nowcastnet-rewritten 0.1.0

A personal reimplementation of NowcastNet inference framework

NowcastNet-Rewritten

1. Introduction

This project is a personal reimplementation of the NowcastNet inference framework. The original research, titled "Skilful nowcasting of extreme precipitation with NowcastNet," by Yuchen Zhang, Mingsheng Long et al., was published in Nature and can be accessed at https://www.nature.com/articles/s41586-023-06184-4. Additionally, the original code by Yuchen Zhang is available at https://doi.org/10.24433/CO.0832447.v1.

Q: Why reimplement? A: Just for learning :)

2. Getting Started

1. Cloning the repository:

   git clone https://github.com/VioletsOleander/nowcastnet-rewritten.git
   

2. Install the dependencies (pick one of the following methods):

   • Sync dependencies using uv:

     uv sync
     

   • Make sure python>=3.10,<3.11, and either install from PyPI:

     pip install -U nowcastnet-rewritten
     

     or install from local:

     pip install .
     

Notes:

• uv is recommended for managing dependencies for full reproducibility.
• You may need to implement your own code to read the dataset. Sample code for reading the radar dataset is provided in the datasets directory.
• To ensure compatibility with this reimplementation's architecture, weights have been modified and are available for download from Hugging Face.

3. Usage

To perform inference, run inference.py with the required arguments.

To view all available arguments, use:

python inference.py -h

It is recommended to use the --config_path option to specify the configuration file, for example:

python inference.py --config_path ../configs/inference.toml

An example configuration file is provided in the configs/ directory.

You can also specify options directly from the command line, for example:

python inference.py \
    --case_type normal \
    --device cuda:0 \
    path_to_weights \
    path_to_data \
    path_to_result

Note: If --config_path is specified, other command line options will be ignored.

4. Example Inference Result

1024 x 1024:

512 x 512: