How openvinotoolkit/training_extensions works
Project Type and Purpose
• This is a low-code computer vision transfer learning framework built with deep learning technologies.
• Its goal is to enable users (even with low expertise in deep learning) to train, infer, optimize, and deploy various computer vision models (classification, detection, segmentation, etc.) via both an API and a CLI.
• The project builds on PyTorch and OpenVINO™ and provides “recipes” (model templates) and auto-configuration to streamline training.
File Structure
• Notice major directories such as “src” (with subdirectories for algorithms, core, cli, data, engine, recipe, etc.). Each subdirectory corresponds to key components:
– The “otx” package in src is the heart of the system. It includes submodules that handle algorithms (for classification, detection, etc.), training recipes, data processing, CLI tools, and engine management.
– “docs” and “for_developers” provide extensive documentation, guides, and design references.
– “tests” are organized thoroughly into unit, integration, regression, and performance tests for all components.
• There are also CI/CD configuration folders (e.g., .ci and .github) and Docker-related files (in both .ci/docker and docker directories), which indicate the project’s continuous integration/deployment strategy.
• Build and packaging files (pyproject.toml, setup-like files, MANIFEST.in, etc.) confirm that the project is structured for distribution and developer contributions.
Elements
– "CLI/API Interface"
– "Training Engine & Auto-Configuration"
– "Data Ingestion & Preprocessing"
– "Model Export & Inference"
– "Core Utilities & Configuration Management"
– "CI/CD Pipelines & Deployment (Docker, GitHub Actions)"
– From CLI/API to the Training Engine
– From Data modules to the Training Engine
– From the Engine to different Algorithm Modules
– Then to Export functionalities and outwards towards deployment
• Optionally, include annotations explaining key processes (like auto-configuration, distributed training, and mixed-precision training).