A toolkit for WeatherBench based on PyTorch

pip install wxbtool

For detailed installation instructions, see the Installation Guide.

wxb data-serve -m wxbtool.specs.res5_625.t850weyn -s Setting3d

wxb train -m wxbtool.zoo.res5_625.unet.t850d3sm_weyn

wxb test -m wxbtool.zoo.res5_625.unet.t850d3sm_weyn

wxb forecast -m wxbtool.zoo.res5_625.unet.t850d3sm_weyn -t 2023-01-01 -o output.png

Note: For deterministic forecast, -t must be in YYYY-MM-DD (date only).

wxb forecast -m wxbtool.zoo.res5_625.unet.t850d3sm_weyn -t 2023-01-01T00:00:00 -G true -s 10 -o output.nc

Note: For GAN forecast, -t must be in YYYY-MM-DDTHH:MM:SS (date and time).

wxb data-serve -m wxbtool.specs.res5_625.t850weyn -s Setting3d -b 0.0.0.0:8088

Note: Use --bind to specify the address. The --port option is currently not used by the implementation.

wxb train -m wxbtool.zoo.res5_625.unet.t850d3sm_weyn -d unix:/tmp/test.sock

wxb backtest -m wxbtool.zoo.res5_625.unet.t850d3sm_weyn -t 2023-01-01 -o output.nc

This will write outputs under output/2023-01-01/ and, when using .nc, also create var_day_rmse.json containing day-by-day RMSE.

wxb data-download -m wxbtool.zoo.res5_625.unet.t850d3sm_weyn --coverage weekly

For more detailed examples and explanations, see the Quick Start Guide.

wxbtool supports flexible dataset layouts beyond the default yearly files. You can configure how data files are discovered by setting two fields in your Setting:

• granularity: one of yearly, quarterly, monthly, weekly, daily, hourly
• data_path_format: a Python format string relative to the variable directory, supporting placeholders {var}, {resolution}, {year}, {month}, {day}, {hour}, {week}, {quarter}

Example for monthly files:

from wxbtool.nn.setting import Setting

class MySetting(Setting):
    def __init__(self):
        super().__init__()
        self.granularity = "monthly"
        self.data_path_format = "{year}/{var}_{year}-{month:02d}_{resolution}.nc"

See details and more examples in the Data Handling Guide.

• Installation Guide
• Quick Start Guide
• Data Handling Guide
• Training Guide
• Evaluation Guide
• Inference Guide
• Troubleshooting Guide

• Architecture Overview
• Model Specifications
• Creating Custom Models

See the comprehensive documentation in the docs directory.

uv build
uv publish

git tag va.b.c master
git push origin va.b.c

• Mingli Yuan (Mountain)
• Ren Lu