OmniBind: Large-scale Omni Multimodal Representation via Binding Spaces

Zehan Wang · Ziang Zhang · Hang Zhang · Luping Liu · Rongjie Huang · Xize Cheng · Hengshuang Zhao · Zhou Zhao

OmniBind provide large-scale 3D point-audio-image-language omni representation models, ranging in scale from 7B to 30B parameters, achieve SoTA on 13 benchmarks.

News

• 2024/07/17: OmniBind is released.
• 2024/05/02: The predecessor FreeBind is accepted by ICML2024, and it is released at Here.

File structure

-assets
      [demo samples, including images, audios and point clouds]
-bpe
      [bpe_simple_vocab_16e6.txt.gz for imagebind]
-checkpoints
      [pretrained weights for OmniBind and experts]
-omni_model
      -eva_clip                     [part of eva_clip code]
      -laion_clap                   [part of laion_clap code]
      -Uni3D                        [part of Uni3D code]
      -wavcaps                      [part of wavcaps code]
      -transformer_430              [part of transformer v4.30.2 code]
      -tokenizers_013               [part of tokenizers v0.13 code]
      projector.py                  [the projector of OmniBind]
      router.py                     [the projector of OmniBind]
      experts.py                    [base feature extractors]
      omni_spaces.py                [combine router and experts together]
      omni_utils.py                 [useful functions]
      paths.py                      [paths of experts repo and weights]
      type.py                       [modality types]

Usage

1. preparing enviornments

Clone this repository and navigate to OmniBind folder.

git clone https://github.com/zehanwang01/OmniBind
cd OmniBind

Install pytorch and other 3rd party dependencies.

See preparation.sh for more details, or just execute the file.

chmod +x preparation.sh
bash ./preparation.sh

Lips: SigLip and laion-clap has environment conflict on transformers. We choose the 4.37.2 transformer version for siglip and built parts of 4.30.2 into the repository for CLAP. To install ImageBind in this environment, make sure libgeos++-dev is installed in your environment(sudo apt install libgeos++-dev), otherwise the you may fail to install package cartorpy for imagebind.

2. Inference

Extract and compare embeddings in OmniBind:

Note: The weights of some expert models and routers will be downloaded when the OmniBind is loaded for the first time.

from omni_model.omni_space import *
from safetensors.torch import load_model
a = OmniBind_Large(pretrained=True)
load_model(a, 'checkpoints/large.safetensors')
a = a.cuda()
with torch.no_grad():
    aud = a.emb_audios(['assets/train.wav', 'assets/toilet.wav'])
    img = a.emb_images(['assets/train.jpeg', 'assets/toilet.jpeg'])
    txt = a.emb_texts(['a photo of train', 'a photo of toilet'])
    pc = a.emb_texts(['assets/train.npy', 'assets/toilet.npy'])
print(aud.shape, img.shape, txt.shape, pc.shape)
print(aud@img.T)
print(aud@txt.T)
print(aud@pc.T)
print(img@txt.T)
print(img@pc.T)
print(txt@pc.T)

3. Pretrained weights

We have made minor changes to the code of CLAP, Wavcaps and Uni3D to make them better initialized in OmniBind, and the relevant code is included in the omni_model directory.

The encoder and projectors of OmniBind have been included in the checkpoint we prepared(Huggingface OmniBind).

The final structure of checkpoints should be like this:

-checkpoints
    -clap                                               [pretrained weights for CLAP]
        630k_clap_fullset_fusion.pt
        music_speech_audioset_epoch_15_esc_89.98.pt
    -projs                                              [space projectors and routers]
        base.pt
        large.pt
        full.pt
    -uni3d-g                                            [pretrained weights for Uni3D]
        -lvis/model.pt
        -mnet40/model.pt
        -scanobjnn/model.pt
    -wavcaps                                            [pretrained weights for wavcaps]
        HTSAT-BERT-PT.pt
        HTSAT-BERT-FT-AudioCaps.pt
        HTSAT-BERT-FT-Clotho.pt
    EVA02_CLIP_E_psz14_plus_s9B.pt                      [pretrained weights for EVA_CLIP_E14p]
    

Citation

If you find this project useful, please consider giving a star and citation:

@misc{wang2024omnibindlargescaleomnimultimodal,
      title={OmniBind: Large-scale Omni Multimodal Representation via Binding Spaces}, 
      author={Zehan Wang and Ziang Zhang and Hang Zhang and Luping Liu and Rongjie Huang and Xize Cheng and Hengshuang Zhao and Zhou Zhao},
      year={2024},
      eprint={2407.11895},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.11895}, 
}

@misc{wang2024freebind,
      title={FreeBind: Free Lunch in Unified Multimodal Space via Knowledge Fusion}, 
      author={Zehan Wang and Ziang Zhang and Xize Cheng and Rongjie Huang and Luping Liu and Zhenhui Ye and Haifeng Huang and Yang Zhao and Tao Jin and Peng Gao and Zhou Zhao},
      year={2024},
      eprint={2405.04883},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

If you have any questions or suggestions, feel free to drop us an email ( wangzehan01@zju.edu.cn, ziangzhang@zju.edu.cn ) or open an issue.