README.md

PyTorch implementation of Action-Centric Relation Transformer for VideoQA.

Steps to run the experiments

Requirements

Here lists the required packages for our model:

• Python 2.7
• PyTorch=1.0.1
• tqdm
• nltk<=3.4
• h5py
• pandas
• typing

You can use the following command to install them: pip install h5py nltk==3.3 pandas torch==1.0.1 tqdm typing

Datasets and word embeddings

• features: Please download([Baidu Yun](链接: https://pan.baidu.com/s/1zz2dfwsnr4G_QhD8WlBSFQ) with code kv8q.
• first put all the things with features.tar.gz into data directory.
• next combine all feature.tar.gz, extract them into one directory features, put it under the directory data

Training on TGIF-QA

Model is trained separately on 4 sub-tasks: Action,Trans,Frame,Count, first get into corresponding directory (taking Action as an example):

• cd TGIF_QA/Action

then begin training

• python main.py

Training on Anet-QA

• cd Anet-QA
• python main.py