MobileNetv4_TSM

This project provides a lightweight, modular video classification framework that integrates the Temporal Shift Module (TSM) into a customized MobileNetV4 backbone, an efficient architecture adapted from the official PyTorch MobileNetV3 to match the design of MobileNetV4 (arXiv:2404.10518).

• MobileNetV4-style backbone: Adapted from the PyTorch MobileNetV3 implementation and modified to follow the basic structure of MobileNetV4.
• Temporal Shift Module (TSM): Added simple TSM logic to introduce temporal modeling without changing parameter count.
• Lightweight by design: This setup is intended for small-scale video classification experiments, especially for short clips or resource-constrained scenarios.

Environment Setup

We recommend using conda to manage dependencies:

conda create -n MobileNetv4_TSM python=3.11
conda activate MobileNetv4_TSM
pip install -r requirements.txt

Data Format

To train the MobileNetv4_TSM model, your dataset should follow this structure:

dataset_root/
├── video1.mp4
├── video2.mp4
└── ...

And a label file in the format:

video1 123 0
video2 146 1
...

Each line contains: video_name (without .mp4) total_frame_count class_label

Acknowledgements

This project is a side project that builds on the work of two research contributions:

• Temporal Shift Module (TSM):
  GitHub | Paper

• MobileNetV4:
  Paper: "MobileNetV4: Universal CNNs for Mobile Vision"