Skip to content

AINPP Precipitation Benchmark

AINPP-PB-LATAM is a benchmark-oriented Python library for precipitation nowcasting in Latin America. The project combines deep learning model experimentation, reproducible scientific evaluation, and documentation that stays close to the source code.

What This Project Covers

  • standardized loading of precipitation datasets stored as .zarr,
  • Hydra-based experiment configuration,
  • training workflows for direct and autoregressive forecasting,
  • evaluation pipelines for benchmark metrics,
  • visualization utilities for scientific analysis and reporting,
  • support for local and HPC-oriented execution.

Benchmark Assumptions

The reference benchmark follows a fixed operational setup:

  • train split: 2018-2022
  • validation split: 2023
  • test split: 2024
  • input window: 12 hourly steps from gsmap_nrt
  • forecast horizon: 6 hourly steps targeting gsmap_mvk
  • grid shape: 880 x 970

These assumptions are represented in the configuration layer and can be extended for new experiments.

Documentation Map

  • Installation: environment creation and dependency installation with uv
  • Architecture: project structure, configuration model, and scientific pipeline
  • Training: model-by-model training guide, Hydra overrides, dataset tuning, and loss selection
  • Usage: how to run train, evaluate, and infer tasks through Hydra
  • Publishing: how the documentation site is deployed to GitHub Pages
  • API: package reference generated with mkdocstrings

Local Preview

To work on the docs locally:

uv venv
source .venv/bin/activate
uv pip install -e .[docs]
uv run mkdocs serve