The paper for this repo has been published: Evaluating Variants of wav2vec 2.0 on Affective Vocal Burst Tasks.
Another version of https://github.com/bagustris/A-VB2022 using CCC loss function (please take a look that repository before trying code in this repository).
Feature extraction for 'w2v2-r-er' and 'w2v2-r-vad' are given there. Also install requirements from there (inside feature_based directory).
We provide a 'feature-based' approach for all four tasks to reproduce results in our ICASSP paper (see citation).
First, make sure you download the data and features, placing them within your working directory. For more info and instructions on how to access the competition data visit competitions.hume.ai.
We suggest creating a virtual environment, and installing the requirements.txt
conda create -n avb-2022-ccc python=3.8
conda activate avb-2022-ccc
pip install -r requirements.txt
After the data, labels, and features are downloaded to your working directory, when running main.py set to -d ./ i.e., where features/ and labels/ are, and run:
A-VB High Baseline
python main.py -d /data/A-VB/ -f w2v2-R-emo-vad -t type -e 100 -lr 0.0005 -bs 8 -p 10 --n_seeds 20
Follow the same procedure for each task altering -t.
| Option | Description |
|---|---|
-d |
Set the path to working dir |
-f |
Feature set e.g., eGeMAPS |
-t |
Task ['high','two','culture','type'] |
-e |
Number of Epochs (default: 20) |
-lr |
Learning Rate (default: 0.001) |
-bs |
Batch Size (default: 8) |
-p |
Early Stopping Patience (default: 5) |
--n_seeds |
Number of Seeds to run for |
--verbose |
Maximum verbosity, (default: quiet) |
B. T. Atmaja and A. Sasou, “Evaluating Variants of wav2vec 2.0 on Affective Vocal Burst Tasks,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Jun. 2023, pp. 1–5, doi: 10.1109/ICASSP49357.2023.10096552.