| # Core (needed for the data/scoring/Indonesia stack -- climatology.py, |
| # indonesia_zones.py, backtest_indonesia.py, era5_data_pipeline.py, |
| # zone_observation.py, crop_risk_scorer.py, hierarchical_search.py) |
| numpy>=1.24 |
|
|
| # RL training stack (train_curriculum.py, train_kaggle.py, |
| # weather_forecast_env.py, gru_weather_policy.py) |
| torch>=2.0 |
| gymnasium>=0.29 |
| stable-baselines3>=2.0 |
| sb3-contrib>=2.0 |
|
|
| # Hyperparameter sweeps (sweep_reward_shaping.py) -- this file was temporarily |
| # created to assess optimal hyperparameters. Train_kaggle.py's best values |
| # dict (learning_rate=6.916624987609979e-05, ent_coef=0.08779238696445962, |
| # etc.) was "found by the Optuna sweep (trial 6 of the 20-trial run against |
| # n_zones=3 / max_steps=250)" and hardcoded as the CLI defaults -- |
| # consistent with this having been a one-time iteration tool whose winning |
| # trial's output was captured inline, rather than a script meant to persist |
| # in the repo. optuna is essential if you wish to reexplore optimal parameters. |
| optuna>=3.5 |
|
|
| # Edge export (mnn_export.py) -- MNN itself has no pip package; build/install |
| # per https://github.com/alibaba/MNN, this only covers the ONNX/export side. |
| onnx>=1.15 |
|
|
| # Real-data fetching (era5_data_pipeline.py, climatology.py) -- all optional; |
| # each degrades to synthetic/cached data gracefully without it, but any real |
| # (non-synthetic) fetch needs at least `requests`. The comment at the top of |
| # this file listing era5_data_pipeline.py under "Core -- numpy>=1.24" is |
| # incomplete: numpy alone is enough for the module to import, not for its |
| # real-data code paths to work. |
| requests>=2.31 |
| cdsapi>=0.6 # ERA5 reanalysis tier only |
| earthengine-api # `import ee` -- IMERG/CHIRPS/SMAP satellite tier only |
| netCDF4>=1.6 # ERA5 NetCDF reads -- tried first |
| xarray>=2023.1 # ERA5 NetCDF reads -- fallback if netCDF4 unavailable |
|
|
| # LocalTimesFMBackend only (timesfm_wrapper.py) -- a deliberately opt-in |
| # forecast tier gated behind a manually downloaded, SHA256-verified |
| # checkpoint (see LocalTimesFMBackend.__post_init__); most users won't hit |
| # this path. Note: timesfm_wrapper.py imports pandas without a try/except |
| # guard (unlike its `import timesfm` a few lines above, which does have |
| # one) -- if pandas is missing, this fails with an unhelpful raw |
| # ImportError rather than the graceful message the rest of this codebase |
| # uses for optional deps. |
| timesfm |
| pandas>=2.0 |
|
|
| # TensorBoard training logs (train_kaggle.py) -- optional; training runs |
| # fine without it, just without tfevents output. Listed here despite being |
| # wrapped in a try/except in code because train_kaggle.py's own quickstart |
| # docstring tells users to install it, and every verified training run in |
| # this project's history had it installed. |
| tensorboard>=2.14 |
|
|
| # Optional -- only needed if you actually connect to a broker |
| # (node_transport.py's MQTTTransport falls back to LocalTransport without it) |
| paho-mqtt>=1.6 |
|
|
| # Testing |
| pytest>=7.0 |
|
|
