Downloads · 30 days
0
mdhamidhosen/tartanimu-unified-hosen42
tartanimu-unified-hosen42 is a machine learning model from mdhamidhosen. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as apache-2.0.
One model, one shared set of weights, for car / dog / drone / human. From raw 6-axis IMU (200 Hz, gravity retained) it predicts the mean 3-D body-frame velocity of every 1 s window. Built for the IROS 2026 TartanIMU C…
Downloads · 30 days
0
Access
Public
Updated Oct 8, 2026
Repo size
23.2 MB
Likes
0
Public
Click a slice to open those files.
.pt23.2 MB · 100%
From the Hugging Face model README
One model, one shared set of weights, for car / dog / drone / human. From raw 6-axis IMU (200 Hz, gravity retained) it predicts the mean 3-D body-frame velocity of every 1 s window. Built for the IROS 2026 TartanIMU Challenge.
| file | what |
|---|---|
model.pt | ranked entry — Kaggle submission 56402554 (submission_final_kaggle_w_uni.csv, private 0.16141 — 7th of 132 on the final leaderboard; public 0.28609). 3.9 M params. SHA-256 in SHA256SUMS. |
model_secondary_final_v4_160.pt | second selected Kaggle entry 56332266 (private 0.16625, public 0.28696): same pipeline, width-128 / 2-layer variant, length-weighted sampling. 1.8 M params. |
infer.py | inference entry point (test trajectory directory + window index CSV → submission CSV) |
model.py | architecture + sliding-chunk trajectory inference |
requirements.txt | pinned environment |
pip install -r requirements.txt
python infer.py --traj_dir /path/to/test --windows /path/to/index/test_windows.csv --out submission.csv
# second entry:
python infer.py --checkpoint model_secondary_final_v4_160.pt --traj_dir ... --windows ... --out submission_v4_160.csv
imu (N,6) = [ax, ay, az, gx, gy, gz] from each .npz and the window order from the CSV. No platform
label, pose, timestamps or metadata. Trajectories are processed independently (never concatenated).--device cpu, tens of minutes).
No internet access needed. TF32 is disabled for determinism; GPU-to-GPU differences are ≤ 1e-3 m/s.fd12d4f25d7b52d44aaa4e4a14534378
(produced with these weights on a Kaggle P100/T4; re-running on another device reproduces it to ≤ 0.005 m/s per window).Raw 200 Hz IMU chunks of 16 s → strided conv stem (200 → 20 Hz tokens) + deterministic strap-down features (gyro-integrated relative orientation, de-gravitated acceleration, integrated velocity change, under both gyro-z sign hypotheses) → 8 residual depthwise-dilated temporal-convolution blocks (≈ 13 s receptive field) → 3-layer transformer encoder over the chunk → per-token velocity head (dense 20 Hz), averaged to one vector per 1 s window. At inference, chunks slide over the whole trajectory with a stride of 2 windows and overlapping predictions are Hann-weight averaged.
Training (challenge train + val only, no external data, no pretrained weights): AdamW + OneCycle, 160 epochs × 250 steps of 64 chunks, EMA (0.998) + SWA over the last 20 epochs; loss = vector-Huber window velocity + 0.5 dense 20 Hz Huber + 0.2 integrated-error (drift) loss + 0.05 auxiliary platform cross-entropy (training signal only, never used for routing); sampling uniform over platforms and uniform over trajectories within a platform; augmentation = random sensor-mount rotations (≤ 15°, ≤ 45° for drone chunks) applied to IMU and target, physically exact time dilation (× 0.77–1.3), IMU bias / scale / white noise.
Team: Hack2Publish (Md. Hamid Hosen, Md Esfer Abdus Sami, Md. Foysal Emon Shanto, Kahakashan Ashraf). Data: TartanIMU Challenge, CMU AirLab.