背景
以前这类云相关的certificate,公司都是全额承担报销的. 现在经济下行周期,消减福利,只报销一般.
去年年末,公司说有AI云服务相关培训, 跟manager申请,承诺不会耽误工作,
自己顶住压力, 仍然坚持报名了training,
年末趁自己尚有余力,开始备考.
2025-11-08 12:30:54
- 考察AddPhrasse 相关API调用
- AddPhraseListAsync
- PhraseistCreateObject
- 考察迁移container的部署能力
- model必须是trained 或者是published的, thus 1.1
- export model using export container (gzip)
- run a container and mount model file
- 选型:
- QnA Maker
- Language Understanding
- Dispatch
- 选型:
- Form Recognizer
- Computer Vision
- Personalizer
- Custom Vision
- 选型和API
- 请求新建资源,用PUT,更新用PATCH
- CognitiveServices是包含OCR,Speech,WebSearch,Decision,Language的
- 几个ai service的原则
- transparency
- faireness
- inclusiveness
- reliability and safety
- privacy and security
- 对docker 分发的理解
- Create
- Build
- Push
- Disctribute and
run
-
对服务和billing end point的理解
-
computeVision vs. computeVision.prediction
-
分辨subscription key vs. query key (只在query 服务中用到)
-
form recogniser的训练
- jpg,png,pdf,tiff
- <= 50mb
- Azure Cognitive Search with server side encryption / CMK
- Index Size Increase
- Query times Increase
- azure key vault is required.
-
AI principle: transparency
-
组网配置. private endpoint with private link
-
组网配置. virtual network + ip firewal rule + private link
-
azure private link 避免virtual network,不论private/public
-
服务选型: anomaly detector
-
考察SoT: audioStreamFormat.getCompressedFormat/SpeechRecogniser
-
技术选型: Face/Speech/Text Analytics
-
QnA会引入两个服务: Azure Storage + Congnitive Search
-
IAM - authoring resources
-
throttle通常解决办法是replica
-
Speech to Text with AutoDetectSourceLanguageConfig/Text to Speech
-
StartRecognizeReceiptsFromUri 参数
-
a new Cognitive Services resource that uses the S0 pricing tier - save cost
-
adding indexes, doesn’t resolve throttle due to increasing query volume
-
adding CMK, doesn’t resolve throttle due to increasing query volume
-
deploy service1 and a private endpoint to vnet1.
-
For inclusiveness: Direct Line Speech is a robust, end-to-end solution for creating a flexible, extensible voice assistant. It is powered by the Bot Framework and its Direct Line
-
api endpoint 构成
-
Purchase Orders (POs) -> Form Recognizer -> OCR -> JSON (extracted info from POs) -> Shaper skill -> JSON -> Table Projection -> JSON -> Power BI
-
Replica used For resolving throttle
-
Face的能力是有facial features and attributes的,比如Face mask, glasses等等.
-
向量搜索建表
- Retrievable: Include the category field in the search results.
- Searchable: Ensure that users can search for words in the category field.
- Facetable: Ensure that users can perform drill down filtering based on category.
- Azure Metrics Advisor
- data mornitoring and nomaly detection in time series data
-
API: describeImageInStreamAsync
-
先调整domain,然后才是retrain 和export
-
涉及finicial situation,的用人介入. (principle of fairness and accountability)
-
Virtual network settings for ta1 而不是NSG
-
Metrics Advisor -> Multivariate Anomaly Detction,多维度异常检测
-
Docker省钱: 尤其是考虑到limited connectivity以及minimise costs
-
create private endpoint + key authentication
-
change points vs. batch detection. batch detection 做dataSet vs. anmomalise
-
建立一个on premise的service的步骤: k8s (on premise) connected to internet; image from MCR; appID, secret + URL
-
Azure 管理后台的API是通过subscription来创建服务的. subscription
-
docker 创建instance的时候,三要素eula,billing,apikey
docker run --rm -it -p 5000:5000 --memory 4g --cpus 1 \
mcr.microsoft.com/azure-cognitive-services/decision/anomaly-detector:latest \
Eula=accept \
Billing={ENDPOINT_URI} \
ApiKey={API_KEY}-
To authenticate with a Microsoft Entra token, the Speech resource must have a custom subdomain and use a private endpoint. The Speech service uses custom subdomains with private endpoints only.
-
6rpm -> 1000 tpm (1 capacity)
-
签名,本地敏感资源通过provisioning在本地直接跑.
- Provision the Language service resource in Azure.
- Deploy a Docker container to an on-premises server.
- Identify the Language service endpoint URL and query the prediction endpoint.
-
API key. AI resources 均衡负载时,不用 bear token (这个是authentication时候用的)
-
模拟环境bot framework sdk -> bot framework emulator
-
deployment name, endpoint, key
-
ipsec rules (两个ip之间已经建立的信道),gateway也可以,但是花实践,virtual network rules(直接限制access可行)
-
Anomaly detector
-
构建command (docker)
-
metric alerts in Azure Monitor, for data audioStreamFormat
-
大规模视频分析,用onedrive download link vs. share link 不如直接下载,重新上传.
-
enable service endpoint/modify virtual network
-
麦克风,人像, 直接speech to text 即可