/tilli-machine-learning

Machine Learning Code for Tilli

Primary LanguagePython

Machine Learning Model Server

This repository contains code for running a machine learning model server using Docker and Python. The server serves predictions based on a trained logistic regression model. It also includes tests, a client server for manual testing, and additional files needed at runtime.

Instructions

  1. To run the server with the machine learning model, follow these steps:

    • Build the Docker image using the provided Dockerfile docker build -t ml-model-server .
    • Run the Docker container to start the server on port 8080. docker run -p 8080:8080 ml-model-server

  2. The tests are located in the tests folder and are executed via the CI/CD pipeline set up in the test.yml file located in the workflows folder (using GitHub Actions). To add new tests, create a new file in the tests folder and add the relevant test cases. Currently, the file tests/test_server.py is created for testing the server.

  3. To train a new model, create a folder named data and add a JSON file (data_1.json) with the required data. (Note: The file format may be updated in the future.)

  4. The client/client.py file contains a client server that can be used for manual testing of the server. Modify the url, emotion, and feedback variables to match your server's endpoint and the desired input values.

  5. The models folder contains additional information and files needed at runtime. Currently, the following files are included:

  • logreg_model.joblib: A logistic regression model trained with the data.
  • col_names.pk1: Column names of the original dataset.
  1. The src/model_dev.py file demonstrates the model development process. It includes the following steps:
  • Load the data using the data_etl() function from the src/util.py file.
  • Convert categorical columns to one-hot encoded columns.
  • Separate the input features and target variable.
  • Create and train a logistic regression model.
  • Predict the activity for a given emotion and feedback.
  • Save the trained model and the column names for future use.
  1. Utility functions are stored in the src/util.py file. These functions include:
  • categorical_to_numeric_df(): Converts categorical values to a numeric range.
  • visualize(): Visualizes the spread of user interactions.
  • data_etl(): Loads and preprocesses the data.
  1. The src/model_evaluation.py file is under development and contains code to evaluate the performance of different models.

Feel free to explore the repository and adapt it to your needs.