Skip to content
/ IAIS Public

[ACL 2021] Learning Relation Alignment for Calibrated Cross-modal Retrieval

License

Notifications You must be signed in to change notification settings

lancopku/IAIS

Repository files navigation

IAIS: Inter-modal Alignment for Intra-modal Self-attentions

PWC PWC PWC

This repository contains the code for our paper Learning Relation Alignment for Calibrated Cross-modal Retrieval (ACL-IJCNLP 2021 main conference).

Overview of IAIS

Some code in this repo are copied/modified from UNITER, and other opensource implementations made available by PyTorch, HuggingFace, OpenNMT, and Nvidia. The image features are extracted using BUTD.

Update

[2023.02] Please refer to Github Release for our fine-tuned checkpoints and logs for MS COCO and Flickr30k.

[2022.12] According to the researchers from CMU, our IAIS algorithm achieves a new SOTA on the Winoground dataset with a 10% improvement on VinVL (Oscar+) and a 52% improvement on UNITER. Thanks for the interesting work. Their paper: link.

Overview

  1. We propose a Relation Consistency Hypothesis: Given a matched image-text pair, the linguistic relation should agree with the visual relation.
  2. We design a novel metric: Intra-modal Self-attention Distance with annotation (ISDa) to measure the consistency between textual and visual relations.
  3. We propose a new regularized training method called Inter-modal Alignment on Intra-modal Self-attentions (IAIS) to calibrate two intra-modal attention distributions mutually via inter-modal alignment, which helps learn better contextualized representations for image-text pairs.

Requirements

We provide Docker image for easier reproduction. Please install the following:

Our scripts require the user to have the docker group membership so that docker commands can be run without sudo. We only support Linux with NVIDIA GPUs. We test on Ubuntu 18.04 and V100 cards. We use mixed-precision training hence GPUs with Tensor Cores are recommended.

Getting Started

  1. Download processed data and pretrained models with the following command.

    bash scripts/download_itm.sh $PATH_TO_STORAGE

    After downloading you should see the following folder structure:

    ├── img_db
    │   ├── coco_train2014
    │   ├── coco_train2014.tar
    │   ├── coco_val2014
    │   ├── coco_val2014.tar
    │   ├── flickr30k
    │   └── flickr30k.tar
    ├── pretrained
    │   ├── uniter-base.pt
    │   ├── uniter-large.pt
    └── txt_db
        ├── itm_coco_train.db
        ├── itm_coco_train.db.tar
        ├── itm_coco_val.db
        ├── itm_coco_val.db.tar
        ├── itm_coco_restval.db
        ├── itm_coco_restval.db.tar
        ├── itm_coco_test.db
        ├── itm_coco_test.db.tar
        ├── itm_flickr30k_train.db
        ├── itm_flickr30k_train.db.tar
        ├── itm_flickr30k_val.db
        ├── itm_flickr30k_val.db.tar
        ├── itm_flickr30k_test.db
        └── itm_flickr30k_test.db.tar
    
  2. Launch the Docker container for running the experiments.

    # docker image should be automatically pulled
    source launch_container.sh $PATH_TO_STORAGE/txt_db $PATH_TO_STORAGE/img_db \
        $PATH_TO_STORAGE/finetune $PATH_TO_STORAGE/pretrained

    The launch script respects $CUDA_VISIBLE_DEVICES environment variable. Note that the source code is mounted into the container under /src instead of built into the image so that user modification will be reflected without re-building the image. (Data folders are mounted into the container separately for flexibility on folder structures.)

  3. Run finetuning for the ITM task.

    All experiments in the paper are conducted on 8 NVIDIA V100 GPUs.

    • Image-Text Retrieval (Flickr30k)
      • finetune with hard negatives
        horovodrun -np 8 python train_itm_hard_negatives.py \
            --config config/train-itm-flickr-base-8gpu-hn.jgon
        
      • finetune with hard negatives + IAIS
        horovodrun -np 8 python train_itm_hard_negatives.py \
            --config config/train-itm-flickr-base-8gpu-hn.jgon --IAIS singular
        
    • Image-Text Retrieval (COCO)
      • finetune with hard negatives
        horovodrun -np 8 python train_itm_hard_negatives.py \
            --config config/train-itm-coco-base-8gpu-hn.json
        
      • finetune with hard negatives + IAIS
        horovodrun -np 8 python train_itm_hard_negatives.py \
            --config config/train-itm-coco-base-8gpu-hn.json --IAIS singular
        

    The argument --IAIS indicates incorporating the auxiliary IAIS loss to the fine-tuning phase. We support

    • singular: Singular Alignment that establishes a one-to-one mapping between linguistic and visual attention weight (Section 3.1 in the paper). Overview of IAIS
    • distributed: Distributed Alignment that establishes a distributed mapping (Section 3.2 in the paper). Overview of IAIS

    The main code for the IAIS method is in the UniterEncoder class in model/model.py.

Contact

If you have any questions related to the code or the paper, feel free to email Shuhuai (renshuhuai007 [AT] gmail [DOT] com).

Citation

If you find this code useful for your research, please consider citing:

@inproceedings{ren2021iais,
    title = "Learning Relation Alignment for Calibrated Cross-modal Retrieval",
    author = "Ren, Shuhuai and Lin, Junyang and Zhao, Guangxiang and Men, Rui and Yang, An and Zhou, Jingren and Sun, Xu and Yang, Hongxia",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    year = "2021",
}

License

MIT