如何将PyTorch Lightning模型部署到生产中

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"大规模服务PyTorch Lightning模型的完整指南。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/2a/2af6ef5180b94a7f67f50706f5e45a72.jpeg","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"纵观机器学习领域，主要趋势之一是专注于将软件工程原理应用于机器学习的项目激增。 例如，"},{"type":"link","attrs":{"href":"https://github.com/cortexlabs/cortex","title":null},"content":[{"type":"text","text":"Cortex"}]},{"type":"text","text":"再现了部署无服务器功能但具有推理管道的体验。类似地，DVC实现了现代版本控制和CI / CD管道，但仅用于ML。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"PyTorch Lightning具有类似的理念，仅适用于训练。框架为PyTorch提供了Python包装器，可让数据科学家和工程师编写干净，可管理且性能卓越的训练代码。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作为构建 整个部署平台的人 ，部分原因是我们讨厌编写样板，因此我们是PyTorch Lightning的忠实拥护者。本着这种精神，我整理了将PyTorch Lightning模型部署到生产环境的指南。在此过程中，我们将研究几种导出PyTorch Lightning模型以包括在推理管道中的选项。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"部署PyTorch Lightning模型进行推理的每种方法"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有三种方法可以导出PyTorch Lightning模型进行投放："}]},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将模型另存为PyTorch检查点"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将模型转换为ONNX"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将模型导出到Torchscript"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们可以通过Cortex为这三个服务。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1.直接打包和部署PyTorch Lightning模块"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从最简单的方法开始，让我们部署一个没有任何转换步骤的PyTorch Lightning模型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"PyTorch Lightning Trainer是一个抽象样板训练代码（思考训练和验证步骤）的类，它具有内置的save_checkpoint（）函数，该函数会将您的模型另存为.ckpt文件。要将模型保存为检查点，只需将以下代码添加到训练脚本中："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/02/025d681f83213cf3d7ce1654847b7be8.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"现在，在我们开始为该检查点提供服务之前，需要注意的是，虽然我一直说“ PyTorch Lightning模型”，但PyTorch Lightning是PyTorch的包装器-项目的README字面意思是“ PyTorch Lightning只是有组织的PyTorch。” 因此，导出的模型是普通的PyTorch模型，可以相应地使用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有了保存的检查点，我们可以在Cortex中轻松地为模型提供服务。如果您不熟悉Cortex，可以 "},{"type":"link","attrs":{"href":"https://docs.cortex.dev/","title":null},"content":[{"type":"text","text":"在这里快速熟悉一下"}]},{"type":"text","text":"，但是Cortex部署过程的简单概述是："}]},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们使用Python为我们的模型编写了一个预测API"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们在YAML中定义我们的API基础结构和行为"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们使用CLI中的命令部署API"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们的预测API将使用Cortex的Python Predictor类定义一个init（）函数来初始化我们的API并加载模型，并使用一个define（）函数在查询时提供预测："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/d6/d6a2224d5d2b032568e54e234730313f.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"很简单 我们从训练代码中重新调整了一些代码的用途，添加了一些推理逻辑，仅此而已。需要注意的一件事是，如果将模型上传到S3（推荐），则需要添加一些逻辑来访问它。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"接下来，我们在YAML中配置基础架构："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/dd/dd55dff424462894bebc088e789b4803.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"再次，简单。我们给我们的API起个名字，告诉Cortex我们的预测API在哪里，并分配一些CPU。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"接下来，我们部署它："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/ab/ab1263a55e354288c69ea8c91acae5ec.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"请注意，我们还可以部署到集群，由Cortex加速和管理："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/b8/b8c81954a32fd5b219ca9ef79961bae0.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在所有部署中，Cortex都会容器化我们的API并将其公开为Web服务。通过云部署，Cortex可以配置负载平衡，自动扩展，监视，更新和许多其他基础架构功能。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"就是这样！现在，我们有一个实时Web API，可根据要求提供模型预测。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2.导出到ONNX并通过ONNX运行时进行投放"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"现在，我们已经部署了一个普通的PyTorch检查点，让事情复杂一些。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"PyTorch Lightning最近添加了一个方便的抽象，用于将模型导出到ONNX（以前，您可以使用PyTorch的内置转换功能，尽管它们需要更多样板）。要将模型导出到ONNX，只需将以下代码添加到您的训练脚本中："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/15/15ab17629a85d7908b2da8580641e3b6.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"请注意，您的输入样本应模仿实际模型输入的形状。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"导出ONNX模型后，就可以使用Cortex的ONNX Predictor为其提供服务。代码基本上看起来是相同的，并且过程是相同的。例如，这是一个ONNX预测API："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/eb/ebe30166a5e0eeedd0170a9658c66515.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基本上一样。唯一的区别是，我们不是通过直接初始化模型，而是通过onnx_client访问该数据，这是Cortex为服务于我们的模型而启动的ONNX运行时容器。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们的YAML看起来也很相似："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/ed/ed9c580485af9b3744717a2c388147b1.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我在此处添加了一个监视标志，目的只是为了显示配置的简便程度，并且有一些ONNX特定字段，但其他方面都是相同的YAML。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"最后，我们使用与之前相同的$ cortex deploy命令进行部署，并且我们的ONNX API已启用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"3.使用Torchscript的JIT编译器进行序列化"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"对于最终部署，我们将把PyTorch Lightning模型导出到Torchscript并使用PyTorch的JIT编译器提供服务。要导出模型，只需将其添加到您的训练脚本中："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/31/313b86adee9ed266007585a717b42bf5.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"用于此目的的Python API与原始PyTorch示例几乎相同："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/3c/3c407844ca45ed4f88d95ef80acfef32.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"YAML保持与以前相同，并且CLI命令当然是一致的。如果需要的话，我们实际上可以更新我们以前的PyTorch API来使用新模型，只需将新的旧的dictor.py脚本替换为新的脚本，然后再次运行$ cortex部署："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/12/12f3991706f8eddc7a82b8b48cd97da9.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Cortex在此处自动执行滚动更新，在该更新中，新的API会被启动，然后与旧的API交换，从而避免了模型更新之间的任何停机时间。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这就是全部。现在，您已经具有用于实时推理的完全可操作的预测API，可根据Torchscript模型提供预测。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"那么，您应该使用哪种方法？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"显而易见的问题是哪种方法效果最好。事实是，这里没有简单的答案，因为这取决于您的模型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"对于BERT和GPT-2等Transformer模型，ONNX可以提供令人难以置信的优化（我们测得的"},{"type":"link","attrs":{"href":"https://www.cortex.dev/post/40x-nlp-inference-with-hugging-face-and-onnx","title":null},"content":[{"type":"text","text":"CPU吞吐量提高"}]},{"type":"text","text":"了"},{"type":"link","attrs":{"href":"https://www.cortex.dev/post/40x-nlp-inference-with-hugging-face-and-onnx","title":null},"content":[{"type":"text","text":"40"}]},{"type":"text","text":"倍 ）。对于其他模型，Torchscript的性能可能比香草PyTorch更好-尽管这也带有一些警告，因为并非所有模型都干净地导出到Torchscript。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"幸运的是，使用任何选项进行部署都很容易，您可以并行测试所有这三个选项，并查看哪种方式最适合您的特定API。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"如果你喜欢本文的话，欢迎点赞转发！谢谢。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"看完别走还有惊喜！"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"我精心整理了计算机/Python/机器学习/深度学习相关的2TB视频课与书籍，价值1W元。关注微信公众号“计算机与AI”，点击下方菜单即可获取网盘链接。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":""}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/90/903237ffd0a3b3ae06272386f26ecb9e.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}