VLM2Vec-Qwen2VL-7B

A new checkpoint trained using Qwen/Qwen2-VL-7B-Instruct with an enhanced training setup (LoRA tuning, batch size of 2048, maximum sub-dataset size of 100k). This model has shown significantly improved performance on MMEB & Flickr30K compared to the previous models using Phi-3.5 and llava-v1.6-mistral as backbone.

This repo contains the code and data for VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks. In this paper, we focus on building a unified multimodal embedding model suitable for a wide range of tasks. Our approach is based on transforming an existing, well-trained Vision-Language Model (VLM) into an embedding model.

Github

• Github

Data

Our model is being trained on MMEB-train and evaluated on MMEB-eval with contrastive learning. We only use in-batch negatives for training.

• Train data: https://huggingface.co/datasets/TIGER-Lab/MMEB-train
• Eval data: https://huggingface.co/datasets/TIGER-Lab/MMEB-eval

Performance

This model outperforms the baselines and previous version of VLM2Vec by a large margin.

How to use VLM2Vec

(More details please refer to our Github repo, here is just a simple demo.)

First you can clone our github

git clone https://github.com/TIGER-AI-Lab/VLM2Vec.git
pip install -r requirements.txt

from src.model import MMEBModel
from src.arguments import ModelArguments
from src.model_utils import load_processor, QWEN2_VL, vlm_image_tokens
from PIL import Image
import torch

model_args = ModelArguments(
    model_name='Qwen/Qwen2-VL-7B-Instruct',
    checkpoint_path='TIGER-Lab/VLM2Vec-Qwen2VL-7B',
    pooling='last',
    normalize=True,
    model_backbone='qwen2_vl',
    lora=True
)

processor = load_processor(model_args)
model = MMEBModel.load(model_args)
model = model.to('cuda', dtype=torch.bfloat16)
model.eval()

# Image + Text -> Text
inputs = processor(text=f'{vlm_image_tokens[QWEN2_VL]} Represent the given image with the following question: What is in the image',
                   images=Image.open('figures/example.jpg'),
                   return_tensors="pt")
inputs = {key: value.to('cuda') for key, value in inputs.items()}
inputs['pixel_values'] = inputs['pixel_values'].unsqueeze(0)
inputs['image_grid_thw'] = inputs['image_grid_thw'].unsqueeze(0)
qry_output = model(qry=inputs)["qry_reps"]

string = 'A cat and a dog'
inputs = processor(text=string,
                   images=None,
                   return_tensors="pt")
inputs = {key: value.to('cuda') for key, value in inputs.items()}
tgt_output = model(tgt=inputs)["tgt_reps"]
print(string, '=', model.compute_similarity(qry_output, tgt_output))
## A cat and a dog = tensor([[0.3301]], device='cuda:0', dtype=torch.bfloat16)

string = 'A cat and a tiger'
inputs = processor(text=string,
                   images=None,
                   return_tensors="pt")
inputs = {key: value.to('cuda') for key, value in inputs.items()}
tgt_output = model(tgt=inputs)["tgt_reps"]
print(string, '=', model.compute_similarity(qry_output, tgt_output))
## A cat and a tiger = tensor([[0.2891]], device='cuda:0', dtype=torch.bfloat16)

Citation

@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}
}