A repository of state-of-the-art model implementations and examples built with TensorFlow, demonstrating best practices for machine learning.
TensorFlow Model Garden is a repository containing implementations of state-of-the-art machine learning models and modeling solutions built with TensorFlow. It provides reference examples that demonstrate best practices for TensorFlow users, enabling them to leverage these models for research and product development. The repository includes officially maintained models, research implementations, and community-curated resources.
Machine learning researchers, data scientists, and developers working with TensorFlow who need reference implementations of SOTA models or want to learn best practices for model development and training.
Developers choose TensorFlow Model Garden because it provides officially maintained, optimized implementations of cutting-edge models with training logs for transparency, along with the Orbit library for flexible training loop customization. It serves as a trusted resource for production-ready model code and best practices directly from the TensorFlow team.
Models and examples built with TensorFlow
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
The official directory provides SOTA implementations kept up-to-date with TensorFlow 2 APIs, ensuring reliability and best practices for production use, as stated in the README.
Training logs on TensorBoard.dev are provided for many models, enhancing reproducibility and allowing users to verify performance claims directly.
Orbit offers a lightweight way to write customized training loops with seamless tf.distribute integration, supporting multi-device training like TPUs, as highlighted in the features.
Includes a curated list of external TensorFlow 2 repositories, expanding the available model implementations beyond official offerings.
Cloning the source requires setting PYTHONPATH and installing dependencies manually, which can be error-prone, especially on Windows, as noted in the installation steps.
Research models may use TensorFlow 1 or 2, leading to compatibility issues and added complexity for users relying solely on TensorFlow 2.
Not all models have provided training logs, limiting transparency and reproducibility for some implementations, as admitted in the README.