UniIR

🌐 Homepage | 🤗 Dataset(M-BEIR Benchmark) | 🤗 Checkpoints(UniIR models) | 📖 arXiv | GitHub

This repo contains the codebase for the ECCV-2024 paper "UniIR: Training and Benchmarking Universal Multimodal Information Retrievers"

🔔News

• 🔥[2024-04-13]: We highlight another valuable and concurrent research on training instruction-following, multi-task multi-modal retrievers with Late-interaction:PreFLMR: Scaling Up Fine-Grained Late-Interaction Multi-modal Retrievers , which was done by the researchers of the University of Cambridge. They also introduced the M2KR benchmark which can be used to train and evaluate multi-modal universal information retrievers. We may combine the M2KR and M-BEIR benchmarks together to facilitate the advance of this field.
• 🔥[2024-03-18]: Release the UniIR(CLIP_SF) large and UniIR(BLIP_FF) large checkpoints 🤗 Checkpoints
• 🔥[2023-12-21]: Our 🤗 M-BEIR Benchmark is now available for use.

Introduction

We propose the UniIR(Universal multimodal Information Retrieval) framework to learn a single retriever to accomplish (possibly) any retrieval task. Unlike traditional IR systems, UniIR needs to follow the instructions to take a heterogeneous query to retrieve from a heterogeneous candidate pool with millions of candidates in diverse modalities.

Content

1. M-BEIR
2. Training
3. Evaluation
4. Model Zoo
5. Citations and Contact

M-BEIR

To train and evaluate universal multimodal retrieval models, we build a large-scale retrieval benchmark named M-BEIR (Multimodal BEnchmark for Instructed Retrieval).

M-BEIR Downloading

We provide the M-BEIR dataset in the 🤗 Dataset. Please follow the instructions provided on the HF page to download the dataset and prepare the data for training and evaluation. You need to set up GiT LFS and directly clone the repo:

git clone https://huggingface.co/datasets/TIGER-Lab/M-BEIR

UniIR Models

We provide the codebase for training and evaluating the UniIR CLIP-ScoreFusion, CLIP-FeatureFusion, BLIP-ScoreFusion, and BLIP-FeatureFusion models.

Environment

Prepare the codebase of the UniIR project and Conda environment using the following commands:

git clone https://github.com/TIGER-AI-Lab/UniIR
cd UniIR

cd src/models/
conda env create -f uniir_env.yml

Training

To train the UniIR models from pretrained CLIP and BLIP checkpoints, please follow the instructions below. The scripts will automatically download the pretrained CLIP and BLIP checkpoints.

1. Download the M-BEIR Benchmark

Please download the M-BEIR benchmark by following the instructions in the M-BEIR section.

2. Scripts

To train UniIR CLIP_SF Large with the default configuration:

cd src/models/uniir_clip/clip_scorefusion/configs_scripts/large/train/inbatch/

Modify inbatch.yaml for hyperparameter tuning and run_inbatch.sh for your own environment and paths.

Note:

1. Modify the UNIIR_DIR in the run_inbatch.sh to the directory where you want to store the checkpoints.
2. Modify the MBEIR_DATA_DIR in the run_inbatch.sh to the directory where you store the M-BEIR benchmark.
3. Modify the SRC_DIR in the run_inbatch.sh to the directory where you store the codebase of the UniIR project(This repo).
4. By default, UniIR models are trained on M-BEIR with in-batch negatives, and the hard negatives provided by the original datasets are not used.
5. We used wandb to log the training process. Please make sure a .env environment with WANDB_API_KEY, WANDB_PROJECT, and WANDB_ENTITY is set.

Then you can run the following command to train the UniIR CLIP_SF Large model.

bash run_inbatch.sh

To train UniIR BLIP_FF Large with the default configuration:

cd src/models/uniir_blip/blip_featurefusion/configs_scripts/large/train/inbatch/

Modify inbatch.yaml for hyperparameter tuning and run_inbatch.sh for your own environment and paths.

bash run_inbatch.sh

Similarly, you can train the UniIR CLIP_FF and BLIP_SF models by modifying the corresponding scripts.

Evaluation

We provide the evaluation pipeline for the UniIR models on the M-BEIR benchmark.

1. Environment

Please create an environment for the FAISS library:

# From the root directory of the project
cd src/common/
conda env create -f faiss_env.yml

2. Download the M-BEIR Benchmark

Please download the M-BEIR benchmark by following the instructions in the M-BEIR section.

