TorNet-Oracle / README.md
Wonder-Griffin's picture
Update README.md
7af00c9 verified
---
library_name: pytorch
license: mit
datasets:
- TorNet
tags:
- weather
- radar
- tornado
- tornado_prediction
- NEXRAD
- MRMS
- HRRR
- lightning
metrics:
- auprc
- f1
- accuracy
- brier
- ece
pipeline_tag: image-classification
language:
- en
---
# Wonder-Griffin/tornado-super-predictor
**TornadoSuperPredictor** from Storm-Oracle, trained on **TorNet (Zenodo)** patches.
Outputs a tornado probability per patch (optionally with atmospheric features).
## Summary
- **Data**: TorNet (official split); optional recent holdout recommended.
- **Architecture**: CNN feature extractor + heads (probability, EF logits, location, timing, uncertainty).
- **Temporal**: 3 volume(s) stacked as channels.
- **Normalization**: zscore.
- **Loss**: bce (pos_weight=2.0).
- **Calibration**: Platt (A,B)=n/a,n/a; Temperature T=n/a.
## Intended Use
- Research on tornado nowcasting from radar patches;
- Evaluation under class imbalance with PR metrics;
- **Not** an operational warning system without further validation & human oversight.
## Dataset
- **Train examples**: 6
- **Eval examples**: 4
- **Class balance**: positives=n/a, negatives=n/a, pos_weight≈2.0
## Evaluation (threshold = 0.5)
Confusion matrix (rows = truth, cols = prediction):
| | Pred 0 | Pred 1 |
|-------:|-------:|-------:|
| True 0 | 0 | 2 |
| True 1 | 0 | 2 |
Metrics:
- **AUPRC**: n/a
- **Accuracy**: n/a
- **(Optional)**: attach PR curve & reliability diagrams
## Training
- Optimizer: AdamW (lr=1e-4, wd=1e-4 by default)
- Batch size: n/a
- Epochs: n/a
- Precision: 16-mixed
- Augmentations: flips/rotations/intensity jitter + optional crops
- Hardware: 1× GPU (FP16 mixed)
## Quickstart
```python
import torch
from transformers import AutoModel
repo = "Wonder-Griffin/TorNet-Oracle"
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()
# Example dummy batch
B, T, H, W = 2, 1, 256, 256 # T time steps -> in_channels = 3*T (reflectivity, velocity, spectrum width?)
radar_x = torch.randn(B, 3*T, H, W)
# Atmospheric dictionary (use only what you have; shapes must be (B, dim))
atmo = {
"cape": torch.randn(B, 1),
"wind_shear": torch.randn(B, 4), # 0–1, 0–3, 0–6, deep
"helicity": torch.randn(B, 2), # 0–1, 0–3
"temperature": torch.randn(B, 3), # sfc, 850, 500
"dewpoint": torch.randn(B, 2), # sfc, 850
"pressure": torch.randn(B, 1),
}
out = model(radar_x=radar_x, atmo=atmo)
print(out.tornado_probability.shape) # (B,)
print(out.ef_scale_probs.shape) # (B, 6)
print(out.location_offset.shape) # (B, 2)
print(out.timing_predictions.shape) # (B, 3)
---
# 3) Notes to avoid common gotchas
- **Export the class names**: Make sure `StormOracleModel` and `StormOracleConfig` are importable at the repo root via `__init__.py`. Hugging Face uses that when `trust_remote_code=True`.
- **Architectures**: The `"architectures"` array in `config.json` **must** include `"StormOracleModel"`.
- **Weights**: You already have `pytorch_model.bin`/**or** `model.safetensors`. Either is fine. Keep the filenames standard.
- **Forward signature**: With remote code, it’s okay that `forward` takes `radar_x` and `atmo`. Users pass them as keyword args as shown.
- **Version pins**: If you rely on features from newer `transformers`, keep the `transformers_version` in `config.json` current.
---
# 4) Optional niceties
- **`hubconf.py`** (for `torch.hub` users):
```python
from .tornado_predictor import TornadoSuperPredictor
def storm_oracle(in_channels=3, pretrained=False, hf_repo=None, map_location="cpu"):
model = TornadoSuperPredictor(in_channels=in_channels)
if pretrained and hf_repo is not None:
from huggingface_hub import hf_hub_download
path = hf_hub_download(hf_repo, filename="pytorch_model.bin")
import torch
state = torch.load(path, map_location=map_location)
model.load_state_dict(state, strict=True)
return model