Quickstart

Four steps from a clone to a first scored cell.

1. Install and point the harness at your data

git clone https://github.com/SPICExLAB/IMU-Bench && cd IMU-Bench
python -m venv imu-bench && source imu-bench/bin/activate && pip install -e .
export IMUBENCH_DATA_ROOT=/data/imubench      # layout in Datasets
imu-bench tasks

2. Run a reference encoder on one task

imu-bench eval --encoder control_random --tasks har_hhar --protocols probe

control_random is the shared backbone with random weights, never trained: the number it reaches is what the decoder alone achieves. The cell writes runs/bench_har_hhar__control_random__probe__s42/ and bench/results/har_hhar/control_random__probe__s42.json.

3. Run InrtlWrist through the plugin interface

imu-bench eval --encoder examples/plugin/inrtlwrist_plugin.yaml --tasks har_hhar pd_pads --protocols probe

The card points at the released checkpoint on the Hugging Face Hub; the harness downloads it. The scores should match the released result files for these cells, which verifies the data preparation.

4. Run your own encoder

Copy examples/plugin/inrtlwrist_plugin.py and .yaml, replace the model, then:

imu-bench eval --encoder my_encoder.yaml            # all ten tasks, both protocols
imu-bench check my_encoder                           # complete? protocol-faithful? which track?
imu-bench package my_encoder --out submission/       # ready for the leaderboard pull request

Details: Submission modes, Evaluation, Tracks and leaderboard submission.