This is the code base for paper "Countering Language Drift with Iterated Learning". Please cite the following if you found this codebase to be useful

@article{lu2020countering,
  title={Countering language drift with seeded iterated learning},
  author={Lu, Yuchen and Singhal, Soumye and Strub, Florian and Pietquin, Olivier and Courville, Aaron},
  journal={arXiv preprint arXiv:2003.12694},
  year={2020}
}

This codebase is tested on python3.7 and torch==1.3.1.

Clone this repo, and at project root do pip install -e . as well as other dependencies.

python preprocess/prepare_text.py -data_dir DATA_DIR

Put the images of multi30k under the directory DATA_DIR/multi30k/imgs, then do

python preprocess/extract_flickr30_imgfeats.py -datadir DATA_DIR

This will produce train_feat.pth and val_feat.pth under DATA_DIR/flickr30k.

python run_pretrain.py --config JSON_PATH --data_dir DATA_DIR --exp_dir EXP_ROOT_DIR

python run_finetune.py --config JSON_PATH --data_dir DATA_DIR --exp_dir EXP_ROOT_DIR

Sample json config can be found under the folder jsons.

1. iwslt_en_de.json, iwslt_fr_en.json: Config for pretraining translation agents.
2. hyperparames_caption.json: Config for language model used for scoring generated English.
3. hyperparames_ranker.json: Config for pretraining image retriever.
4. gumbel/: Configs for Gumbel finetuning (vanilla Gumbel, SIL, S2P)