Multitask Emotion Recognition Model with Knowledge Distillation and Task Discriminator

Source code of Multitask Emotion Recognition Model with Knowledge Distillation and Task Discriminator

Requirements

1. requirements

   • Anaconda must be installed and GPU-enabled.
   • conda_requirements.yml should be fixed to fit your anaconda path.

   name: imlab_ABAW
   channels:
     - conda-forge
       .
       .
       .
   # specify your anaconda path with env name
   # ex)/home/{user_name}/anaconda3/envs/imlab_ABAW
   prefix: /home/{your_user_name}/anaconda3/envs/imlab_ABAW

   • Then, you can create new conda environment.

   $ conda env create -f conda_requirements.yml
   $ conda activate imlab_ABAW
   $ pip install -r requirements.txt

Download data

1. Download data in ./data directory(data/)

   • Download extracted feature from SoundNet, FER model. You can download here (image_feature, audio_feature)
   • You should download idx pickle including file_name, labels etc. You can download here
   • You should also download Multi_Task_Learning_Challenge_test_set_release.txt. download here

   data_path
   	-features
   	-idx
   	-testset
   	-result(will be generated automatically)

2. Change ‘data_path’ in config.py to your data_path data downloaded.

Train

1. When you run trainer.py, it is stored in a directory called result in the data_path as the date and time of the file execution.
   • If you want a pre-trained file, you can download here

Evaluate

1. Change "eval_path" on config.py to the absolute path of generated folder.
2. Running submit.py creates a submission folder in that path and creates a submission file.

Citation

@misc{https://doi.org/10.48550/arxiv.2203.13072, doi = {10.48550/ARXIV.2203.13072}, url = {https://arxiv.org/abs/2203.13072}, author = {Jeong, Euiseok and Oh, Geesung and Lim, Sejoon}, title = {Multitask Emotion Recognition Model with Knowledge Distillation and Task Discriminator}, publisher = {arXiv}, year = {2022}, copyright = {Creative Commons Zero v1.0 Universal} }