This repository represents the implementation of the paper:
Oneata, Dan, Leanne Nortje, Yevgen Matusevych, and Herman Kamper. The mutual exclusivity bias of bilingual visually grounded speech models. Interspeech, 2025.
The implementation relies on PyTorch, which can be installed via conda:
conda create -n me-vgs python=3.12
conda activate me-vgs
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidiaThen we can install the code as a library:
pip install -e .To train a visually grounded speech model, we can run the mevgs/train.py script followed by the configuration name;
for example:
python mevgs/train.py en-nl_links-no_size-md_a The list of configurations is in the mevgs/config.py file.
To obtain predictions for the mutual exclusivity tests, we can run the mevgs/predict.py script,
again followed by the configuration name;
for example:
python mevgs/predict.py en-nl_links-no_size-md_a The results are then obtained with the mevgs/scripts/evaluate.py script:
python mevgs/scripts/evaluate.py en-nl_links-no_size-md_a The results in the Table 1 from the paper are obtained with the following command:
python mevgs/scripts/show_interspeech25_table_1.py