1. Install dependencies:

pip install -r requirements.txt

2. Donwload and extract data from Kaggle

3. Preprocess the data:

python preprocess_data.py <path-to-train.tsv> <path-to-dataset-folder>

4. Run experiment with standard cross-entropy loss:

python run_experiments.py <path-to-dataset-folder>

Add --gpu to train on GPU

5. Run experiment with complement cross-entropy loss:

python run_experiments.py <path-to-dataset-folder> --use_cce

6. Accuracy scores will appearn on screen. More details will be available in Tensorboard:

tensorboard --logdir=./runs