This code can be runned only on a pc with a GPU and with Python 3.8.
Open a terminal.
You need first to create a conda environment:
conda create -n test_env python=3.8 && conda activate test_env && conda install pip
Clone this repository
git clone https://github.com/antoniopurificato/apatt_at_semeval.git
Move the current directory to this repository and install the required packages:
cd apatt_at_semeval && pip install -r requirements.txt
Download the dataset. The dataset is private and could be accessed only after a request to Semeval 2023 Task 3 organizers from this site. Once you have the dataset run the following command:
tar -xzvf semeval2023task3bundle-v4.tgz
Then you have to create two folders required for the execution of the code:
mkdir lightning_logs && cd ..
After this you have to save the path of the dataset folder, it will be useful later!
cd semeval2023task3bundle-v4 && pwd && cd ..
Now you can start training your first model! You have first to train every single model and then you can create an ensemble, so you have to use train_single_model.py.
python3 train_single_model.py --epochs [NUMBER OF EPOCHS] --model [NAME OF THE MODEL] --language [LANGUAGE] --threshold [THRESHOLD VALUE EXPLAINED IN THE PAPER] --save [SAVE OR NOT THE MODEL] --path [PATH OF THE DATASET FOLDER (OUTPUT OF pwd IN THE PREVIOUS STEP)]
For other information, as the name of the avalaible models you can use:
python3 train_single_model.py --help
Once you have all the models for a language you can run the ensemble (make sure to have the right paths on the script train_ensemble.py).
python3 train_ensemble.py --epochs [NUMBER OF EPOCHS] --language [LANGUAGE] --threshold [THRESHOLD VALUE EXPLAINED IN THE PAPER] --save [SAVE OR NOT THE MODEL] --path [PATH OF THE DATASET FOLDER (OUTPUT OF pwd IN THE PREVIOUS STEP)]
For other information, as the name of the avalaible models you can use:
python3 train_ensmble.py --help
You can also test the results on the validation dataset!
python3 test.py --language [LANGUAGE] --threshold [THRESHOLD VALUE EXPLAINED IN THE PAPER] --model [NAME OF THE MODEL] --ensemble [IS AN ENSEMBLE OR NOT]