This is the TensorFlow Model Garden, a large monorepo of model implementations, training loops, and research code. The dominant character is an application/framework: it provides official TF2 model code under `official/`, a custom training library under `orbit/`, and a large body of research code under `research/` and `community/`. The codebase is not a single deployable service but a collection of training/eval/export scripts, model factories, config systems, and data pipelines. It uses Gin configs, dataclass-based configs, and a registry pattern to compose models, tasks, and data loaders. Tests exist across the repo, and CI workflows are present. The repository is hybrid in that it contains research-oriented experiments and data-processing code, but the core is a software framework for building and training TensorFlow models.
NeuralGCM is a research-grade hybrid machine learning and physics-based atmospheric model for weather and climate simulation, implemented primarily in JAX and Haiku. The repository provides a well-structured library (`neuralgcm`) with a clean inference API (`PressureLevelModel`) that supports loading pretrained checkpoints, encoding ERA5-like data into the model's sigma-coordinate state, advancing the model in time, and decoding back to pressure-level output with xarray integration. The code is highly modular and configurable via the gin-config system, allowing flexible composition of encoders, decoders, advanced neural parameterization modules, stochastic random fields, and data transforms. The repository also includes reference training code (designated as pseudocode), paper configuration files, Jupyter notebooks demonstrating data preparation and model deep dives, and pre-trained deterministic and stochastic checkpoints at multiple resolutions. While the inference path is fully reproducible given the provided checkpoints and public ERA5 data, the training pipeline relies on internal Google infrastructure and proprietary code, limiting full reproducibility. The library demonstrates strong software engineering practices for a research codebase, with extensive unit tests, modular design, and clear versioning, though it exhibits some technical drift due to an ongoing migration of components to a separate repository (`terrax`) and the use of hardcoded normalization statistics tied to specific training datasets.