A larger 3D CNN variant trained on temporal dual-polarimetric NEXRAD radar sequences. This is the 64M parameter version -- see resnet3d-18-tornet for the smaller, equally performant 33M version.

Model Description

Same dual-head architecture as ResNet3D-18 but with deeper residual blocks (layers=[3,4,6,3]).

Parameter	Value
Backbone	ResNet3D-34 (BasicBlock, layers=[3,4,6,3])
Parameters	63.7M
Input shape	(B, 24, 8, 128, 128)
Output shape	(B, 4)

Performance

Test Set (3,685 events)

Head	AUC	CSI	F1	Precision	Recall
Detection	0.898	0.540	0.701	0.596	0.851
Prediction	0.988	0.865	0.928	0.924	0.932
Combined	0.943	--	--	--	--

Key Finding

ResNet3D-34 (64M params) performs nearly identically to ResNet3D-18 (33M params) on this task (combined AUC 0.943 vs 0.942). The smaller model is recommended for deployment due to faster inference and lower memory usage.

Training

Parameter	Value
Optimizer	AdamW (lr=1e-3)
Batch size	512
Epochs	20
GPU	NVIDIA H200 (144GB)

Usage

Same as ResNet3D-18 -- see resnet3d-18-tornet for full usage instructions and deployment guidance.

python

from model_resnet3d import DualHeadResNet3D
model = DualHeadResNet3D(in_channels=24, arch="resnet34")
state = torch.load("best.pt", map_location="cpu")
model.load_state_dict(state["model_state_dict"])

Citation

bibtex

@model{resnet3d-34-tornet,
  title={ResNet3D-34 for Temporal Radar Tornado Detection},
  author={DeepGuess},
  year={2026},
  url={https://huggingface.co/deepguess/resnet3d-34-tornet},
}