DrGil/grw-smoothing-movinet

GRW Smoothing (MoViNet) — Model Weights

This repository contains MoViNet checkpoint weights trained with GRW-smoothing from the paper:

Paper

• https://huggingface.co/papers/2511.20928
• https://arxiv.org/abs/2511.20928

Code (reproduction instructions)

For step-by-step instructions to reproduce results (environment setup, data prep, evaluation commands, exact settings), please refer to:

• https://github.com/cmusatyalab/grw-smoothing

Files (checkpoints)

This repo provides the following checkpoint files:

• a0s_grw.pt
• a1s_grw.pt
• a2s_grw.pt
• a3b_grw.pt

Associated dataset

A frozen evaluation dataset used for reproducible testing is provided here:

• https://huggingface.co/datasets/DrGil/k600_test_ds

Intended use

These weights are intended for research and benchmarking in video classification, especially when studying efficiency/accuracy trade-offs and temporal smoothness regularization.

Limitations

• Performance and behavior depend on the training/evaluation setup described in the paper and code.
• Models may not generalize well outside the evaluation distribution.
• If used on real-world or sensitive video content, apply appropriate privacy and governance practices.

Citation

If you use this model, please cite:

@inproceedings{goldman2025grwsmoothing,
  title     = {Smooth Regularization for Efficient Video Recognition},
  author    = {Gil Goldman and Raja Giryes and Mahadev Satyanarayanan},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2025},
  url       = {https://arxiv.org/abs/2511.20928}
}