A Unimodal Valence-Arousal Driven Contrastive Learning Framework for Multimodal Multi-Label Emotion Recognition

Wenjie Zheng, Jianfei Yu, and Rui Xia

📄 Paper

This repository contains the code for UniVA, a framework that proposes unimodal valence-arousal driven contrastive learning for the multimodal multi-label emotion recognition task.

Dependencies

conda env create -f environment.yml

---

Data preparation

Download link. Also, there are two files that, due to upload size limitations, have been placed at the link.

---

Evaluating UniVA on the MOSEI dataset

you can check our UniVA-RoBERTa on 4 NVIDIA 3090 GPUs by running the script below

nohup bash run_MOSEI/run_MOSEI_TAV_ours.sh &

you can get the following results: Acc51.3 HL0.182 miF160.5 maF144.4

To evaluate the performance of UniVA-Glove on 1 NVIDIA 3090Ti GPU, run the script below

nohup bash run_MOSEI/run_MOSEI_TAV_ours_glove.sh &

you can get the following results: Acc49.2 HL0.205 miF157.2 maF137.2

Evaluating UniVA on the M3ED dataset

you can check our UniVA-RoBERTa on 4 NVIDIA 3090 GPUs by running the script below

nohup bash run_M3ED/run_M3ED_TAV_ours.sh &

you can get the following results: Acc50.6 HL0.149 miF153.4 maF140.2

To evaluate the performance of UniVA-Glove on 1 NVIDIA 3090Ti GPU, run the script below

nohup bash run_M3ED/run_M3ED_TAV_ours_glove.sh &

you can get the following results: Acc46.4 HL0.159 miF149.1 maF124.2

---

Citation

Please consider citing the following if this repo is helpful to your research.

@inproceedings{zheng2024univa,
  title={A Unimodal Valence-Arousal Driven Contrastive Learning Framework for Multimodal Multi-Label Emotion Recognition},
  author={Zheng, Wenjie and Yu, Jianfei and Xia, Rui},
  booktitle={Proceedings of the 32st ACM International Conference on Multimedia},
  year={2024}
}

Please let me know if I can future improve in this repositories or there is anything wrong in our work. You can ask questions via issues in Github or contact me via email wjzheng@njust.edu.cn. Thanks for your support!