v0.1.0

v0.1.0

This release establishes mlcast as a usable foundation for machine-learning weather nowcasting experiments. It brings together configurable experiment construction, Zarr-backed meteorological data loading, a PyTorch Lightning training loop, probabilistic ConvGRU nowcasting, and reproducible run configuration. With this release, users can train the built-in ConvGRU baseline from the command line, customize experiments through Fiddle configs and fiddlers, or replace the network with their own nowcasting architecture while keeping the surrounding data, training, logging, and reproducibility machinery.

Added

• Fiddle-based experiment configuration centered on the default training_experiment graph, covering dataset setup, data module, network, Lightning module, optimizer, scheduler, callbacks, logger, and trainer.
• mlcast train CLI for launching the default experiment, applying ordered set: overrides and fiddler: mutations, loading saved YAML configs, switching config functions, and inspecting resolved configs before training.
• Python API for constructing, editing, validating, building, and running experiment configs programmatically.
• Zarr-backed source-data pipeline with CF-standard-name variable lookup, normalization, precomputed CSV sampling, random on-the-fly sampling, train/validation/test splitting, spatial augmentations, optional target masks, and anonymous S3 access.
• Precipitation conversion and normalization utilities for rainfall rate, rainfall flux, and 5-minute rainfall amount.
• Generic PyTorch Lightning nowcasting module that wraps any compatible network and handles training, validation, testing, optimizer/scheduler construction, ensemble forecasting, masked losses, image logging, checkpoint loading, and prediction.
• Built-in ConvGRU encoder-decoder nowcasting model with multi-scale spatial encoding/decoding, latent-space forecast rollout, padding for arbitrary spatial sizes, and optional stochastic ensemble generation.
• Loss functions for deterministic and probabilistic nowcasting, including MSE, MAE, CRPS, almost-fair CRPS, temporal consistency regularization, and masking of invalid target pixels.
• Semantic fiddlers for common multi-parameter configuration changes, including random sampling, variable selection, masking, ratio splits, anonymous S3 datasets, and MLflow logging.
• Reproducibility and experiment tracking support through saved or uploaded Fiddle YAML configs, flattened config hyperparameters, TensorBoard logging, MLflow logging, W&B-compatible config artifact handling, and MLflow system metadata/metrics logging.
• Documentation covering installation, CLI workflows, Python workflows, custom network integration, available fiddlers, project structure, and the ConvGRU architecture.

Changed

• Training now centers on a single Fiddle-based experiment graph rather than the earlier ad hoc configuration flow.
• The CLI now uses mlcast train with nested --config overrides and named fiddlers for more reproducible experiment changes.
• The default training setup now uses the ConvGRU ensemble nowcasting pipeline as the primary supported baseline.
• Saved experiment configs are now persisted alongside runs for exact reproduction.