This repo contains the code and data for VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks. In this paper, we aimed at building a unified multimodal embedding model for any tasks.
- [2024-10] VLM2Vec has been integrated into vLLM.
- [2024-10] The technical report, code, data, and model for VLM2Vec are all available online.
Our model is based on converting an existing well-trained VLM (Phi-3.5-V) into an embedding model. The basic idea is to take the last token in the end of the sequence as the representation of the multimodal inputs.
Our model is being trained on MMEB-train and evaluated on MMEB-eval with contrastive learning. We only use in-batch negatives for training. Our best results were based on Lora training with batch size of 1024. We also have checkpoint with full training with batch size of 2048. Our results on 36 evaluation datasets are:
- Train data: https://huggingface.co/datasets/TIGER-Lab/MMEB-train
- Eval data: https://huggingface.co/datasets/TIGER-Lab/MMEB-eval
Our model can outperform the existing baselines by a huge margin.
Please refer to demo as the minimum example.
Download the image file zip from huggingface
git lfs install
git clone https://huggingface.co/datasets/TIGER-Lab/MMEB-train
cd MMEB-train
python unzip_file.py
cd ../
For GPUs with small memory, use GradCache to reduce memory usage, i.e. setting small values to --gc_q_chunk_size
and --gc_p_chunk_size
.
Use --lora --lora_r 16
to enable LoRA tuning.
torchrun --nproc_per_node=2 --master_port=22447 --max_restarts=0 train.py \
--model_name microsoft/Phi-3.5-vision-instruct --bf16 --pooling last \
--dataset_name TIGER-Lab/MMEB-train \
--subset_name ImageNet_1K N24News HatefulMemes InfographicsVQA ChartQA Visual7W VisDial CIRR NIGHTS WebQA MSCOCO \
--num_sample_per_subset 50000 \
--image_dir MMEB-train \
--max_len 256 --num_crops 4 --output_dir $OUTPUT_DIR --logging_steps 1 \
--lr_scheduler_type linear --learning_rate 2e-5 --max_steps 2000 \
--warmup_steps 200 --save_steps 1000 --normalize True \
--temperature 0.02 --per_device_train_batch_size 8 \
--grad_cache True --gc_q_chunk_size 2 --gc_p_chunk_size 2
Download the image file zip from huggingface
wget https://huggingface.co/datasets/TIGER-Lab/MMEB-eval/resolve/main/images.zip
unzip images.zip -d eval_images/
- For full-finetuned models, we use
python eval.py --model_name TIGER-Lab/VLM2Vec-Full \
--encode_output_path outputs/ \
--num_crops 4 --max_len 256 \
--pooling last --normalize True \
--dataset_name TIGER-Lab/MMEB-eval \
--subset_name N24News CIFAR-100 HatefulMemes VOC2007 SUN397 ImageNet-A ImageNet-R ObjectNet Country211 \
--dataset_split test --per_device_eval_batch_size 16 \
--image_dir eval_images/
- For LoRA-based models, we use
python eval.py --lora --model_name microsoft/Phi-3.5-vision-instruct --checkpoint_path TIGER-Lab/VLM2Vec-LoRA \
--encode_output_path outputs/ \
--num_crops 4 --max_len 256 \
--pooling last --normalize True \
--dataset_name TIGER-Lab/MMEB-eval \
--subset_name N24News CIFAR-100 HatefulMemes VOC2007 SUN397 ImageNet-A ImageNet-R ObjectNet Country211 \
--dataset_split test --per_device_eval_batch_size 16 \
--image_dir eval_images/
- We have adapted code from Tevatron, a flexible and efficient toolkit that supports training and inference for neural retrieval models.
@article{jiang2024vlm2vec,
title={VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks},
author={Jiang, Ziyan and Meng, Rui and Yang, Xinyi and Yavuz, Semih and Zhou, Yingbo and Chen, Wenhu},
journal={arXiv preprint arXiv:2410.05160},
year={2024}
}