1. 简介
概览
在此 Codelab 中，您将创建一个使用 Node.js 编写的 Cloud Run 作业，用于提供视频中每个场景的直观说明。首先，您的作业将使用 Video Intelligence API 检测每次场景变化的时间戳。接下来，您的作业将使用名为 ffmpeg 的第三方二进制文件捕获每个场景切换时间戳的屏幕截图。最后，Vertex AI 可视化字幕用于提供屏幕截图的视觉说明。
此 Codelab 还演示了如何在 Cloud Run 作业中使用 ffmpeg 从视频中捕获给定时间戳的图片。由于 ffmpeg 需要单独安装，因此此 Codelab 会介绍如何创建一个 Dockerfile，以便在 Cloud Run 作业中安装 ffmpeg。
下图说明了 Cloud Run 作业的工作原理：
学习内容
- 如何使用 Dockerfile 创建容器映像以安装第三方二进制文件
- 如何通过为 Cloud Run 作业创建服务账号来调用其他 Google Cloud 服务来遵循最小权限原则
- 如何通过 Cloud Run 作业使用 Video Intelligence 客户端库
- 如何调用 Google API 以获取 Vertex AI 中每个场景的直观描述
2. 设置和要求
前提条件
- 您已登录 Cloud 控制台。
- 您之前已部署了 Cloud Run 服务。例如，您可以按照快速入门：部署 Web 服务开始操作。
激活 Cloud Shell
- 在 Cloud Console 中，点击激活 Cloud Shell
。
如果这是您第一次启动 Cloud Shell，系统会显示一个中间屏幕，说明它是什么。如果您看到中间屏幕，请点击继续。
预配和连接到 Cloud Shell 只需花几分钟时间。
这个虚拟机装有所需的所有开发工具。它提供了一个持久的 5 GB 主目录，并在 Google Cloud 中运行，大大增强了网络性能和身份验证功能。您在此 Codelab 中的大部分（即使不是全部）工作都可以通过浏览器完成。
在连接到 Cloud Shell 后，您应该会看到自己已通过身份验证，并且相关项目已设为您的项目 ID。
- 在 Cloud Shell 中运行以下命令以确认您已通过身份验证：
gcloud auth list
命令输出
Credentialed Accounts
ACTIVE ACCOUNT
* <my_account>@<my_domain.com>
To set the active account, run:
$ gcloud config set account `ACCOUNT`
- 在 Cloud Shell 中运行以下命令，以确认 gcloud 命令了解您的项目：
gcloud config list project
命令输出
[core] project = <PROJECT_ID>
如果不是上述结果，您可以使用以下命令进行设置：
gcloud config set project <PROJECT_ID>
命令输出
Updated property [core/project].
3. 启用 API 并设置环境变量
在开始使用此 Codelab 之前，您需要先启用多个 API。此 Codelab 需要使用以下 API。您可以通过运行以下命令来启用这些 API：
gcloud services enable run.googleapis.com \
storage.googleapis.com \
cloudbuild.googleapis.com \
videointelligence.googleapis.com \
aiplatform.googleapis.com
然后，您可以设置要在整个 Codelab 中使用的环境变量。
REGION=<YOUR-REGION> PROJECT_ID=<YOUR-PROJECT-ID> PROJECT_NUMBER=$(gcloud projects describe $PROJECT_ID --format='value(projectNumber)') JOB_NAME=video-describer-job BUCKET_ID=$PROJECT_ID-video-describer SERVICE_ACCOUNT="cloud-run-job-video" SERVICE_ACCOUNT_ADDRESS=$SERVICE_ACCOUNT@$PROJECT_ID.iam.gserviceaccount.com
4. 创建服务账号
您将为 Cloud Run 作业创建一个服务账号，用于访问 Cloud Storage、Vertex AI 和 Video Intelligence API。
首先，创建服务账号。
gcloud iam service-accounts create $SERVICE_ACCOUNT \ --display-name="Cloud Run Video Scene Image Describer service account"
然后向服务账号授予对 Cloud Storage 存储分区和 Vertex AI API 的访问权限。
# to view & download storage bucket objects gcloud projects add-iam-policy-binding $PROJECT_ID \ --member serviceAccount:$SERVICE_ACCOUNT_ADDRESS \ --role=roles/storage.objectViewer # to call the Vertex AI imagetext model gcloud projects add-iam-policy-binding $PROJECT_ID \ --member serviceAccount:$SERVICE_ACCOUNT_ADDRESS \ --role=roles/aiplatform.user
5. 创建 Cloud Storage 存储分区
使用以下命令创建一个 Cloud Storage 存储分区，您可以在其中上传视频以供 Cloud Run 作业进行处理：
gsutil mb -l us-central1 gs://$BUCKET_ID/
[可选] 您可以下载此示例视频，将其下载到本地，加以使用。
gsutil cp gs://cloud-samples-data/video/visionapi.mp4 testvideo.mp4
现在，将您的视频文件上传到存储分区中。
FILENAME=<YOUR-VIDEO-FILENAME> gsutil cp $FILENAME gs://$BUCKET_ID
6. 创建 Cloud Run 作业
首先，为源代码创建一个目录，然后通过 cd 命令进入该目录。
mkdir video-describer-job && cd $_
