设计师专属 SD
使用 Vertex AI 触发、运行和管理微调、训练和部署自定义稳定扩散模型的完全自动化工作流程
描述
Sd-aa-S 是一个全自动 MLOps 流水线，用于在 GCP 上触发、管理和跟踪稳定扩散微调作业，使用 Google Cloud Storage、Cloud Build、Cloud PubSub、Firestore、Cloud Run、Cloud Functions 和 Vertex AI 等 GCP 组件。它旨在简化使用不同技术（从 Dreambooth 开始）调整稳定扩散的机器学习工作流程。即将支持 Lora、ControlNet 等平台。该项目面向机器学习/数据工程师、数据科学家以及任何对构建大规模稳定扩散微调平台感兴趣或正在努力构建该平台的人士。
三部分
1. 应用程序部分
1. Set up your Cloud Environment
2. Create a backend service for handling uploads to a GCS bucket
- Receive images from clients and store them under a predefined GCS bucket path
- Track the status of individual uploads in a Firestore collection
- Track the status of the overall upload job in a separate Firestore collection
- Once the job is compelted, publish the jobID as the message on a predefined PubSub topic
3. Deploy this backend service as a Cloud Run endpoint using Cloud build
4. Create a frontend portal to upload images using ReactJs
5. Deploy the frontend service on Cloud Run
2. Vertex AI 部分
1. Set up your Cloud Environment
2. Create a new custom container artifact for running the pipeline components
3. Create a new custom container artifact for running the training job itself
4. Create a Jupyter notebook outlining the Pipeline flow & components
5. Compile a YAML file from a Vertex AI workbench and store the precompiled YAML file under a GCS bucket path
3. 管道部分
1. Set up your Cloud Environment
2. Create a cloud function that gets triggered every time the jobID is published on a predefinied topic (from 1st part)
3. Within the cloud function, the python code subscribes to the topic and triggers a Vertex AI pipeline job using the precomiled YAML file (from 2nd part)
4. The pipeline jobs finetunes the stable diffusion model using Dreambooth, uploads the new custom model to Model registy & deploys an endpoint
5. The job also updates Firestore with the status of the pipeline job from start to end