SmoothI

The model is implemented using PyTorch. Python environment requirements can be found at 'requirements.txt' file.

The src/ contains training or evaluation scripts.

The commands to prepare the pickle files, to train the model with different losses and finally evaluate the trained model are as follows:

To start, change current directory to src/

cd src/

Prepare data (train/vali/test) pickle files

python prepare_data_pickles.py PATH_TO_DATA_FOLDER

e.g. ('python prepare_data_pickles.py ../data/')

Train model by maximizing P@10

python train.py --model_dir ../experiments/ --fold 1 -c smoothi_pk -k 10

Train model by maximizing NDCG@1

python train.py --model_dir ../experiments/ --fold 1 -c smoothi_ndcg -k 1

Train model by maximizing NDCG

python train.py --model_dir ../experiments/ --fold 1 -c smoothi_ndcg -k 0

Evaluate model

python evaluate.py --model_dir ../experiments/ --fold 1

Citation

If you use this work, please cite:

@article{arxiv2021-smoothI,
  author = {Thonet, Thibaut and Cinar, Yagmur Gizem and Gaussier, Eric and Li, Minghan and Renders, Jean-Michel},
  title = {SmoothI: Smooth Rank Indicators for Differentiable IR Metrics},
  year = {2021},
  journal = {arXiv},
  volume = {abs/2105.00942}
}