CCRK: Improving the Consistency in Cross-Lingual Cross-Modal Retrieval with 1-to-K Contrastive Learning
Cross-lingual Cross-modal Retrieval (CCR) is an essential task in web search, which aims to break the barriers between modality and language simultaneously and achieves image-text retrieval in the multi-lingual scenario with a single model. In recent years, excellent progress has been made based on cross-lingual cross-modal pre-training; particularly, the methods based on contrastive learning on large-scale data have significantly improved retrieval tasks. However, these methods directly follow the existing pre-training methods in the cross-lingual or cross-modal domain, leading to two problems of inconsistency in CCR: The methods with cross-lingual style suffer from the intra-modal error propagation, resulting in inconsistent recall performance across languages in the whole dataset. The methods with cross-modal style suffer from the inter-modal optimization direction bias, resulting in inconsistent rank across languages within each instance, which cannot be reflected by Recall@K. To solve these problems, we propose a simple but effective 1-to-K contrastive learning method, which treats each language equally and eliminates error propagation and optimization bias. In addition, we propose a new evaluation metric, Mean Rank Variance (MRV), to reflect the rank inconsistency across languages within each instance. Extensive experiments on four CCR datasets show that our method improves both recall rates and MRV with smaller-scale pre-trained data, achieving the new state-of-art.
Requirements
- Install python3 environment
conda create -n ccrk python=3.8.8
conda activate ccrk
pip3 install -r requirements.txt
Checkpoints
We pretrain the model for only 30 epochs on 2 A100 GPUs. The batch size is set to 128.
Checkpoint | Pretrain Dataset |
---|---|
CCR-10-2M-30epoch | CC2M 10lan |
CCR-10-3M-30epoch | CC2M+COCO+VG+SBU 10lan |
Notes:
2M
:CC2M
3M
:CC2M+SBU+VG+COCO
, SBU/VG/COCO splits are borrowed from CCLM.CC2M
: Because of many broken links, we only collect 1863804 images of Conceptual Captions Dataset.10lan
:zh,en,de,fr,ja,cs,id,es,ru,tr
data
- Download data from corresponding websites
- If running pre-training scripts:
- download pre-trained models for parameter initialization
- image encoder: swin-transformer-base
- text encoder: xlm-roberta-large
- download pre-trained models for parameter initialization
- Organize these files like this:
CCRK/
data/
xlm-roberta-large/...
swin_base_patch4_window7_224_22k.pth
finetune/
mscoco/...
multi30k/...
iglue/
datasets/...
images/
flickr30k-images/*.jpg
coco/
train2014/*.jpg
val2014/*.jpg
test2015/*.jpg
image_data_train/
image_pixels/*.csv
wit_test/
*.csv
pretrain_data/
translated_4M/
cc3m-mm-data-all/
part_*.data
vg-mm-data-all/
part_*.data
coco-mm-data-all/
part_*.data
cc3m-mm-data-all/
part_*.data
ccrk_2m_10lan_epoch_29.th
ccrk_3m_10lan_epoch_29.th
Pretrain
# CCRK 2M 6lan
python3 run.py --task "pretrain" --dist "1" --output_dir "output/CCRK-2m-6lan" --seed 42 --config configs/Pretrain_2m.yaml --pret_para "--language_chosen zh,ja,en,de,fr,cs" --device "c2"
# CCRK 2M 10lan
python3 run.py --task "pretrain" --dist "1" --output_dir "output/CCRK-2m-10lan" --seed 42 --config configs/Pretrain_2m.yaml --device "c2"
# CCRK 3M 6lan
python3 run.py --task "pretrain" --dist "1" --output_dir "output/CCRK-3m-6lan" --seed 42 --config configs/Pretrain_3m.yaml --pret_para "--language_chosen zh,ja,en,de,fr,cs" --device "c2"
# CCRK 3M 10lan
python3 run.py --task "pretrain" --dist "1" --output_dir "output/CCRK-3m-10lan" --seed 42 --config configs/Pretrain_3m.yaml --device "c2"
For distributed training across nodes, see run.py for more details.
Data
To facilitate research on multi-lingual multi-modal pre-training, we provide the text translation of COCO+VG+SBU+CC3M
, which contains 10 language: zh/en/de/fr/ja/cs/id/tr/ru/es
If you want to translate more languages, please refer to translation/README.md
Please cite the corresponding papers appropriately and download the images from their websites.
For more details, please read the code dataset/pretrain_dataset_multilingual.py (more specifically ImageMultiTextDataset) to see what format is needed.
Finetune
We finetune the model for every dataset on 4 V100 GPUs.
Data: MSCOCO and Multi30K
Please download MSCOCO, Multi30K from the corresponding websites. We provide some links for reference.
MSCOCO
- Multi30k
For these two datasets, you need to additionally reformulate the train json files like this:
[
{
"caption": "A woman wearing a net on her head cutting a cake. ",
"image": "coco/val2014/COCO_val2014_000000522418.jpg",
"image_id": 522418
}, ...
