Implementation of portable neural networks using docker containerization

Student: Sara Mikoč

This thesis explores the implementation of portable neural networks using Docker technology to eliminate the issue of incompatibility between different computing environments. For the purposes of research and system demonstration, a Multi-Layer Perceptron (MLP) model was trained within the PyTorch framework using the publicly available Pima Indians Diabetes Database. Data preparation included feature standardization, the application of the KNN method for missing value imputation, and the introduction of class weights to address class imbalance. The model evaluation was conducted by analyzing the confusion matrix and calculating metrics such as accuracy, precision, recall, and F1-score, which confirmed the reliability of the predictions. The main part of the implementation relates to the containerization of the developed model. By creating a Dockerfile configuration, the model, along with all necessary libraries and dependencies, was packaged into an isolated Docker image, ensuring that the system operates identically on any platform without the need for manual environment setup. Additionally, a FastAPI interface was implemented to enable interaction with the model and obtain real-time predictions from within the container itself. This approach demonstrates that the use of Docker ensures complete consistency and reliability of neural networks during distribution, ensuring that the model retains its functionality and accuracy regardless of the host's hardware or software infrastructure.

Repository: https://repozitorij.fsb.unizg.hr/object/fsb:13796