Awesome

Cross-modal Prompts: Adapting Large Pre-trained Models for Audio-Visual Downstream Tasks, NeurIPS 2023

This is the Pytorch implementation of our paper:

Cross-modal Prompts: Adapting Large Pre-trained Models for Audio-Visual Downstream Tasks

[Paper] [arXiv] [Video] [Poster] [Slides]

Haoyi Duan, Yan Xia, Mingze Zhou, Li Tang, Jieming Zhu, Zhou Zhao

In NeurIPS 2023

📝Requirements and Installation

• Getting Started

git clone https://github.com/haoyi-duan/DG-SCT
cd DG-SCT
pip install -r requirements.txt

• Download HTS-AT Backbone

  Download checkpoints.zip from Google Drive or Baidu Disk (pwd: 2023), and extract it into the directory ./DG-SCT/.

AVE

• Download Data

  Download frames.zip from Google Drive or Baidu Disk (pwd: 2023), wave.zip from Google Drive or Baidu Disk (pwd: 2023), and extract them into the directory ./data/AVE/.

• Usage

  Go to AVE task directory.

  cd DG-SCT/AVE
  

  • Training

    bash train.sh
    

  • Testing

    ./models/best_82.18.pt: Google Drive or Baidu Disk (pwd: 2023)

    bash test.sh
    

• Results

  

AVS

• Download Data

  • Download Dataset

    The updated AVSBench dataset is available here (AVSBench-object). You may request the dataset by filling the Google Form.

    The downloaded data should be placed to the directory ./data/.

  • Download Wave

    Download wave for task S4 (Google Drive or Baidu Disk (pwd: 2023)) and task MS3 (Google Drive or Baidu Disk (pwd: 2023)), and extract them into the directory ./data/AVSBench_data/Single-source/s4_data/ and ./data/AVSBench_data/Multi-sources/ms3_data/, respectively.

• Download pretrained backbones

  The pretrained ResNet50/PVT-v2-b5 (vision) and VGGish (audio) backbones can be downloaded from here and placed to the directory ./DG-SCT/AVS/pretrained_backbones/.

• Usage

  Go to AVS task directory.

  # for S4 task:
  cd DG-SCT/AVS/avs_scripts/avs_s4
  
  # for MS3 task:
  cd DG-SCT/AVS/avs_scripts/avs_ms3
  

  • Training

    bash train.sh
    

  • Testing

    checkpoint for S4 task: ./DG-SCT/AVS/avs_scripts/avs_s4/train_logs Google Drive or Baidu Disk (pwd:2023)

    checkpoint for MS3 task: ./DG-SCT/AVS/avs_scripts/avs_ms3/train_logs Google Drive or Baidu Disk (pwd:2023)

    bash test.sh
    

• Results

  

AVVP

• Download Data

  Download extracted feats, frame and wave of LLP dataset from Baidu Disk (pwd: 2023), and extract it into the directory ./data/AVVP/.

• Usage

  Go to AVVP task directory:

  cd DG-SCT/AVVP
  

  • Training

    bash train.sh
    

  • Testing

    ./models/MGN_Net.pt: Google Drive or Baidu Disk (pwd:2023)

    bash test.sh
    

• Results

  

AVQA

• Download Data

  Download frames.zip from Google Drive or Baidu Disk (pwd: 2023), audio_wave.zip from Google Drive or Baidu Disk (pwd: 2023), and extract them into the directory ./data/AVQA/.

• Usage

  Go to AVQA task directory.

  cd DG-SCT/AVQA
  

  • Audio-Visual Grounding Generation

    python grounding_gen/main_grd_gen.py
    

    You can download the ./grounding_gen/models_grounding_gen/lavish_grounding_gen_best.pt from Google Drive or Baidu Disk (pwd: 2023) to skip the Audio-Visual Grounding Generation process.

  • Training

    bash train.sh
    

  • Testing

    ./net_grd_avst/avst_models/avst.pt: Google Drive or Baidu Disk (pwd: 2023)

    bash test.sh
    

• Results

  

Few-shot/Zero-shot

We use audio-text backbones in CLAP: 630k-audioset-fusion-best.pt, and 630k-fusion-best.pt. Please download and place them into the directory ./pretrain/models/.

• Few-shot

  • Go to Few-shot Directory

    cd few-shot
    

  • AVE

    • 1 shot

      python main_AVE.py --dataset_name AVE --shot 1 --alpha 0.2 --beta 0.05 --gamma 0.01 --weak 0 --classification 0
      

    • 2 shots

      python main_AVE.py --dataset_name AVE --shot 2 --alpha 0.2 --beta 0.05 --gamma 0.01 --weak 0 --classification 0
      

    • 4 shots

      python main_AVE.py --dataset_name AVE --shot 4 --alpha 0.2 --beta 0.05 --gamma 0.01 --weak 0 --classification 0
      

    • 8 shots

      python main_AVE.py --dataset_name AVE --shot 8 --alpha 0.2 --beta 0.05 --gamma 0.01 --weak 0 --classification 0
      

    • 16 shots

      python main_AVE.py --dataset_name AVE --shot 16 --alpha 0.2 --beta 0.05 --gamma 0.01 --weak 0 --classification 0
      

  • AVE Classification

    • 1 shot

      python main_AVE_class.py --dataset_name AVE --shot 1 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 2 shots

      python main_AVE_class.py --dataset_name AVE --shot 2 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 4 shots

      python main_AVE_class.py --dataset_name AVE --shot 4 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 8 shots

      python main_AVE_class.py --dataset_name AVE --shot 8 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 16 shots

      python main_AVE_class.py --dataset_name AVE --shot 16 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

  • LLP Classification

    • 1 shot

      python main_LLP_class.py --dataset_name LLP --shot 1 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 2 shots

      python main_LLP_class.py --dataset_name LLP --shot 2 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 4 shots

      python main_LLP_class.py --dataset_name LLP --shot 4 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 8 shots

      python main_LLP_class.py --dataset_name LLP --shot 8 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

    • 16 shots

      python main_LLP_class.py --dataset_name LLP --shot 16 --alpha 0.2 --beta 0.05 --gamma 0.05 --weak 1 --classification 1
      

• Zero-shot

  • Download Data

    Download VGG-Sound(40K) from Baidu Disk (pwd: 2023), and extract it into the directory ./data/.

  • Usage

    • Pretrain on VGG-Sound(40K)

      cd pretrain
      bash train.sh
      

      The pretrained model will be placed at pretrain/models/.

    • Zero-shot

      MODEL_NAME="name of the pretrained model in pretrain/models/."
      # AVE
      python zero_shot.py --test_dataset_name AVE --backbone $MODEL_NAME --is_event_score 1
      
      # AVE classification
      python zero_shot.py --test_dataset_name AVE --backbone $MODEL_NAME --is_event_score 0
      
      # LLP classification
      python zero_shot.py --test_dataset_name LLP --backbone $MODEL_NAME --is_event_score 0
      

• Results

🎓Cite

If you find this work useful, please consider citing it.

@inproceedings{duan2023cross,
  title={Cross-modal Prompts: Adapting Large Pre-trained Models for Audio-Visual Downstream Tasks},
  author={Duan, Haoyi and Xia, Yan and Zhou, Mingze and Tang, Li and Zhu, Jieming and Zhao, Zhou},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}

👍Acknowledgments

Our code is based on CMBS, AVSBench, MGN, MUSIC-AVQA, and LAVisH.

✏Model Checkpoints