Frédéric Debraine, 30/05/2020

[Slides]

• Train/test split based on speaker id to avoid data leakage or overfitting on voices.
• Replace stratified CV by Group CV to avoid data leakage between train and validation sets.
• Replace ROC-AUC metric by F1 (otherwise we need probability outputs for all models, not possible for ensemble)
• Making sure the score average mode is on 'weighted' to take into account class imbalance.

• modules: Python scripts related to ML pipeline.
• data: Contains Emo-DB dataset.
• app.py: Flask API defining end points to train and predict.
• Dockerfile: Defines steps to build container (install dependencies, download and extract data, expose ports,...)
• docker-compose.yml: Automatize the build and run of two containers: one for the jupyter notebook and the other for the Flask API.
• Report_SER.ipynb: Notebook used for fast prototyping and presenting preliminary results and visualizations.

• docker-compose build
• docker-compose up

• Open in browser: http://localhost:8081/tree?
• Password: password

• curl http://0.0.0.0:8080/train
• You can also select the number of top base models to consider from 2 to 5 (default is n=3): curl http://0.0.0.0:8080/train?n=3

• curl http://0.0.0.0:8080/test_ids | json_pp

• curl http://0.0.0.0:8080/predict?id=10b09Lb

• https://github.com/ritikraj660/Speech-emotion-recognition/
• https://librosa.github.io/librosa/index.html
• https://www.datasciencelearner.com/
• https://towardsdatascience.com/