Official PyTorch code for training and inference pipeline for
DepMamba: Progressive Fusion Mamba for Multimodal Depression Detection

📰 News

• 25/09/2024 Released the arxiv version paper.
• 25/09/2024 Released the training code.

📕Introduction

In this work, we propose an audio-visual progressive fusion Mamba model for efficient multimodal depression detection, termed DepMamba. Specifically, the DepMamba features two core designs: hierarchical contextual modeling and progressive multimodal fusion. First, we introduce CNN and Mamba blocks to extract features from local to global scales, enriching contextual representation within long-range sequences. Second, we propose a multimodal collaborative State Space Model (SSM) that extracts intermodal and intramodal information for each modality by sharing state transition matrices. A multimodal enhanced SSM is then employed to process concatenated audio-visual features for improved modality cohesion.

📖 Usage:

1. Clone Repository

git clone https://github.com/Jiaxin-Ye/DepMamba.git

2. Requirements

Our code is based on Python 3.8 and CUDA 11.7. There are a few dependencies for running the code. The major libraries including Mamba and PyTorch are listed as follows:

conda create -n DepMamba -c conda-forge python=3.8
conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia
conda install packaging
git clone https://github.com/Dao-AILab/causal-conv1d.git 
cd causal-conv1d 
git checkout v1.1.3 
CAUSAL_CONV1D_FORCE_BUILD=TRUE pip install .
cd ..
git clone https://github.com/state-spaces/mamba.git
cd ./mamba
git checkout v1.1.3
MAMBA_FORCE_BUILD=TRUE pip install .
pip install -r requirement.txt

If you want to record training log, you need to login your own wandb account.

wandb login

Change these lines in main.py to your own account.

wandb.init(
    project="DepMamba", entity="<your-wandb-id>", config=args, name=wandb_run_name,
)

3. Prepare Datasets

We use the D-Vlog and LMVD dataset, proposed in this paper. For the D-Vlog dataset, please fill in the form at the bottom of the dataset website, and send a request email to the author. For the LMVD dataset, please download features on the released Baidu Netdisk website or figshare.

Following D-Vlog's setup, the dataset is split into train, validation and test sets with a 7:1:2 ratio. For the LMVD without official splitting, we randomly split the LMVD with an 8:1:1 ratio and the specific division is stored in `./datasets/lmvd_labels.csv'.

Furthermore, you can run extract_lmvd_npy.py to obtain .npy features to train the model.

5. Training and Testing

Training

$ python main.py --train True --train_gender both --test_gender both --epochs 120 --batch_size 16 --learning_rate 1e-4 --model DepMamba --dataset dvlog --gpu 0

$ python main.py --train True --train_gender both --test_gender both --epochs 120 --batch_size 16 --learning_rate 1e-4 --model DepMamba --dataset lmvd --gpu 0

Testing

$ python main.py --train False --test_gender both --epochs 120 --batch_size 16 --learning_rate 1e-4 --model DepMamba --dataset dvlog --gpu 0

$ python main.py --train False --test_gender both --epochs 120 --batch_size 16 --learning_rate 1e-4 --model DepMamba --dataset lmvd --gpu 0

📖 Citation

• If you find this project useful for your research, please cite our paper:

@article{yedepmamba,
  title={DepMamba: Progressive Fusion Mamba for Multimodal Depression Detection},
  author = {Jiaxin Ye and Junping Zhang and Hongming Shan},
  journal      = {CoRR},
  volume       = {abs/2409.15936},
  year         = {2024},
  eprinttype    = {arXiv},
  eprint       = {2409.15936},
}

🙌🏻 Acknowledgement

• We acknowledge the wonderful work of Mamba and Vision Mamba.
• We borrow their implementation of Mamba and bidirectional Mamba.
• We acknowledge AllenYolk and ConMamba.
• The training pipelines are adapted from SpeechBrain.