The file structure is as below.
├── gain-code
   ├──readme.txt
   ├──COIN.json
   ├──preprocess.py
   └── features
       ├── 1
           ├── video_1.npy
           ├── video_2.npy
           └── ...
       ├── 2
       └── ...
   ├──edtcn
       ├── feature_dataset.py
       ├── model.py
       └── train.py

COIN.json: the official annoations of COIN, which can be downloaded at https://github.com/coin-dataset/annotations.

preprocess.py: to preprocess the data and generate the label files and the lists of training and tesing data.

features: this file contains the extracted S3D features. Features from a task are gathered in a file named by the task ID.

The file edtcn contains:
feature_dataset.py: the implementation of the dataset with extracted features and the causal inventation.
model.py: the implementation of ED-TCN
train.py: to train the model with default parameters by using:
python train.py --action train --causal yes --ckpt_dir $dir for checkpoint to save$
          to evaluate the model by using:
python train.py --action predict --ckpt_dir $dir for checkpoint to load$