Downloads · 30 days
9
31% of all-time downloads
suyashdiamond11/tartanimu-ln-metric1024-tomjerry
tartanimu-ln-metric1024-tomjerry is a machine learning model from suyashdiamond11. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
The original architecture and initialization are by arattan7272. This package contains one continued checkpoint, used identically for every platform, and a standalone runner that regenerates predictions from raw IMU.
Downloads · 30 days
9
31% of all-time downloads
All-time downloads
29
Public
Repo size
21.5 MB
Likes
0
Public
Click a slice to open those files.
.pt21.5 MB · 93%
From the Hugging Face model README
The original architecture and initialization are by arattan7272. This package contains one continued checkpoint, used identically for every platform, and a standalone runner that regenerates predictions from raw IMU.
Parent: arattan7272/tartanimu-lnresnet, revision 4b5f23afb14da3f3e70dac1a675fbf7f24597aed, Apache-2.0. Parent checkpoint SHA-256: b42e7aedae87a067c93c56400ddb33434d9566029f3e1f4435ec94a839b3b949. Original copyright, license and attribution are retained. We do not claim the original network as our invention.
Tom&Jerry adds 1,024 low-rate updates using all 395 training and 80 validation recordings. Both matched experimental arms use velocity supervision and consistency with the frozen parent; the candidate also uses an aligned 20-metre trajectory loss. The frozen parent is used only during training. The submitted checkpoint is the final EMA, decay 0.995, of the candidate run. There is no intermediate checkpoint choice or model mixture at inference.
The original admission gate failed. Macro velocity error was 1.0339% worse than the matched control, exceeding a predeclared 1% cap. We recorded a post-hoc exception for one exploratory Kaggle evaluation because the actual composite objective, all four platform composite scores, and macro trajectory error improved against both parent and control. The original failed result remains preserved. This is not a prospective gate pass. The parent had already fitted every released labeled trajectory, so these development measurements are not independent validation or a private-score estimate.
exploratory_decision.json and validation_summary.json disclose the decision and audit. All 4,096 raw training crops and labels were reconstructed and hash-checked. The training plan, history and source are audit artifacts; that source depends on the campaign cache and is not advertised as a standalone from-scratch trainer. Inference is standalone.
The network predicts 64 target windows with seven context windows on each side. Overlap starts have stride 16, retain the original native starts and include the final full context. An eight-window edge taper weights predictions; accumulation uses FP64. Short trajectories retain one native context. The network uses FP32 with TF32 disabled. No rotation augmentation, external platform dispatch or multiple-checkpoint inference is used.
Use Python 3.11 and the pinned requirements with an appropriate official PyTorch wheel:
python reproduce.py --traj-dir /path/to/test \
--windows /path/to/index/test_windows.csv \
--out-dir result --device cuda
The output directory must be new. The runner verifies source and checkpoint hashes, loads with weights_only=True, and reads only raw IMU and window ordering. It requires no internet. submission.csv is the exact scored artifact and is never read to generate predictions. Competition data and hidden labels are not redistributed.
reproduction_summary.json records exact full-validation reproduction, a fixed CPU/GPU comparison, and two independent test runs with identical arrays and CSV bytes. These checks establish implementation reproducibility, not hidden-test accuracy. Official reporting-service results are not used for model tuning or selection.