dayuml / mssnet

KSSNet: Multi-Label Classification with Label Graph Superimposing

KSSNet: Multi-Label Classification with Label Graph Superimposing [arXiv]

Overview

The PyTorch implementation of the KSSNet.

Prerequisites

The code is built with following libraries:

• Python 3.5 or higher
• PyTorch 0.4.1 or higher
• torchvision 0.2.0 or higher
• PIL
• torchnet
• tqdm

Data Preparation

We have trained on Charades dataset with this code. Most of data preprocessing have been done, while the remained precedure is the data preparation:

• Download tar file from Charades and extract files into "data/Charades_v1/Charades_v1_rgb". The preprocessed annotations have already been contained and in this repo. If necessary, The preprocessing code will be add in this repo.

Code

This code is based on the ML-GCN codebase. Thanks Megvii-Nanjing for their work.

Testing

The scripts will test the checkpoint provided in this repo by running:

python test_i3d_charades.py

The superparameters are set in the 'main' function of test_i3d_charades.py. Sorry for the rough code!

Training

To train I3D with this repo: sh python train_i3d_charades.py

The superparameters are set in the 'main' function of test_i3d_charades.py, too.

Reference

This project is based on ML-GCN