Model Optimization for Segmentation

The aim of this project was to a take a scene segmentation model (DeepLabV3) used in a self-driving car and optimize it using Quantization and Knowledge Distillation techniques.

Dependencies

1. thop : PyTorch OpCounter
2. pytorch

Project Structure

• outputs contain output videos
• scripts/models contains the model architecture used for inference
• scipts contains the helper files for reading the dataset, tabulating metrics for inference and model performance
• main contains the notebooks for quantization and knowledge distillation

Output