Mr. Right: Multimodal Retrieval on Representation of ImaGe witH Text

Mr. Right is a novel retrieval dataset containing multimodal documents (images and texts) and multi-related queries. It also provides a multimodal framework for evaluation and compares with previous text-to-text retrieval models and image-text retrieval models. Dataset and model checkpoints are released.

For more details, please checkout our Mr. Right paper.

Dataest

• Dataset json files
  • multimodal_documents.json
    • amount: 806,357
    • id: document id
    • title: document title
    • doc_text: document text
    • img_url: document image URL
  • multimodal_pretrain_pairs.json
    • amount: 351,979
    • id: document id
    • title: document title
    • doc_text: document text
    • org_doc_text: wiki original document text (no remove snippets)
    • query_text: text-related query from auto generation
    • query_img: image-related query from auto generation
    • img_url: document image URL
  • multimodal_finetune_pairs.json
    • amount: 900
    • id: document id
    • title: document title
    • doc_text: document text
    • query_text: text-related query from human annotation
    • query_img: image-related query from human annotation
    • query_multi: mixed query from human annotation
    • img_url: document image URL
  • multimodal_val_queries.json
    • amount: 100
    • id: document id
    • title: document title
    • query_text: text-related query from human annotation
    • query_img: image-related query from human annotation
    • query_multi: mixed query from human annotation
  • multimodal_test_queries.json
    • amount: 2,047
    • id: document id
    • title: document title
    • doc_text: document text
    • query_text: text-related query from human annotation
    • query_img: image-related query from human annotation
    • query_multi: mixed query from human annotation

Requirements for evaluation

• python3.8
• pytorch=1.10.1+cu113
• pytorch-lightning=1.5.10
• transformers=4.6.1
• timm=0.4.12
• opencv-python
• einops
• wandb

conda create --name multimodal python=3.8 pandas numpy 
conda activate multimodal
pip install -r requirements.txt
wandb login

Preprocess

1. Download Mr. Right dataset.
2. Extract the Mr_Right.tar.gz to data/ directoy.
3. Download images and create path for each image (Be sure that your storage is more than 1.5TB)
4. Add image path to your json files: {id:0, ......,"doc_image": "xxx.jpg"}, including multimodal_documents.json, multimodal_train_pairs.json, and multimodal_finetune_pairs.json

bash download_dataset.sh

Model Checkpoint

We train our model based on ALBEF, METER, and ViLT.

bash ./checkpoints/download_checkpoints.sh

• Download pretrain models from their repositories if you want to train by yourself. Remember to set the path in config file
  • ALBEF: ALBEF_4M.pth
  • ViLT: vilt_200k_mlm_itm.ckpt
  • METER: meter_clip16_288_roberta_pretrain.ckpt

Edit Configs

• In configs/ALBEF.yaml, configs/METER.yaml, or ViLT.yaml set the paths for the json files and the image path.
• In our task, there are large numbers of documents. To improve the efficiency of validation per training epoch, we suggest that you should split small numbers of doucments with multimodal_val_queries.json. Remember to reset the document id and set the document path in config file.

# dir root: data
python extract_multimodal_val.py --mul_doc multimodal_documents.json \
--mul_val multimodal_val_queries.json \ 
--val_amount 10000 \ 
--output multimodal_val_documents.json

Fine-tune Multimodal model

• embeds_feats: average embedding or compute cls embedding
• pl_checkpoint: resume from checkpoint

CUDA_VISIBLE_DEVICES=0 python main.py \
--num_gpus [number of gpus] \
--num_workers [number of workers] \
--wandb_task_name [Name of task] \
--batch_size 16 \ 
--pretrain [ALBEF | ViLT | METER] \ 
--embeds_feats [avg | cls] \ 
--pl_checkpoint [path for resumed model] \
--save_checkpoint [path for saving checkpoints] \
--neg_matching
--ctx_prediction
--re_ranking

Evaluate

We evaluate our models on a V100 32GB GPU. However, when we calculate the score of TR, IR, and MR simultaneously, the memory size is not enough. Therefore, we store the embeddings checkpoint and calculate the score seperately.

# Run model
CUDA_VISIBLE_DEVICES=0 python main.py \
--num_gpus 1 \
--mode test \
--wandb_task_name [Name of task] \ 
--pickle_output [Directory of testing pickle files] \
--test_output [Json results of model] \
--batch_size 128 \ 
--pretrain [ALBEF | ViLT | METER] \ 
--pl_checkpoint checkpoints/[ albef.ckpt | vilt.ckpt | meter.ckpt] \

# Calculate the score
python compute_pickle.py \
--pickle_input [Embeddings of different retrieval tasks]

Benchmark

License

This data is available under the Creative Commons Attribution Share Alike 4.0 license.

Contact

For any questions please contact [email protected] or [email protected]