]
and the valid and test files like this:
[
{
"image": "coco/val2014/COCO_val2014_000000391895.jpg",
"caption": [
"A man with a red helmet on a small moped on a dirt road. ",
"Man riding a motor bike on a dirt road on the countryside.",
"A man riding on the back of a motorcycle.",
"A dirt path with a young person on a motor bike rests to the foreground of a verdant area with a bridge and a background of cloud-wreathed mountains. ",
"A man in a red shirt and a red hat is on a motorcycle on a hill side."
],
"image_id": 391895
}, ...
]
Data: IGLUE
For IGLUE, you just need to clone this repo and place it in the root path of our repo as follows. Our code works on the original annotations of IGLUE without any preprocess.
CCRK/
iglue/
datasets/...
For WIT, please download the image_data_train.tar
and test images from its kaggle webpage, and extract them to images
, images/wit_test
seperately.
Tips for WIT:
- The download link of
image_data_train.tar
is in Data Description. - You need to extract the files again in
images/image_data_train/image_pixels
andiglue/datasets/wit/annotations/train_en.jsonl.zip
)
Retrieval Tasks: Multi30K and MSCOCO
# English-only Fine-tune
## Multi30K
python3 run.py --dist 1 --task itr_multi30k --config configs/cclm-base-ft/Retrieval_multi30k_en_ft.yaml --output_dir output/multi30k --bs 64 --seed 42 --epoch 10 --checkpoint ../pretrain_data/ccrk_3m_10lan_epoch_29.th --device "c4"
## MSCOCO
python3 run.py --dist 1 --task itr_coco --config configs/cclm-base-ft/Retrieval_coco_en_ft.yaml --output_dir output/mscoco --bs 64 --seed 42 --epoch 10 --checkpoint ../pretrain_data/ccrk_3m_10lan_epoch_29.th --device "c4"
## split train and test to speed up
python3 run.py --dist 1 --task itr_coco --config configs/cclm-base-ft/Retrieval_coco_en_ft.yaml --output_dir output/mscoco --bs 64 --seed 42 --epoch 10 --checkpoint output/mscoco/checkpoint_best.pth --device "c4" --evaluate
# Single-Language Fine-tune
## Multi30K, optional language: cs/de/fr
python3 run.py --dist 1 --task itr_multi30k --config configs/cclm-base-ft/Retrieval_multi30k_cs_ft.yaml --output_dir output/multi30k/cs --bs 64 --seed 42 --epoch 10 --checkpoint output/multi30k/checkpoint_best.pth --device "c4"
## MSCOCO, optional config: ja/zh
python3 run.py --dist 1 --task itr_coco --config configs/cclm-base-ft/Retrieval_coco_ja_ft.yaml --output_dir output/mscoco/zh --bs 64 --seed 42 --epoch 10 --checkpoint output/mscoco/checkpoint_best.pth --device "c4"
IGLUE: Zero-Shot
We provide examples of fine-tuning on English train set and evaluating on the test sets of other languages.
# xFlickr&CO
python3 run.py --dist 1 --task xflickrco --output_dir output/xflickrco --checkpoint ../pretrain_data/ccrk_3m_10lan_epoch_29.th --bs 64 --seed 42 --device "c4"
# WIT
python3 run.py --dist 1 --task wit --output_dir output/wit --bs 80 --seed 42 --checkpoint ../pretrain_data/ccrk_3m_10lan_epoch_29.th --device "c4"
IGLUE: Few-Shot
We also evaluate CCLM on IGLUE max-shot settings. Note that you need to finetune the pretrained model on English first, then load the checkpoints for few-shot learning.
# xFlickr&CO, optional language: de/es/id/ja/ru/tr/zh
python3 run.py --dist 1 --task xflickrco --output_dir output/xflickrco/zh --checkpoint output/xflickrco/checkpoint_best.pth --bs 64 --seed 42 --fewshot de,100 --lr 1e-6 --device "c4"
The value after language in --fewshot
settings of xFlickr&CO is the number of few-shot samples, where we always use the maximum values.
MRV
After all language of models are finetuned, we can evaluate MRV using following command:
python3 run.py --task xflickrco --dist "1" --output_dir output/xflickrco/test --bs 64 --seed 42 --device "c4" --checkpoint output/xflickrco/checkpoint_best.pth --evaluate --fewshot de,100 --ft_para " --model_cap ccrk_3m_10lan --checkpoint_fmt output/xflickrco/format/checkpoint_best.pth"
Specific implementation of MRV can be found in the analysis_ranks
function of xFlickrCO.py
. We calculate mean rank variation of four models en,de,ja,zj
.
Citation
If this work is helpful, please kindly cite as:
@article{nie2024improving,
title={Improving the Consistency in Cross-Lingual Cross-Modal Retrieval with 1-to-K Contrastive Learning},
author={Nie, Zhijie and Zhang, Richong and Feng, Zhangchi and Huang, Hailang and Liu, Xudong},
journal={arXiv preprint arXiv:2406.18254},
year={2024}
}
Acknowledgement
About code, Our project is based on CCLM.
About pretraining datasets, zh,ja,de,fr,cs
texts in cc3m
are translated by UC2 while zh,ja,de,fr,cs
texts in sbu/coco/vg
are translated by CCLM. For other languages id,es,ru,tr
, we use m2m_100_1.2B
model developed by Meta AI and EasyNMT as a tool to translate all datasets from English.
Thanks for their great jobs!