Downloads · 30 days
0
transformer-ship-traj-pred
mark000071/transformer-ship-traj-pred
transformer-ship-traj-pred is a machine learning model from mark000071. 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 mit.
These are the checkpoints for the code at https://github.com/mark000071/transfomer-ship-traj-pred. That repository holds the model code, the evaluation scripts and the full README, including the known discrepancies. T…
Access
Public
Updated Sep 24, 2026
Repo size
5.3 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt5.2 GB · 99%
At a glance
- License
- mit
- Access
- Public
- Created
- Sep 24, 2026
- Updated
- Sep 24, 2026
- SHA
- d4b618d0
Datasets
Model card
From the Hugging Face model README
EnvShip-DR weights: SDF-map Transformer for 10-minute vessel trajectory prediction
These are the checkpoints for the code at https://github.com/mark000071/transfomer-ship-traj-pred. That repository holds the model code, the evaluation scripts and the full README, including the known discrepancies. The data is mark000071/envship_v2_datasets (Track A).
# from the GitHub repo root
python scripts/download_weights.py # 12 headline checkpoints
python scripts/download_weights.py --subset all # everything below
| Folder | Content | Test ADE (m) |
|---|---|---|
envship_dr/dma/v4_env_spatial_e90_s{42,1,2}/ | headline, DMA | 84.38 / 85.05 / 85.65 (recorded)¹ |
envship_dr/piraeus/v4_env_spatial_e90_s{42,1,2}/ | headline, Piraeus | 155.94 / 157.31 / 151.19 |
envship_dr/norway/v4_env_spatial_e90_s{42,1,2}/ | headline, Norway | 123.57 / 119.28 / 112.20 |
envship_dr/noaa/v4_env_spatial_e90_s{42,1,2}/ | headline, NOAA | 105.38 / 110.51 / 108.33 |
envship_dr/{region}/v4_env_spatial_e30_s42/ | 30-epoch schedule | DMA 94.34 · Piraeus 218.52 · Norway 174.07 · NOAA 167.07 |
ablation_dma/v4_{variant}_e30_s42/ | DMA ablation, 30 ep | vanilla 97.99 · env_pool 94.43 · env_spatial 90.41 · social 99.35 · social_env_pool 97.67 · social_env_spatial 91.36 |
lstm_env_pool/{region}_s{seed}/ | LSTM + env-pool baseline² | Piraeus 146.43 / 151.86 / 149.99 · Norway 123.64 / 126.76 / 137.61 · NOAA 105.65 / 107.20 / 103.43 |
per_sample_errors/ | errors.npz (d: N × 30 per-step errors), for paired tests | — |
¹ On the public (post-refill) DMA rasters these checkpoints score 84.31 / 84.97 / 85.64. See the GitHub README, §B. ² This baseline reads the binary land and water masks (env channels 0–1), not the SDF. See the GitHub README, §A.
Each folder contains best.pt and the exact training config.yaml. A checkpoint is a dict
{"model": state_dict, "config", "epoch", "val_ADE_greedy30", ...}. Integrity:
sha256sum -c SHA256SUMS.
Research use only; not for autonomous collision avoidance. See docs/MODEL_CARD.md on GitHub.
- License
- mit
- Created
- Sep 24, 2026
- Updated
- Sep 24, 2026