How open-edge-platform/training_extensions works
Project type and purpose
• This is an open-source, low-code transfer learning framework for computer vision built on PyTorch and OpenVINO.
• It exposes both a CLI (otx) and an API (the Engine class) that let users train, infer, optimize, and export models across tasks (classification, detection, segmentation, anomaly).
• Key subdomains:
• Data ingestion & preprocessing (src/otx/core/data, src/otx/data/torch)
• Model architectures & algorithms (src/otx/algo)
• Training engine & scheduling (src/otx/engine, callbacks, strategies)
• Configuration & recipes (src/otx/core/config, src/otx/recipe, src/otx/tools/templates)
• Export & deployment (src/otx/core/exporter)
File structure
• Top-level “src/otx” breakup:
• cli: user entry point for CLI
• engine: orchestration of training, scheduling, adaptive batch size
• core: configuration resolution, data abstractions, metrics, model wrappers, exporter
• data: dataset factories, caching, tiling, transform libraries
• algo: task-specific model definitions, backbones, heads, losses, samplers, explainability
• recipe & tools: YAML-based recipes and code/templates for one-stop model setups
• utils: cross-cutting utilities (logging, device selection, import helpers)
• Supporting folders: docs, tests (unit, integration, perf, regression), Docker configs, CI pipelines
README for insights
• Reinforces the “recipe → engine → algo → export” workflow.
• Auto-configuration and templating are first-class features.
• External dependencies: PyTorch, OpenVINO toolkit, Datumaro data frontend.
Main components
a. User Interfaces
• CLI (otx entrypoint)
• API (Python import of Engine)
b. Engine & Orchestration
• Engine core (src/otx/engine/engine.py)
• Adaptive BS search, callbacks, schedulers
c. Configuration Layer
• Recipe loader (src/otx/recipe)
• Core config resolver (src/otx/core/config)
• Template generator (src/otx/tools/templates)
d. Data Pipeline
• Dataset factory & caching (src/otx/core/data)
• Torch data wrapper & transforms (src/otx/data/torch)
• Tiling & pre-filtering modules
e. Algorithm & Model Layer
• Task-specific folders under algo (classification, detection, segmentation, anomaly, explain)
• Common modules: backbones, heads, losses, layers, samplers
• Integration with PyTorch models & OpenVINO conversion
f. Export & Deployment
• Native exporter (src/otx/core/exporter/native.py)
• Exportable-code demo package
g. External Services & Dependencies
• PyTorch, OpenVINO, Datumaro
• GPU/XPU drivers
• Docker for environment encapsulation
Relationships & interactions
• CLI/API → Engine
• Engine ← Recipe & Config → Template
• Engine → Data Pipeline → Raw dataset sources
• Engine → Model Building (algo modules) → Training loop
• Training loop uses Callbacks & Strategies → progress, metrics
• Upon completion, Engine → Exporter → Deployment artifacts
• Tools (e.g., converter) act as side-utilities on models or recipes
Architectural patterns & design principles
• Modular layering: UI | orchestration | config | data | algorithms | export
• Plugin-style recipe/template system for extensibility
• Separation of concerns: core handles abstractions, algo implements CV logic
• Dependency injection via config/resolver for tasks, devices, hyper-params
• CI/CD pipelines ensure quality: tests, docs, code scanning
Essential elements
• Labels: name each component (e.g., “Recipe Loader,” “Dataset Factory,” “classification Model,” “Exporter”)