jakeryderv/tornet-detector-torch

TorNet tornado detector, PyTorch weights

The published TorNet baseline detector (tornet-ml/tornado_detector_baseline_v1, MIT Lincoln Laboratory, MIT license) converted from its Keras 3 checkpoint to a plain PyTorch state dict. No Keras is needed to use it; no weights were changed.

tornet_detector_v1.pt is a torch.saved state dict for tornado_onset.tornet.TorNetDetector (26 tensors, 4,665,409 parameters): a CoordConv CNN encoder (four blocks, 48 to 384 filters) and the max-pool head (1×1 convolutions 512, 256, 1; max over cells). Inputs are channels-first: the 12 normalised radar channels plus 2 range-folded channels of one scan, and the 2 CoordConv coordinate channels, exactly as the Keras model builds them internally.

Conversion: onset-convert-weights in tornado-onset reads model.weights.h5 out of the .keras zip with h5py and maps layer by layer. Verified against the Keras model run in fp32 over the 3,119-sample 2013–2014 TorNet test partition: maximum logit difference 1.9e-5, identical AUC.

from huggingface_hub import hf_hub_download
from tornado_onset.tornet import load_detector

detector = load_detector(hf_hub_download("jakeryderv/tornet-detector-torch", "tornet_detector_v1.pt"))

Cite the TorNet paper when you use it: Veillette et al., TorNet: A benchmark dataset for tornado detection and prediction using full-resolution polarimetric weather radar data, AIES, 2024, https://doi.org/10.1175/AIES-D-24-0006.1.