metadata
language: en
tags:
- weather-forecasting
- diffusion-models
- rectified-flow
- meteorology
- pytorch
- deep-learning
license: mit
datasets:
- meteolibre
MeteoLibre Rectified Flow Model
In the folder models_shortcut/:
Model Description
- Model type: Rectified Flow Diffusion Model
- Architecture: 3D U-Net with FiLM conditioning
- Input: Meteorological data patches (12 channels + 1 lightning channels, 3D spatio-temporal)
- Output: Generated weather forecast data
- Training data: MeteoLibre meteorological dataset
- Language(s): Python
- License: MIT
Training
The model was trained using:
- Framework: PyTorch with Hugging Face Accelerate
- Optimizer: Adam (lr=5e-4) OR SOAP
- Batch size: 64
- Epochs: 200
- Precision: Mixed precision (bf16)
- Distributed training: Multi-GPU support
And there is different video exemple for the inference.