Dimensionality reduction of CNNs for image processing
deep learning • model pruning • CNNs • Streamlit • PyTorch
This project was developed as part of the Automation Engineering Project at TU Berlin, within the department of Industrial Automation Technology. The goal was to reduce the dimensions of convolutional neural networks through model pruning, improving efficiency and speed while preserving accuracy as much as possible.
More broadly, the project tackles the challenge that many deep learning models are excessively large and resource-intensive, raising the critical question of optimizing them without compromising performance.
Simplified representation of a pruning pipeline
The outcome was a Streamlit-based tool with an implemented pipeline, allowing users to test, apply and compare CNN model pruning techniques, and save pruned models.
My contributions included implementing the app, integrating and merging code from various pruning methods developed by different team members, deploying the application with Docker, and coordinating the group. Additionally, I cocreated the results page, which visually presents and analyzes the pruning outcomes within the app.