This is the repository of our paper https://arxiv.org/abs/2009.10434.
For now, only the testing model and the corresponding code is available. We will release the whole project once the paper is accepted.
The pretrained models are provided for Charades-STA which shall be stored in 'Home_path/checkpoints/charades_sta_train' and TaCoS which shall be stored in 'Home_path/checkpoints/tacos_train'.
The extracted I3D features for TaCoS and for Charades-STA is provided for both of them, which should be stored in 'Home_path/proposal_free/preprocessing/tacos' and 'Home_path/preprocessing/charades-sta', respectively.
The above models and features are stored in Baiduyun disk, where the extraction key is th08 for all of them.
The pre-trained glove embedding that we use is glove.840B.300d.zip trained with the Common Crawl (840B tokens, 2.2M vocab, cased, 300d vectors, 2.03 GB download), which shall be unzipped and put in 'Home_path/data/TMLGA'.
The code is based on https://github.com/crodriguezo/TMLGA. And thanks to their features.
run the program with 'python main.py --config-file experiments/tacos_train.yaml'