tartanimu-publicrefit4096-tomjerry

Tom&Jerry: credited V25b all-labeled refit

Reproducible single-model inference for TartanIMU 2026 submission 56343614, Kaggle account tomjerry101, public score 0.42596 (lower is better). This fixed refit improves our preceding 0.42731 entry by about 0.32%. The September 19 snapshot ranks it 50/124; top rank has not been reached and the private result is unknown.

The model derives from DBX / leaonwang's public V25b code and checkpoint release. The upstream architecture and weights are credited to their publisher. Code retains Apache-2.0 and weights, including derivatives, retain CC BY-NC 4.0; see both license files. Challenge data is not included.

Run

Use Python 3.11 and install requirements.txt. The recorded reproduction used Python 3.11.15, PyTorch 2.5.1+cu118, NumPy 2.4.3 and a GTX 1650 Ti. Install the appropriate official PyTorch wheel for your device; cross-runtime bit identity is not guaranteed. The runner supports cpu and cuda:0.

python inference.py --trajectory-root /path/to/test --windows /path/to/index/test_windows.csv --out /new/output/folder --device cuda:0

The output folder must not exist. Outputs are submission.csv, unrounded predictions.npz, and a receipt containing input and artifact hashes. No network connection is needed during inference. Load only the numeric NPZ weight archive; no pickled model object is required.

Inputs and model

One shared TransportState checkpoint contains 4,199,322 parameters in 166 FP32 tensors. It combines a temporal convolution encoder, rotary attention, learned sensor calibration and gyro transport. Four internal velocity offsets and a learned gate are trained within the same model. The gate receives raw IMU features, with no externally supplied platform identity or separately selected models.

The only signal input is 200 Hz six-axis IMU (accelerometer followed by gyroscope, SI units, gravity retained). The runner decodes only the imu array. Window and trajectory identifiers are used for ordering and grouping. Ground-truth orientation, position, velocity and platform metadata are not model inputs. Frozen sensor normalization is in inference_plan.json.

Inference is offline and noncausal: 32-second contexts, 16-second overlap, eight deterministic yaw views of the same weights, inverse rotation and edge-weighted context blending. Trajectory-averaged logits softly mix the internal offsets. Twenty dense estimates form each one-second output. Short sequences repeat the final input frame with a validity mask. This is not a single forward pass and not a causal real-time deployment claim.

Training and evaluation

The publisher reports fitting all 395 original training and 80 validation trajectories, with checkpoint averaging and EMA upstream. Our public3072 parent then received 3,072 train-only updates. This release starts from that accepted parent and fits all 475 labeled trajectories for exactly 4,096 further updates, with fresh AdamW, cosine learning rate 1e-5 to 1e-6, weight decay 0.01, gradient clipping 1, batch size 4 and seed 2026091902. One trajectory per platform and a valid 32-second crop are sampled uniformly from the combined labeled pool. The loss and augmentation are unchanged from public3072: corrected end-position/duration trajectory supervision, yaw and small constant sensor-bias perturbations, and inherited normalization. The refit introduces no new EMA or checkpoint averaging.

There is no independent validation score. All train and validation recordings were used for fitting. Post-fit validation predictions only check the inference pipeline, and were not used for accuracy selection. The 4,096-update endpoint, seed, learning-rate schedule and all inference settings were frozen before public scoring. The independent audit replayed every sampling and augmentation draw, verified all 166 optimizer states at step 4,096 and confirmed all 475 labeled recordings were sampled. Integrity checks do not prove generalization. Test labels and official report-service per-sequence metrics were not used to fit or select this model.

Fresh raw validation reproduced all 23,714 reference predictions. Two independent full test processes produced identical 30,644-window arrays and CSV bytes on the recorded runtime. These are inference reproduction checks, not a full retraining replication. Model/source hashes are verified by the runner.

CSV MD5: 4232ad9803c1567d68afd01fe452816d. The exact accepted prediction file is included as submission.csv for organizer verification, as required by the live competition rules. The inference runner never reads that file. Private ranking remains unknown until the organizers release and verify final results.

Scope and limitations

Research use under the stated noncommercial weight license. Unseen sensors, mounting orientations and motion regimes may degrade accuracy. No independent validation set remains for this checkpoint. Long temporal context, eight views and full-trajectory gating impose latency and memory costs. No uncertainty calibration, physical-safety guarantee or autonomous-navigation readiness is claimed.

Credit the TartanIMU challenge and data authors and the original TartanIMU research, in addition to the V25b publisher. The challenge technical report and final verification are separate from this model artifact.