Exploring Frequency-Domain Feature Modeling for HRTF Magnitude Upsampling
Paper • 2602.11670 • Published
FD-Conformer
Pretrained checkpoints and preprocessed SONICOM data for Exploring Frequency-Domain Feature Modeling for HRTF Magnitude Upsampling.
Files
| File | Description |
|---|---|
sonicom.npz |
Preprocessed SONICOM data: H_log, float32, shape (200, 793, 2, 106), in dB |
checkpoints.zip |
Conformer checkpoints for 3, 5, 19, and 100 measurement directions |
configs.zip |
Matching evaluation configurations and artifact information |
SHA256SUMS.txt |
File checksums |
Usage
Download the files from Files and versions. In the GitHub repository, place sonicom.npz under data/ and extract both ZIP files into the repository root.
For example, evaluate the 5-point checkpoint:
python src/eval.py --config configs/sonicom_m5.yaml --ckpt checkpoints/m5/best_model.pt
Use the corresponding m3, m19, or m100 configuration and checkpoint for other measurement counts. Follow the GitHub README for environment setup and detailed instructions.
Artifacts match source commit d753ec1.
License and attribution
- Checkpoints and configurations: MIT.
sonicom.npz: CC BY-SA 3.0, following the source SOFA license notices.
The processed data are derived from the SONICOM HRTF Dataset. Please cite Engel et al. (2023), “The SONICOM HRTF Dataset,” JAES, 71(5), 241–253, and the FD-Conformer paper when using these resources.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support