Our paper has been accepted at ECIR'23
HADA is a framework that combines any pretrained SOTA models in image-text retrieval (ITR) to produce a better result. A special feature in HADA is that this framework only introduces a tiny number of additonal trainable parameters. Thus, it does not required multiple GPUs to train HADA or external large-scale dataset (although pretraining may further improve the performance).
In this repo, we used HADA to combine 2 SOTA models including ALBEF and LightningDOT. The total recall was increase by 3.6% on the Flickr30k dataset. Therefore, it needs to clone ALBEF and LightningDOT and extract their feature first.
We uploaded the extracted feature and pretrained model here. Or you can run the extraction again by using files in ALBEF and DOT directory.
We used mlflow-ui to keep track the performance between configurations. Please modify or remove this related-part if you do not want to use.
Remember to update the path in the config files in HADA folders. Then you can train or evaluate by the file run_exp.py
# Train
python run_exp.py -cp HADA_m_extend/Config/C5.yml -rm train
# Test
python run_exp.py -cp HADA_m_extend/Config/C5.yml -rm testWe created a sub-repository for applying HADA using LAVIS as backbones here.
For any issue or comment, you can directly email me at manh.nguyen5@mail.dcu.ie