然后，创建一个包含以下内容的 package.json 文件：
{
"name": "video-describer-job",
"version": "1.0.0",
"private": true,
"description": "describes the image in every scene for a given video",
"main": "app.js",
"author": "Google LLC",
"license": "Apache-2.0",
"scripts": {
"start": "node app.js"
},
"dependencies": {
"@google-cloud/storage": "^7.7.0",
"@google-cloud/video-intelligence": "^5.0.1",
"axios": "^1.6.2",
"fluent-ffmpeg": "^2.1.2",
"google-auth-library": "^9.4.1"
}
}
为提高可读性，此应用包含多个源文件。首先，创建一个包含以下内容的 app.js 源文件。此文件包含作业的入口点以及应用的主要逻辑。
const bucketName = "<YOUR_BUCKET_ID>";
const videoFilename = "<YOUR-VIDEO-FILENAME>";
const { captureImages } = require("./helpers/imageCapture.js");
const { detectSceneChanges } = require("./helpers/sceneDetector.js");
const { getImageCaption } = require("./helpers/imageCaptioning.js");
const storageHelper = require("./helpers/storage.js");
const authHelper = require("./helpers/auth.js");
const fs = require("fs").promises;
const path = require("path");
const main = async () => {
try {
// download the file to locally to the Cloud Run Job instance
let localFilename = await storageHelper.downloadVideoFile(
bucketName,
videoFilename
);
// PART 1 - Use Video Intelligence API
// detect all the scenes in the video & save timestamps to an array
// EXAMPLE OUTPUT
// Detected scene changes at the following timestamps:
// [1, 7, 11, 12]
let timestamps = await detectSceneChanges(localFilename);
console.log(
"Detected scene changes at the following timestamps: ",
timestamps
);
// PART 2 - Use ffmpeg via dockerfile install
// create an image of each scene change
// and save to a local directory called "output"
// returns the base filename for the generated images
// EXAMPLE OUTPUT
// creating screenshot for scene: 1 at output/video-filename-1.png
// creating screenshot for scene: 7 at output/video-filename-7.png
// creating screenshot for scene: 11 at output/video-filename-11.png
// creating screenshot for scene: 12 at output/video-filename-12.png
// returns the base filename for the generated images
let imageBaseName = await captureImages(localFilename, timestamps);
// PART 3a - get Access Token to call Vertex AI APIs via REST
// needed for the image captioning
// since we're calling the Vertex AI APIs directly
let accessToken = await authHelper.getAccessToken();
console.log("got an access token");
// PART 3b - use Image Captioning to describe each scene per screenshot
// EXAMPLE OUTPUT
/*
[
{
timestamp: 1,
description:
"an aerial view of a city with a bridge in the background"
},
{
timestamp: 7,
description:
"a man in a blue shirt sits in front of shelves of donuts"
},
{
timestamp: 11,
description:
"a black and white photo of people working in a bakery"
},
{
timestamp: 12,
description:
"a black and white photo of a man and woman working in a bakery"
}
]; */
// instantiate the data structure for storing the scene description and timestamp
// e.g. an array of json objects,
// [{ timestamp: 5, description: "..." }, ...]
