命令中添加<code>--validation_prompt</code>以跟踪结果。这对于调试模型和查看中间结果非常有用。</p></blockquote> <!> <p>训练完成后，您可以使用新训练的模型进行推理！</p> <!> <!> <p>恭喜您训练了一个Wuerstchen模型！要了解更多关于如何使用您的新模型的信息，请参
以下内容可能有所帮助：</p> <ul><li>查看 <a href="../api/pipelines/wuerstchen#text-to-image-generation">Wuerstchen</a> API 文档，了解更多关于如何使用该管道进行文本到图像生成及其限制的信息。</li></ul> <!> <p></p>`,1);function H(Z,W){x(W,!1),C(()=>{new URLSearchParams(window.location.search).get("fw")}),R();var t=A();G("q4kcug",f=>{var U=Y();B(U,"content",Q),j(f,U)});var s=e(N(t),2);I(s,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var n=e(s,2);a(n,{title:"Wuerstchen",local:"wuerstchen",headingTag:"h1"});var o=e(n,10);l(o,{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=e(o,4);l(p,{code:"Y2QlMjBleGFtcGxlcyUyRnd1ZXJzdGNoZW4lMkZ0ZXh0X3RvX2ltYWdlJTBBcGlwJTIwaW5zdGFsbCUyMC1yJTIwcmVxdWlyZW1lbnRzLnR4dA==",highlighted:`<span class="hljs-built_in">cd</span> examples/wuerstchen/text_to_image
pip install -r requirements.txt`,lang:"bash",wrap:!1});var i=e(p,6);l(i,{code:"YWNjZWxlcmF0ZSUyMGNvbmZpZw==",highlighted:"accelerate config",lang:"bash",wrap:!1});var c=e(i,4);l(c,{code:"YWNjZWxlcmF0ZSUyMGNvbmZpZyUyMGRlZmF1bHQ=",highlighted:"accelerate config default",lang:"bash",wrap:!1});var r=e(c,4);l(r,{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 d=e(r,6);a(d,{title:"脚本参数",local:"脚本参数",headingTag:"h2"});var h=e(d,6);l(h,{code:"YWNjZWxlcmF0ZSUyMGxhdW5jaCUyMHRyYWluX3RleHRfdG9faW1hZ2VfcHJpb3IucHklMjAlNUMlMEElMjAlMjAtLW1peGVkX3ByZWNpc2lvbiUzRCUyMmZwMTYlMjI=",highlighted:`accelerate launch train_text_to_image_prior.py \\
  --mixed_precision=<span class="hljs-string">&quot;fp16&quot;</span>`,lang:"bash",wrap:!1});var M=e(h,4);a(M,{title:"训练脚本",local:"训练脚本",headingTag:"h2"});var m=e(M,6);l(m,{code:"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",highlighted:`<span class="hljs-keyword">with</span> ContextManagers(deepspeed_zero_init_disabled_context_manager()):
    pretrained_checkpoint_file = hf_hub_download(<span class="hljs-string">&quot;dome272/wuerstchen&quot;</span>, filename=<span class="hljs-string">&quot;model_v2_stage_b.pt&quot;</span>)
    state_dict = torch.load(pretrained_checkpoint_file, map_location=<span class="hljs-string">&quot;cpu&quot;</span>)
    image_encoder = EfficientNetEncoder()
    image_encoder.load_state_dict(state_dict[<span class="hljs-string">&quot;effnet_state_dict&quot;</span>])
    image_encoder.<span class="hljs-built_in">eval</span>()`,lang:"py",wrap:!1});var y=e(m,4);l(y,{code:"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",highlighted:`prior = WuerstchenPrior.from_pretrained(args.pretrained_prior_model_name_or_path, subfolder=<span class="hljs-string">&quot;prior&quot;</span>)

optimizer = optimizer_cls(
    prior.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 b=e(y,4);l(b,{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>):
    images = [image.conver
t(<span class="hljs-string">&quot;RGB&quot;</span>) <span class="hljs-keyword">for</span> image <span class="hljs-keyword">in</span> examples[image_column]]
    examples[<span class="hljs-string">&quot;effnet_pixel_values&quot;</span>] = [effnet_transforms(image) <span class="hljs-keyword">for</span> image <span class="hljs-keyword">in</span> images]
    examples[<span class="hljs-string">&quot;text_input_ids&quot;</span>], examples[<span class="hljs-string">&quot;text_mask&quot;</span>] = tokenize_captions(examples)
    <span class="hljs-keyword">return</span> examples`,lang:"py",wrap:!1});var u=e(b,4);l(u,{code:"cHJlZF9ub2lzZSUyMCUzRCUyMHByaW9yKG5vaXN5X2xhdGVudHMlMkMlMjB0aW1lc3RlcHMlMkMlMjBwcm9tcHRfZW1iZWRzKQ==",highlighted:"pred_noise = prior(noisy_latents, timesteps, prompt_embeds)",lang:"py",wrap:!1});var _=e(u,4);a(_,{title:"启动脚本",local:"启动脚本",headingTag:"h2"});var g=e(_,8);l(g,{code:"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",highlighted:`<span class="hljs-built_in">export</span> DATASET_NAME=<span class="hljs-string">&quot;lambdalabs/naruto-blip-captions&quot;</span>

accelerate launch  train_text_to_image_prior.py \\
  --mixed_precision=<span class="hljs-string">&quot;fp16&quot;</span> \\
  --dataset_name=<span class="hljs-variable">$DATASET_NAME</span> \\
  --resolution=768 \\
  --train_batch_size=4 \\
  --gradient_accumulation_steps=4 \\
  --gradient_checkpointing \\
  --dataloader_num_workers=4 \\
  --max_train_steps=15000 \\
  --learning_rate=1e-05 \\
  --max_grad_norm=1 \\
  --checkpoints_total_limit=3 \\
  --lr_scheduler=<span class="hljs-string">&quot;constant&quot;</span> \\
  --lr_warmup_steps=0 \\
  --validation_prompts=<span class="hljs-string">&quot;A robot naruto, 4k photo&quot;</span> \\
  --report_to=<span class="hljs-string">&quot;wandb&quot;</span> \\
  --push_to_hub \\
  --output_dir=<span class="hljs-string">&quot;wuerstchen-prior-naruto-model&quot;</span>`,lang:"bash",wrap:!1});var J=e(g,4);l(J,{code:"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",highlighted:`<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoPipelineForText2Image
<span class="hljs-keyword">from</span> diffusers.pipelines.wuerstchen <span class="hljs-keyword">import</span> DEFAULT_STAGE_C_TIMESTEPS

pipeline = AutoPipelineForText2Image.from_pretrained(<span class="hljs-string">&quot;path/to/saved/model&quot;</span>, torch_dtype=torch.float16).to(<span class="hljs-string">&quot;cuda&quot;</span>)

caption = <span class="hljs-string">&quot;A cute bird naruto holding a shield&quot;</span>
images = pipeline(
    caption,
    width=<span class="hljs-number">1024</span>,
    height=<span class="hljs-number">1536</span>,
    prior_timesteps=DEFAULT_STAGE_C_TIMESTEPS,
    prior_guidance_scale=<span class="hljs-number">4.0</span>,
    num_images_per_prompt=<span class="hljs-number">2</span>,
).images`,lang:"py",wrap:!1});var w=e(J,2);a(w,{title:"下一步",local:"下一步",headingTag:"h2"});var X=e(w,6);V(X,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/zh/training/wuerstchen.md"}),F(2),j(Z,t),E()}export{H as component};