Pre-trained checkpoints for Reachability Is Not Generalization

This repository hosts 142 pre-trained model checkpoints covering every experiment reported in the anonymous submission

Reachability Is Not Generalization: Understanding Verb–Noun Decomposition in Assembly Action Recognition.

The bundle contains one best.pth per (dataset × configuration × view × hand) configuration across three assembly datasets (MECCANO, HA-ViD, IMPACT) and five recognition formulations (atomic, shared, independent, C2C-shared, C2C-indep). Together the checkpoints consume ~49 GB, distributed as six ~9 GB tar shards (checkpoints.tar.part00checkpoints.tar.part05). The file MANIFEST.json inside the bundle maps every checkpoint to the paper exhibit (table row / figure) it backs.

Download

pip install huggingface_hub
huggingface-cli download assembly-verbnoun/assembly-ckpt1 \
    --repo-type model --local-dir checkpoints/

# Reassemble + extract in one stream (no 49 GB intermediate file)
cd checkpoints/
cat checkpoints.tar.part* | tar -xf -

# Optional: remove the shards once extraction succeeds
rm checkpoints.tar.part*

After extraction, checkpoints/ contains one directory per experiment configuration tag — e.g. meccano_comp_shared/best.pth, impact_comp_front_left/best.pth, havid_gran_k2_view0_lh_aa/best.pth, ... — mirroring the layout the training and evaluation scripts expect. No path rewriting is needed.

Code and reproduction

The full code release — training, evaluation, table / figure regeneration, dataset preparation, and compositional-split construction — is available at the companion GitHub repository:

🔗 Anonymized code release: https://anonymous.4open.science/r/assembly-F7F2

The repository contains:

  • All training and evaluation scripts (code/)
  • Compositional split construction scripts for all three datasets (data/scripts/)
  • Scripts to regenerate every paper table and figure from the evaluation JSONs
  • Dataset download links and preparation instructions

License

This checkpoint bundle is released under the MIT license (see the GitHub repository for the full text). The three datasets used in the experiments remain under their original licenses and must be obtained from their respective authors.

Citation

Citation block will be added in the camera-ready release. For the review period, please refer to the submission by its anonymous title above.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support