GreenAI Waste Classification Project

This project aims to train a lightweight image classification model for wasted sorting while evaluating the environmental impact of the training process. We use a ResNet-50 pretrained on ImageNet and we fine-tuned the final layers on a waste-classification dataset from Kaggle. We also include a Green AI analysis with real GPU power measurements to estimate energy use and CO₂ emissions.

Project Structure

green-ai-project/
├── notebooks/
│   ├── download_data.ipynb     # Kaggle API download (need a kaggle API key)
│   ├── train_resnet50.ipynb    # training notebook + stats logging
│   ├── test.ipynb              # evaluation + visualisations
│   └── greenai_eval.ipynb      # energy, CO₂ and FLOPs analysis
├── data/                       # dataset (ignored in git)
├── checkpoints/                # weights + training_stats.json + gpu_per_sec.txt               
├── .gitignore
└── README.md

Dataset

We use the Kaggle dataset:

Trash Type Image Dataset https://www.kaggle.com/datasets/farzadnekouei/trash-type-image-dataset

Download it using the provided notebook:

notebooks/download_data.ipynb