How microsoft/nni works
High-Level Architecture Components
a. User Interface Layer
• Web UI (ts/webui)
• Command Line Interface (nni/tools/nnictl)
• Python API (nni/*.py)
b. Core Service Layer
• Experiment Management (nni/experiment)
• Training Service Management (nni/training_service)
• Runtime Management (nni/runtime)
c. Algorithm Layer
• HPO (Hyperparameter Optimization)
• NAS (Neural Architecture Search)
• Model Compression
• Feature Engineering
Key Component Relationships
a. Communication Channels
• REST API (between UI and Core)
• Command Channels (between Core and Training)
• WebSocket (for real-time updates)
b. Storage Components
• Database (for experiment data)
• File Storage (for logs and checkpoints)
• Shared Storage Services
Training Service Integration
a. Supported Platforms
• Local
• Remote
• Cloud (AML)
• Kubernetes-based (OpenPAI, Kubeflow, etc.)
b. Platform-specific Components
• Trial Keeper
• GPU Scheduler
• Resource Manager
Algorithm Framework Integration
A section displaying:
a. Framework Support
• PyTorch
• TensorFlow
• Scikit-learn
• Others
b. Algorithm Categories
• Tuners
• Assessors
• Advisors
• Core Components (e.g., blue)
• External Services (e.g., green)
• User Interfaces (e.g., orange)
• Algorithm Components (e.g., purple)
• Storage Components (e.g., yellow)
Important Implementation Details
Architectural patterns:
• Modular Design (shown by clear component boundaries)
• Plugin Architecture (for algorithms)
• REST API Design
• Event-driven Communication
Deployment View
A section showing:
• Container Support (Docker)
• Multi-platform Deployment
• Scaling Capabilities
• Component Types
• Connection Types
• Version Information
The scalability of the system
This structure reflects NNI's sophisticated architecture while making it accessible to understand how the different components interact to provide automated machine learning capabilities.