DCAF

Acknowledgement

This repository is based on the original implementation of ConFEDE. We sincerely thank the authors of ConFEDE for open-sourcing their code.

• Original ConFEDE Framework: Contrastive Feature Decomposition for Multimodal Sentiment

---

Dataset

The dataset used in this project is available for download at:
GitHub Repository – MMSA

---

Requirements

Python: 3.10.16

Python packages:

• matplotlib==3.6.3
• pytorch-metric-learning==0.9.99
• pytorch-pretrained-bert==0.6.2
• torch==2.6.0
• torchaudio==2.6.0
• torchsummary==1.5.1
• torchvision==0.21.0
• transformers==4.47.1
• ujson==4.0.2

Tip: you can also create a requirements.txt with the list above and run:

pip install -r requirements.txt

---

Checkpoints

Original DCAF Model Checkpoints:
Google Drive Link

---

Training

To train the model, you need to first train the encoders and then train the fusion network. Run the following commands sequentially:

# 1) Train the Encoders
python main.py

# 2) Train the Fusion Network
python main_fusion.py