CNN + TCN network that predicts a mouse's breathing trace from video.
This repository is a uv workspace of two packages:
| Package | Import | Command | What it is |
|---|---|---|---|
zephyr |
zephyr |
zephyr |
The network: clip lists, preprocessing, training, evaluation, inference. The only released package. |
zephyr-benchmarks |
zephyr.benchmarks |
zephyr-benchmarks |
Benchmarks on our dataset and the methods zephyr is compared against (TS-CAN, PhysNet, pixel, Facemap-style). Workspace only. |
zephyr never imports zephyr.benchmarks; a test enforces it.
git clone https://github.com/AllenNeuralDynamics/zephyr
cd zephyr
uv sync --locked --all-packages --extra cpu # or --extra gpu for CUDAThe dataset (face-camera videos, frame timestamps and thermistor traces) is on
a public S3 bucket; no AWS credentials are needed. The download commands are in
the benchmarks README, under Dataset.
Everything goes into data/ at the repository root.
examples/has one clip list, fold and experiment to start from; the zephyr README walks through them.packages/zephyr-benchmarks/configs/holds the benchmark's configs; its README covers downloading the data and reproducing the published results.data/(inputs),benchmarks/(reference runs) andruns/(outputs) live at the repository root and are ignored by Git.