Audio Sentiment Classifier Capstone Project

Summary

This repo contains my Capstone Project developed during my
Machine Learning Engineer studies. This project addresses the business needs of a hypothetical call center which must quickly determine whether its calls require escalation / intervention due to customer dissatisfaction. For this project I selected an audio dataset and then performed data cleaning, wrangling, and exploratory data analysis. Next I performed feature selection/development, assessed several ML algorithms, and performed various hyperparameter searches. (Details found here) With a performant model in hand, I developed an ETL pipeline to reliably reproduce the training results from the initial data. Next wrote an API to accept audio data and respond with a binary sentiment classification. Finally, I containerized the API and deployed it to AWS.

Usage

Navigate to Prod Prediction App in a browser.

Click the 'Browse' button to upload a file using the file picker and then click submit. The file must be of .wav format and less than 16 mb is size. The response will contain a binary value specifying whether the audio's sentiment is negative (value 1 returned) or positive (value 0 returned).

You can try some sample files from the MELD dataset I've included here.

Dataset

The MELD dataset may be downloaded directly from: https://affective-meld.github.io/ See 'Download Raw Data'. In particular, this project uses the Raw video files and their associated labels. Here is a backup in case the original is down.

Development Installation

Python 3.7.* is required to support the librosa audio package. Due to dependency compability issues the production prediction API runs with its own (minimal) set of requirements distinct from those of the ETL pipeline.

Prodution Prediction API

The Production Prediction API can be installed from the top-level requirements.txt with conda:

conda create --name prediction_app python=3.7 --file requirements.txt

conda activate prediction_app

Install the sentiment_classifier Python package:

pip install -e .

Run the Flask API locally:

export FLASK_APP='src/sentiment_classifier/prediction/wsgi.py' && export FLASK_ENV=development && flask run

Training ETL Pipeline

The training ETL Pipeline uses a separate set of requirements:

conda create --name etl_pipe python=3.7 --file requirements/etl_requirements.txt

conda activate etl_pipe

Run the unit tests: pytest tests

Download the raw video files downloaded to the data/raw/dev|train|test directories. Run the ML training pipeline to prepare the data and train a model: python src/sentiment_classifier/run_dag.py

To run a Bayesian hyperparameter search instead of training a model: python src/sentiment_classifier/run_dag.py --search

in either case, the results of the run are stored to data/results/

Creating requirements.txt

For reference the requirements.txt file was created by: pip list --format=freeze > *.txt

This command is used due to a bug in pip freeze to obtain requirements with pinned version numbers which is installable using either pip or conda.

Dockerizing the Prediction App

Must have Docker engine set up and running.

Create docker image from the Dockerfile: docker image build -t prediction_app .

Run a docker container using the image docker run -d -p 5000:5000 prediction_app

Then navigate in a browser to: localhost:5000

AWS Deployment

Configure AWS CLI: https://docs.aws.amazon.com/cli/latest/userguide/cli-configure-quickstart.html

AWS lightsail instructions: https://aws.amazon.com/getting-started/hands-on/serve-a-flask-app/

docker build -t flask-container .

aws lightsail create-container-service --service-name flask-service --power small --scale 1

aws lightsail push-container-image --service-name flask-service --label flask-container --image flask-container

# note container number in output

aws lightsail create-container-service-deployment \
--service-name flask-service \
--containers file://deploy/containers.json \
--public-endpoint file://deploy/public-endpoint.json

# check status
aws lightsail get-container-services --service-name flask-service

# Cleanup and delete Lightsail resources
aws lightsail delete-container-service --service-name flask-service

Sphinx documentation

Using html theme: conda install -y sphinx_rtd_theme

Configure sphinx: sphinx-quickstart docs

Generate docs: sphinx-apidoc -f -o docs/source src/sentiment_classifier

cd docs && make html && cd ..

Generated docs are viewable from: docs/build/html/index.html

ryanjmccall / prod_mle_capstone