pai-python-sdk/inference/model_deploy_container/model_deploy_container.ipynb (392 lines of code) (raw):
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 使用镜像部署模型\n",
"\n",
"PAI支持用户使用镜像的方式部署模型，通过镜像，开发者可以自定义模型部署的环境，包括Python、使用的机器学习框架、依赖的第三方库等，能够支持用户灵活的部署需求。详细的介绍可以参考PAI帮助文档：[使用镜像部署模型](https://help.aliyun.com/zh/pai/user-guide/deploy-a-model-service-by-using-a-custom-image)。\n",
"\n",
"PAI Python SDK提供了便利的API，支持用户能够使用自定义镜像，或是PAI提供的预置推理，将一个本地，或是OSS上的模型快捷得部署为模型在线服务。\n",
"\n",
"本文档将介绍，用户如何通过PAI Python SDK通过自定义镜像的方式部署模型。"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"## 费用说明\n",
"\n",
"本示例将会使用以下云产品，并产生相应的费用账单：\n",
"\n",
"- PAI-EAS：部署推理服务，详细计费说明请参考[PAI-EAS计费说明](https://help.aliyun.com/zh/pai/product-overview/billing-of-eas)\n",
"- OSS：存储模型和推理等，详细计费说明请参考[OSS计费概述](https://help.aliyun.com/zh/oss/product-overview/billing-overview)\n",
"\n",
"> 通过参与云产品免费试用，使用**指定资源机型**，可以免费试用PAI产品，具体请参考[PAI免费试用](https://help.aliyun.com/zh/pai/product-overview/free-quota-for-new-users)。\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"## 安装和配置SDK\n",
"\n",
"我们需要首先安装PAI Python SDK以运行本示例。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!python -m pip install --upgrade pai"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"\n",
"SDK需要配置访问阿里云服务需要的AccessKey，以及当前使用的工作空间和OSS Bucket。在PAI SDK安装之后，通过在 **命令行终端** 中执行以下命令，按照引导配置密钥、工作空间等信息。\n",
"\n",
"\n",
"```shell\n",
"\n",
"# 以下命令，请在 命令行终端 中执行.\n",
"\n",
"python -m pai.toolkit.config\n",
"\n",
"```\n",
"\n",
"我们可以通过以下代码验证配置是否已生效。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import pai\n",
"from pai.session import get_default_session\n",
"\n",
"print(pai.__version__)\n",
"\n",
"sess = get_default_session()\n",
"\n",
"# 获取配置的工作空间信息\n",
"assert sess.workspace_name is not None\n",
"print(sess.workspace_name)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 部署模型推理服务\n",
"\n",
"模型在线服务包含了模型的文件、模型的推理服务代码、以及推理服务运行环境。\n",
"本示例将使用一个简单的`PyTorch`模型，通过`Flask`和`PAI`提供的`PyTorch`基础镜像，部署模型在线服务。\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"下载示例使用的简单PyTorch模型。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# 下载模型到本地 \"model\" 目录\n",
"\n",
"!mkdir -p model/\n",
"!wget https://pai-sdk.oss-cn-shanghai.aliyuncs.com/pai/resources/toy_model.pt -P model/"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 准备推理服务代码\n",
"\n",
"在部署模型之前，我们首先需要准备推理服务的代码，它提供HTTP接口，负责接收预测请求，使用模型进行推理，返回预测结果。\n",
"\n",
"当前示例我们将使用 ``Flask`` 编写一个简单的推理服务，保存为 ``infer_src/app.py`` 文件。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!mkdir -p infer_src"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"%%writefile infer_src/app.py\n",
"import json\n",
"from flask import Flask, request\n",
"import os\n",
"import torch\n",
"import numpy as np\n",
"\n",
"app = Flask(__name__)\n",
"model = None\n",
"# 默认的模型文件路径\n",
"MODEL_PATH = \"/eas/workspace/model/\"\n",
"\n",
"def load_model():\n",
" \"\"\"加载模型\"\"\"\n",
" global model\n",
" model = torch.jit.load(os.path.join(MODEL_PATH, \"toy_model.pt\"))\n",
" model.eval()\n",
"\n",
"@app.route(\"/\", methods=[\"POST\"])\n",
"def predict():\n",
" data = np.asarray(json.loads(request.data)).astype(np.float32)\n",
" output_tensor = model(torch.from_numpy(data))\n",
" pred_res = output_tensor.detach().cpu().numpy()\n",
" return json.dumps(pred_res.tolist())\n",
"\n",
"if __name__ == \"__main__\":\n",
" load_model()\n",
" app.run(host=\"0.0.0.0\", port=int(os.environ.get(\"LISTENING_PORT\", 8000)))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 获取PAI提供的预置推理镜像\n",
"\n",
"PAI提供了一系列预置的推理镜像，镜像内预置了机器学习框架、常用的第三方库、Python、NVIDIA CUDA库等。我们可以通过以下代码列出所有的预置镜像。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from pai.image import list_images, ImageScope\n",