let scenes = [];
// for each timestamp, send the image to Vertex AI
console.log("getting Vertex AI description for each timestamps");
scenes = await Promise.all(
timestamps.map(async (timestamp) => {
let filepath = path.join(
"./output",
imageBaseName + "-" + timestamp + ".png"
);
// get the base64 encoded image bc sending via REST
const encodedFile = await fs.readFile(filepath, "base64");
// send each screenshot to Vertex AI for description
let description = await getImageCaption(
accessToken,
encodedFile
);
return { timestamp: timestamp, description: description };
})
);
console.log("finished collecting all the scenes");
console.log(scenes);
} catch (error) {
//return an error
console.error("received error: ", error);
}
};
// Start script
main().catch((err) => {
console.error(err);
});
接下来，创建 Dockerfile。
# Copyright 2020 Google, LLC. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # Use the official lightweight Node.js image. # https://hub.docker.com/_/node FROM node:20.10.0-slim # Create and change to the app directory. WORKDIR /usr/src/app RUN apt-get update && apt-get install -y ffmpeg # Copy application dependency manifests to the container image. # A wildcard is used to ensure both package.json AND package-lock.json are copied. # Copying this separately prevents re-running npm install on every code change. COPY package*.json ./ # Install dependencies. # If you add a package-lock.json speed your build by switching to 'npm ci'. # RUN npm ci --only=production RUN npm install --production # Copy local code to the container image. COPY . . # Run the job on container startup. CMD [ "npm", "start" ]
创建一个名为 .dockerignore 的文件，以忽略某些文件的容器化。
Dockerfile .dockerignore node_modules npm-debug.log
现在，创建一个名为 helpers 的文件夹。此文件夹将包含 5 个帮助程序文件。
mkdir helpers cd helpers
接下来，创建一个包含以下内容的 sceneDetector.js 文件。此文件使用 Video Intelligence API 检测视频中的场景何时发生变化。
const fs = require("fs");
const util = require("util");
const readFile = util.promisify(fs.readFile);
const ffmpeg = require("fluent-ffmpeg");
const Video = require("@google-cloud/video-intelligence");
const client = new Video.VideoIntelligenceServiceClient();
module.exports = {
detectSceneChanges: async function (downloadedFile) {
// Reads a local video file and converts it to base64
const file = await readFile(downloadedFile);
const inputContent = file.toString("base64");
// setup request for shot change detection
const videoContext = {
speechTranscriptionConfig: {
languageCode: "en-US",
enableAutomaticPunctuation: true
}
};
const request = {
inputContent: inputContent,
features: ["SHOT_CHANGE_DETECTION"]
};
// Detects camera shot changes
const [operation] = await client.annotateVideo(request);
console.log("Shot (scene) detection in progress...");
const [operationResult] = await operation.promise();
// Gets shot changes
const shotChanges =
operationResult.annotationResults[0].shotAnnotations;
console.log(
"Shot (scene) changes detected: " + shotChanges.length
);
// data structure to be returned
let sceneChanges = [];
// for the initial scene
sceneChanges.push(1);
// if only one scene, keep at 1 second
if (shotChanges.length === 1) {
return sceneChanges;
}
// get length of video
const videoLength = await getVideoLength(downloadedFile);
shotChanges.forEach((shot, shotIndex) => {
if (shot.endTimeOffset === undefined) {
shot.endTimeOffset = {};
}
if (shot.endTimeOffset.seconds === undefined) {
shot.endTimeOffset.seconds = 0;
}
if (shot.endTimeOffset.nanos === undefined) {
shot.endTimeOffset.nanos = 0;
}
// convert to a number
let currentTimestampSecond = Number(
shot.endTimeOffset.seconds
);
let sceneChangeTime = 0;
// double-check no scenes were detected within the last second
if (currentTimestampSecond + 1 > videoLength) {
sceneChangeTime = currentTimestampSecond;
} else {
// otherwise, for simplicity, just round up to the next second
sceneChangeTime = currentTimestampSecond + 1;
}
sceneChanges.push(sceneChangeTime);
});
return sceneChanges;
}
};
async function getVideoLength(localFile) {
let getLength = util.promisify(ffmpeg.ffprobe);
let length = await getLength(localFile);
console.log("video length: ", length.format.duration);
return length.format.duration;
}
现在，创建一个名为 imageCapture.js 且包含以下内容的文件。此文件使用节点软件包 fluent-ffmpeg 从节点应用中运行 ffmpeg 命令。
const ffmpeg = require("fluent-ffmpeg");
const path = require("path");
