How tuandung222/convert-phowhisper-asr-from-encdec-to-ctc works
Project type and purpose
• This is a full-stack, containerized ML application for Vietnamese Automatic Speech Recognition.
• The system comprises:
• A FastAPI backend serving a CTC-based PhoWhisper model for inference
• Interactive UIs (Streamlit and Gradio) for end users
• Observability stack (Prometheus, Grafana, Jaeger)
• Deployment automation via Docker Compose, Kubernetes (DigitalOcean), and Terraform
• CI/CD pipelines using GitHub Actions and Jenkins
File structure
• Top-level directories:
• api/ – FastAPI server, model loading, metrics and tracing clients
• ui/ – Streamlit and legacy UI code, static assets
• src/ – Core inference & training code (models, trainer, utilities, Gradio demo)
• configs/ – YAML configurations for training, model parameters, inference
• docker/ – Docker Compose overlays for API, UI, monitoring, full stack
• k8s/ – Kubernetes manifests and setup scripts (base, ingress, monitoring)
• terraform/ – IaC for DigitalOcean Kubernetes cluster and resources
• monitoring/ – Stand-alone Prometheus, Grafana, Alertmanager configs
• jenkins/ – Jenkinsfile, Dockerfile for Jenkins master, deployment scripts
• tests/ – Integration and unit tests for API, UI, model scripts
• notebooks/ – Jupyter notebooks for experimentation and refactoring
README insights
• User → UI → API → Model inference pipeline
• API pushes metrics to Prometheus and traces to Jaeger
• Grafana visualizes Prometheus metrics
• Deployment options: Docker Compose vs. Kubernetes (provisioned by Terraform)
• CI/CD flow: GitHub → Jenkins → Build & Deploy to Docker Compose or K8s
Detail
• Don’t omit the training code under src/training or notebook experiments; show them as optional developer workflows
• CI/CD pipeline steps: code → build → push → deploy
• Infrastructure provisioning: Terraform → DOKS → kubeconfig → kubectl apply
• How API metrics & traces flow into the monitoring stack