- Dataset: CMU-MOSEI
- Source: textual feature, visual feature, acoustic feature
- Target: sentiment label (-3, 3), emotion label (6-class)
- Emotion 6-class: {happiness, sadness, anger, fear, disgust, surprise}
- Clone the repo.
git clone git@github.com:SoyeonHH/MMDA.git
- Set your CUDA device in here
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
word_emb_path = '/data1/multimodal/glove.840B.300d.txt'
- Modify CMU-Multimodal SDK file path in here
sdk_dir = Path('/data1/multimodal/CMU-MultimodalSDK')
- Modify MOSI and MOSEI file path in here with downloaded dataset from Google Drive.
data_dir = Path('/data1/multimodal')
data_dict = {'mosi': data_dir.joinpath('MOSI'), 'mosei': data_dir.joinpath('MOSEI')}
bash train.sh
If you want to use additional network, called 'ConfidNet', in previous architecture, run this code:
bash train_confid.sh
The Dataset source is from CMU-Multimodal SDK, kniter1/TAILOR, and declare-lab/Multimodal-Infomax.