This repository contains the implementation of Activity Graph Transformers. The paper is available here or at the project website.
See the following image for an overview of the architecture of Activity Graph Transformer.
- Clone the respository.
https://github.com/Nmegha2601/activitygraph_transformer.git
- Use Conda to install Python3. Create a new environment in conda, and install dependencies from
requirements.txt
.
conda create --name agt_env --file requirements.txt
-
Extract I3D features. We used this to extract the features for each of the dataset in
.npy
format. Place the extracted features indata/thumos/i3d_feats
-
Run the code using the following command. This will start training a model on the dataset. The checkpoints and training log will be saved in an automatically created directory
output
.
bash run_scripts/run_agt.sh
For further questions, please email Megha Nawhal.