How keras-team/keras-cv works
Project type and purpose
• This is an open-source Python library (KerasCV) providing a modular computer-vision SDK on top of Keras 2/3, with multi-backend support for TensorFlow, JAX, and PyTorch.
• Its purpose is to expose standardized APIs (layers, models, metrics, callbacks, datasets, custom ops) for building, training, and deploying vision pipelines.
File structure
• Top-level directories & files:
• keras_cv/ (package root)
– api/ ← public Python API entrypoints
– src/ ← implementation code (organized by submodules: backend abstraction, layers, models, datasets, metrics, losses, custom ops, utils, training, visualization, tests, tools)
• examples/ ← usage snippets / demos
• benchmarks/ ← performance tests
• build_deps/ ← Bazel-based pip-wheel build config (for custom ops)
• .github/, .kokoro/, WORKSPACE, BUILD.bazel ← CI/CD & build definitions
• setup.py, pyproject.toml, requirements*.txt ← packaging & dependencies
• MVC-like layering:
– API layer (keras_cv/api)
– Core implementation (keras_cv/src)
– Backend adapters (tensorflow, JAX, PyTorch abstraction in src/backend)
– Extensions: custom C++ ops built via Bazel under src/custom_ops
– Utilities & tools for weight conversion, training scripts, and resource loading
• CI pipelines: GitHub workflows + Kokoro configs handle tests, nightly builds, custom-ops wheel builds.
README for additional insights
• Multi-backend support: configured via KERAS_BACKEND
• Highlights major modules: layers, metrics, callbacks, models, datasets
• Provides instructions for custom-op builds via Bazel
• Emphasizes extension points: adding backbones, meta-architectures, new tasks