Janie1996 / MSRFG

The code for Multi-Scale Receptive Field Graph Model for Emotion Recognition in Conversations

Multi-Scale Receptive Field Graph Model for Emotion Recognition in Conversations

PyTorch implementation for the paper:

• Title: Multi-Scale Receptive Field Graph Model for Emotion Recognition in Conversations

• Authors: Jie Wei, Guanyu Hu, Luu Anh Tuan, Xinyu Yang, Wenjing Zhu

• Submitted to: ICASSP2023

Getting Started

git clone https://github.com/Janie1996/MSRFG.git

Requirements

You can create an anaconda environment with:

conda env create -f environment.yaml
conda activate MSRFG

Usage

1. Preparation

a. Download dataset from google-drive. Unzip it and put them under ./data/

b. Download model checkpoint from google-drive. Unzip it and put them under ./checkpoints/

2. Test

• Run IEMOCAP

  python eval_iemocap.py

• Run MELD

  python eval_meld.py

3. Train

1. the proposed model training

• Run IEMOCAP

  python train\train_iemocap.py

• Run MELD

  python train\train_meld.py

2. fine-tune the Utterance Encoder

• Wav2Vec 2.0
• RoBERTa-Large

If you have questions, feel free to contact [e-mail]

Acknowledgements

• IEMOCAP: Interactive emotional dyadic motion capture database
• MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations
• Directed Acyclic Graph Network for Conversational Emotion Recognition