Downloads · 30 days
0
xunden/tartanimu-xunden-iros2026-rich150
tartanimu-xunden-iros2026-rich150 is a robotics model from xunden. Use it for the robotics 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 cc-by-nc-sa-4.0.
This repository is the second of the two single-model checkpoints that team Xunden selected on the Kaggle leaderboard. The first one lives at xunden/tartanimu-xunden-iros2026 (RICH2200tv, 31-channel inputs, 200 epochs…
Downloads · 30 days
0
Access
Public
Updated Sep 24, 2026
Repo size
361 MB
Likes
0
Public
Click a slice to open those files.
.pt361 MB · 99%
From the Hugging Face model README
This repository is the second of the two single-model checkpoints that team Xunden
selected on the Kaggle leaderboard. The first one lives at
xunden/tartanimu-xunden-iros2026
(RICH2_200_tv, 31-channel inputs, 200 epochs). The two checkpoints are independent
single models: each generated its own selected CSV on its own, and no prediction in
either CSV is an ensemble, blend or average of the two.
| Quantity | Value |
|---|---|
| Kaggle public leaderboard (selected entry #2) | 0.33684 |
| Submitted CSV md5 | af459049ce7a07416ef2c161be8e9814 |
The official per-sequence scoring service was not run on this particular CSV; its recorded result is the Kaggle public leaderboard entry above. The richer companion model in the sister repository reports the official per-sequence service numbers.
| Architecture | resnet_bilstm_attn_mh — dilated ResNet window encoder, BiLSTM, multi-head self-attention, shared velocity head + softmax-routed platform-conditioned expert mixture (same family as the sister checkpoint) |
| Parameters | 90.34 M (single checkpoint, best.pt) |
| Input features | rich, 17 channels per 0.5 s token: raw IMU (6) + first differences (6) + channel norms (2) + attitude-grounded linear acceleration (3) |
| Difference vs. sister checkpoint | feature mode only: this run uses the 17-channel rich set and a 150-epoch schedule instead of the 31-channel rich2 set and 200 epochs; everything else (architecture, augmentations, losses, inference code) is identical |
| Window / context | 1 s windows split into 2 tokens; 20-window (20 s) sequences |
| Training data | train + val (475 labelled trajectories), 150 epochs, cosine schedule |
| Checkpoint epoch | 142 (selected by the training-time proxy metric) |
The checkpoint is self-contained: it stores per-channel mean/std for input
standardization plus all architecture hyper-parameters (imu_feature_mode, 17
input_channels, sub_window, seq_len, model kind / scale / encoder kind), so
inference needs no extra flags or training-time artefacts.
python -m tartan_imu.predict_tcn \
--data-root /path/to/data/root \
--checkpoint best.pt \
--output submission.csv \
--split test --stride 20 --num-workers 8
--stride 20 with the default single view (no test-time view averaging) is the exact
inference configuration of this submission. Determinism: model.eval() +
torch.no_grad(), no dropout and no stochastic input augmentation.
Re-executing this checkpoint with the code in code/ reproduces the submitted CSV
bit-exactly — identical md5 af459049ce7a07416ef2c161be8e9814, verified as part of
this release.
--data-root … --model-kind resnet_bilstm_attn_mh --model-scale 4.0 --encoder-kind dilated --sub-window 2 --seq-len 20 --stride 5 --imu-feature-mode rich --lr 1e-4 --lr-schedule cosine --warmup-epochs 4 --min-lr-ratio 0.02 --vibration-aug 3.2 --padding-aug 0.5 --train-splits train val --eval-split val --epochs 150 --batch-size 128 --selection-metric proxy_score
| Path | Contents |
|---|---|
best.pt | The submitted single checkpoint (weights + standardizer + hyper-parameters), epoch 142. |
reproduce.sh | Regenerates and format-validates the submission CSV. |
code/tartan_imu/ | Minimal inference code (predict_tcn, sequence_data, data, tcn_model) — the same code that regenerated the submission bit-exactly. |
code/requirements.txt, code/run_inference.sh | Dependencies and a one-line runner. |
--stride 20, view offset 0). For trajectories whose window
count is not a multiple of the stride, tail windows are covered by two overlapping
sequences of the same model whose predictions are averaged (1.8 % of test windows,
overlap factor ≤ 2); this is a property of the sequence builder, not an ensemble.submission.csv is the exact CSV behind our selected Kaggle entry (original
filename submission_RICH150_tv.csv, MD5 af459049ce7a07416ef2c161be8e9814); reproduce.sh regenerates a byte-identical
file from best.pt. requirements.txt pins the exact environment used.