const util = require("util");
module.exports = {
captureImages: async function (localFile, scenes) {
let imageBaseName = path.parse(localFile).name;
try {
for (scene of scenes) {
console.log("creating screenshot for scene: ", +scene);
await createScreenshot(localFile, imageBaseName, scene);
}
} catch (error) {
console.log("error gathering screenshots: ", error);
}
console.log("finished gathering the screenshots");
return imageBaseName; // return the base filename for each image
}
};
async function createScreenshot(localFile, imageBaseName, scene) {
return new Promise((resolve, reject) => {
ffmpeg(localFile)
.screenshots({
timestamps: [scene],
filename: `${imageBaseName}-${scene}.png`,
folder: "output",
size: "320x240"
})
.on("error", () => {
console.log(
"Failed to create scene for timestamp: " + scene
);
return reject(
"Failed to create scene for timestamp: " + scene
);
})
.on("end", () => {
return resolve();
});
});
}
最后，创建一个名为 imageCaptioning.js 且包含以下内容的文件。此文件使用 Vertex AI 获取每张场景图片的直观描述。
const axios = require("axios");
const { GoogleAuth } = require("google-auth-library");
const auth = new GoogleAuth({
scopes: "https://www.googleapis.com/auth/cloud-platform"
});
module.exports = {
getImageCaption: async function (token, encodedFile) {
// this example shows you how to call the Vertex REST APIs directly
// https://cloud.google.com/vertex-ai/generative-ai/docs/image/image-captioning#get-captions-short
// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/image-captioning
let projectId = await auth.getProjectId();
let config = {
headers: {
"Authorization": "Bearer " + token,
"Content-Type": "application/json; charset=utf-8"
}
};
const json = {
"instances": [
{
"image": {
"bytesBase64Encoded": encodedFile
}
}
],
"parameters": {
"sampleCount": 1,
"language": "en"
}
};
let response = await axios.post(
"https://us-central1-aiplatform.googleapis.com/v1/projects/" +
projectId +
"/locations/us-central1/publishers/google/models/imagetext:predict",
json,
config
);
return response.data.predictions[0];
}
};
创建名为 auth.js 的文件。 此文件将使用 Google 身份验证客户端库来获取直接调用 Vertex AI 端点所需的访问令牌。
const { GoogleAuth } = require("google-auth-library");
const auth = new GoogleAuth({
scopes: "https://www.googleapis.com/auth/cloud-platform"
});
module.exports = {
getAccessToken: async function () {
return await auth.getAccessToken();
}
};
最后，创建一个名为 storage.js 的文件。此文件将使用 Cloud Storage 客户端库从 Cloud Storage 下载视频。
const { Storage } = require("@google-cloud/storage");
module.exports = {
downloadVideoFile: async function (bucketName, videoFilename) {
// Creates a client
const storage = new Storage();
// keep same name locally
let localFilename = videoFilename;
const options = {
destination: localFilename
};
// Download the file
await storage
.bucket(bucketName)
.file(videoFilename)
.download(options);
console.log(
`gs://${bucketName}/${videoFilename} downloaded locally to ${localFilename}.`
);
return localFilename;
}
};
7. 部署和执行 Cloud Run 作业
首先，确保您位于此 Codelab 的根目录 video-describer-job 中。
cd .. && pwd
然后，您可以使用此命令部署 Cloud Run 作业。
gcloud run jobs deploy $JOB_NAME --source . --region $REGION
现在，您可以通过运行以下命令执行 Cloud Run 作业：
gcloud run jobs execute $JOB_NAME
作业执行完毕后，您可以运行以下命令来获取日志 URI 的链接。（或者，您也可以使用 Cloud 控制台并直接转到 Cloud Run 作业来查看日志。）
gcloud run jobs executions describe <JOB_EXECUTION_ID>
您应该会在日志中看到以下输出：
[{ timestamp: 1, description: 'what is google cloud vision api ? is written on a white background .'},
{ timestamp: 3, description: 'a woman wearing a google cloud vision api shirt sits at a table'},
{ timestamp: 18, description: 'a person holding a cell phone with the words what is cloud vision api on the bottom' }, ...]
8. 恭喜！
恭喜您完成此 Codelab！
建议您查看有关 Video Intelligence API、Cloud Run 和 Vertex AI 视觉字幕的文档。
所学内容
- 如何使用 Dockerfile 创建容器映像以安装第三方二进制文件
- 如何通过为 Cloud Run 作业创建服务账号来调用其他 Google Cloud 服务来遵循最小权限原则
- 如何通过 Cloud Run 作业使用 Video Intelligence 客户端库
- 如何调用 Google API 以获取 Vertex AI 中每个场景的直观描述
9. 清理
为避免产生意外费用（例如，如果此 Cloud Run 作业被意外调用的次数超过免费层级中的每月 Cloud Run 调用次数），您可以删除该 Cloud Run 作业或删除您在第 2 步中创建的项目。
如需删除 Cloud Run 作业，请前往 https://console.cloud.google.com/run/ 前往 Cloud Run Cloud 控制台，然后删除 video-describer-job 函数（如果您使用的是其他名称，则删除 $JOB_NAME）。
如果您选择删除整个项目，可以前往 https://console.cloud.google.com/cloud-resource-manager，选择您在第 2 步中创建的项目，然后选择“删除”。如果删除项目，则需要在 Cloud SDK 中更改项目。您可以通过运行 gcloud projects list 来查看所有可用项目的列表。