"\n",
"\n",
"data = [\n",
" [\n",
" \"ImageUri\",\n",
" \"FrameworkName\",\n",
" \"FrameworkVersion\",\n",
" \"AcceleratorType\",\n",
" \"PythonVersion\",\n",
" ]\n",
"]\n",
"\n",
"# 列出常用的PyTorch推理镜像\n",
"for img in list_images(framework_name=\"PyTorch\", image_scope=ImageScope.INFERENCE):\n",
" data.append(\n",
" [\n",
" img.image_uri,\n",
" img.framework_name,\n",
" img.framework_version,\n",
" img.accelerator_type,\n",
" img.python_version,\n",
" ]\n",
" )\n",
"\n",
"# 列出常用的TensorFlow推理镜像\n",
"for img in list_images(framework_name=\"TensorFlow\", image_scope=ImageScope.INFERENCE):\n",
" data.append(\n",
" [\n",
" img.image_uri,\n",
" img.framework_name,\n",
" img.framework_version,\n",
" img.accelerator_type,\n",
" img.python_version,\n",
" ]\n",
" )"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from IPython.display import HTML, display\n",
"\n",
"display(\n",
" HTML(\n",
" \"<table><tr>{}</tr></table>\".format(\n",
" \"</tr><tr>\".join(\n",
" \"<td>{}</td>\".format(\"</td><td>\".join(str(_) for _ in row))\n",
" for row in data\n",
" )\n",
" )\n",
" )\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"通过SDK提供的 `pai.image.retrieve` API，可以获取指定框架版本的镜像。在当前示例中，我们将使用PAI提供的PyTorch 1.12版本的CPU推理镜像"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from pai.image import retrieve, ImageScope\n",
"\n",
"# # 获取PyTorch 1.10 GPU推理镜像\n",
"# print(retrieve(\n",
"# framework_name=\"PyTorch\", # 框架名称\n",
"# framework_version=\"latest\", # 框架版本\n",
"# accelerator_type=\"gpu\", # 选择支持Nvidia CUDA GPU的镜像\n",
"# image_scope=ImageScope.INFERENCE, # 镜像类型，推理镜像\n",
"\n",
"# # ).image_uri)\n",
"\n",
"# 获取最新的PyTorch CPU推理镜像\n",
"torch_image_uri = retrieve(\n",
" framework_name=\"PyTorch\", # 框架名称\n",
" framework_version=\"1.12\", # 框架版本，latest表示使用PAI支持的最新版本\n",
" # accelerator_type=\"cpu\", # 默认使用CPU镜像\n",
" image_scope=ImageScope.INFERENCE, # 镜像类型，推理镜像\n",
").image_uri\n",
"print(torch_image_uri)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 部署推理服务\n",
"使用以上的推理服务代码，以及PyTorch推理镜像，我们将一个PyTorch模型部署为模型在线服务。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from pai.model import Model, container_serving_spec\n",
"\n",
"\n",
"m = Model(\n",
" model_data=\"./model/\", # 模型文件，可以是一个本地文件或是OSS Bucket路径(例如 oss://<BucketName>/path/to/model )，\n",
" inference_spec=container_serving_spec(\n",
" image_uri=torch_image_uri, # 推理服务使用的镜像\n",
" command=\"python app.py\", # 模型推理服务启动命令\n",
" source_dir=\"./infer_src/\", # 推理服务代码所在目录\n",
" requirements=[\"flask==2.0.0\", \"Werkzeug==2.3.4\"], # 推理服务依赖的Python包\n",
" ),\n",
")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from pai.common.utils import random_str\n",
"\n",
"# 部署模型服务\n",
"p = m.deploy(\n",
" service_name=f\"toy_model_{random_str(6)}\", # 模型服务名称, 地域内唯一\n",
" instance_type=\"ecs.c6.large\", # 模型服务使用的机器实例规格\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 调用推理服务\n",
"\n",
"部署服务后返回的`pai.predictor.Predictor`对象可以用于调用推理服务，发送预测请求。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import numpy as np\n",
"\n",
"# 构造一个随机数组输入\n",
"dummy_input = np.random.rand(1, 10, 10).tolist()\n",
"print(dummy_input)\n",
"\n",
"result = p.raw_predict(\n",
" data=dummy_input,\n",
")\n",
"\n",
"# 打印推理结果\n",
"print(result.json())"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"在测试完成之后，删除推理服务"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"p.delete_service()"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "base",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.8.3"
}
},
"nbformat": 4,
"nbformat_minor": 2
}