submission_platform_experts_16models

TartanIMU platform-expert ensemble

This repository contains a self-contained, inference-only PyTorch ensemble for TartanIMU velocity estimation. It combines four base models with four experts for each platform (20 checkpoints total), a saved platform router, and the Test feature cache used for the historical submission. The historical submission has 30,644 rows and a user-reported Public score of 0.26571.

This is a custom PyTorch ensemble, not a Transformers checkpoint. Use the bundled infer.py entry point; the Hugging Face inference widget cannot run it automatically.

Repository layout

The four platform IDs are 0=car, 1=dog, 2=drone, and 3=human. The router predicts a platform from IMU summary features. When its confidence passes the platform threshold, the four matching experts are averaged and blended with the four base predictions using B + 0.5 * (E - B); otherwise the base ensemble is used.

Quick start

Python 3.10+ is recommended. Install the pinned runtime dependencies, then verify every bundled asset before loading a checkpoint:

python -m pip install -r requirements.txt
sha256sum -c SHA256SUMS
python infer.py --verify-only
python infer.py --smoke --device cpu

To reproduce a submission, use a GPU when available. Outputs are never overwritten:

python infer.py --device cuda:0 --output outputs/reproduced.csv

Use --frozen-base to preserve the historical base CSV text while recomputing only the routed experts. Each output gets a sidecar .audit.json containing the device, output hash, and numerical difference from the bundled historical submission.

Upload to Hugging Face Hub

Create an empty model repository, install huggingface_hub, log in with hf auth login, and upload this directory as-is:

hf upload your-user/tartanimu-platform-experts . --repo-type model

The .gitattributes file sends checkpoint and router binaries through Git LFS. No conversion to transformers format is needed because this is a custom PyTorch ensemble.

Evaluation and provenance

The reported Public score is a historical user report, not a score recomputed by this repository. The fixed fold1 result for the separately trained experts was Score 0.3487334649, Macro AVE 0.2695168333, and Macro ATE20 1.0042230308; its scope is holdout_evaluation_with_unlabeled_test_adaptation. Final checkpoints were trained on the official Train+Val set and must not be presented as unseen-data generalization results.

The training process used unlabeled Test IMU distribution matching. It did not read Test truth, Test platform labels, or fit normalization on Test, and it did not perform supervised Test training. These details are retained in provenance/ for auditability.

The package has been verified for strict loading of all 20 checkpoints, a real feature CPU smoke forward, full router replay, byte-identical historical prediction replay, and one complete CPU Test re-inference producing 30,644 rows. The CPU output is not byte-identical to the historical CUDA/BF16 output; its maximum absolute difference was 0.009473085. A full GPU re-inference was not completed on the source machine because its NVIDIA driver was unavailable; outputs can vary with PyTorch, CUDA, BF16, and hardware versions.

License

Repository code and packaging files follow Apache-2.0; see LICENSE. Review the provenance and any upstream dataset or checkpoint terms before redistributing bundled artifacts.

This repository intentionally contains only the code, inference assets, and English upload documentation needed to run the model.