3. Download the UniIR Checkpoints

You can train the UniIR models from scratch or download the pre-trained UniIR checkpoints by following the instructions in the Model Zoo section.

4. Scripts

To evaluate UniIR CLIP_SF Large with the default configuration:

cd src/models/uniir_clip/clip_scorefusion/configs_scripts/large/eval/inbatch/

Modify embed.yaml, index.yaml, retrieval.yaml and run_eval_pipeline_inbatch.sh for your own environment, paths and evaluation settings.

Note:

1. If you download our pretrained UniIR model, please modify the UNIIR_DIR in the run_eval_pipeline_inbatch.sh to the directory where you want to store large files including the checkpoints, embeddings, index and retrieval results. Then you can place the clip_sf_large.pth file in the following path:

   $UNIIR_DIR/checkpoint/CLIP_SF/Large/Instruct/InBatch/clip_sf_large.pth

   This the default path specified by model.ckpt_config in the embed.yaml file.
2. Modify the MBEIR_DATA_DIR in the run_eval_pipeline_inbatch.sh to the directory where you store the M-BEIR benchmark.
3. Modify the SRC_DIR in the run_eval_pipeline_inbatch.sh to the directory where you store the codebase of the UniIR project(This repo).

The default configuration will evaluate the UniIR CLIP_SF Large model on both the M-BEIR (5.6M heterogeneous candidate pool) and the M-BEIR_local (homogeneous candidate pool) benchmarks. UNION in the yaml files refers to the M-BEIR (5.6M heterogeneous candidate pool). You can follow the comments in the yaml files and modify the configurations to evaluate the model on the M-BEIR_local benchmark only.

bash run_eval_pipeline_inbatch.sh

embed, index, logger and retrieval_results will be saved in the $UNIIR_DIR directory.

To evaluate UniIR BLIP_FF Large with the default configuration:

cd src/models/unii_blip/blip_featurefusion/configs_scripts/large/eval/inbatch/

Similarly, if you download our pretrained UniIR model, you can place the blip_ff_large.pth file in the following path:

$UNIIR_DIR/checkpoint/BLIP_FF/Large/Instruct/InBatch/blip_ff_large.pth

The default configuration will evaluate the UniIR BLIP_FF Large model on both the M-BEIR and the M-BEIR_local benchmarks.

bash run_eval_pipeline_inbatch.sh

UniRAG evaluation

UniRAG evaluation is very similar to the default evaluation with the following differences:

• It stores jsonl files containing queries and their retrieved candidates under retrieval_results. This is useful when retrieved results will be used in downstream applications like RAG.
• When retrieve_image_text_pairs in retrieval.yaml is set to True, a complement candidate will be fetched for each candidate with text or image only modality. With this setting, the candidate and its complement will always have image, text modality. Complement candidates are fetched by using the original candidates as queries (e.g., query_text -> candidate_image -> complement candidate_text).
• To run evaluations in UniRAG mode follow the instructions provided above replacing InBatch and inbatch with UniRAG and unirag, respectively.

You can train and evaluate the UniIR CLIP_FF and BLIP_SF models by modifying the corresponding scripts.

Model Zoo

We provide the UniIR model checkpoints in the 🤗 Checkpoints. You can directly use the checkpoints for retrieval tasks or fine-tune the models for your own retrieval tasks.

Available Checkpoints

Model Name	Version	Model Size	Model Link
UniIR(CLIP-SF)	Large	5.13 GB	Download Link
UniIR(BLIP-FF)	Large	7.49 GB	Download Link

You can download them by

git clone https://huggingface.co/TIGER-Lab/UniIR

Citation and Contact

• Cong Wei: c58wei@uwaterloo.ca
• Yang Chen: yangc@gatech.edu
• Alan Ritter: alan.ritter@cc.gatech.edu
• Wenhu Chen: wenhuchen@uwaterloo.ca

BibTeX:

@article{wei2023uniir,
  title={Uniir: Training and benchmarking universal multimodal information retrievers},
  author={Wei, Cong and Chen, Yang and Chen, Haonan and Hu, Hexiang and Zhang, Ge and Fu, Jie and Ritter, Alan and Chen, Wenhu},
  journal={arXiv preprint arXiv:2311.17136},
  year={2023}
}