Vision Model Profiler using Waste Classification

Comparing different image classification models for sorting common household waste (cardboard, glass, metal, paper, plastic, trash).

Features

• Multiple CNN architectures (ResNet, ConvNeXt, EfficientNet, MobileNet)
• Training on multiple datasets (TrashNet, Garbage Classification v2)
• Training scripts, experiments and live demo

Installation

Prerequisites

• Python 3.8+
• A GPU or Apple Silicon (MPS) is recommended but not required

Setup

git clone <your-repo-url>
cd project-final

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

pip install -r requirements.txt

Download Data

TrashNet (6 classes):

wget https://github.com/garythung/trashnet/raw/master/data/dataset-resized.zip
unzip dataset-resized.zip -d data/

Garbage Classification v2 (10 classes): download from Kaggle and place the folders under data/garbage-classification-v2/ following the existing class subfolder structure.

How to Run

• Verify data and visualizations

python -m scripts.verify_data

• Train on TrashNet (6 classes)

python -m scripts.train_universal --dataset trashnet --architecture resnet50

• Train on Garbage Classification v2 (10 classes)

python -m scripts.train_universal --dataset garbage_v2 --architecture resnet50

• Live demo with webcam (requires a trained checkpoint)

python -m scripts.live_demo \
  --checkpoint outputs/.../best_model.pth \
  --dataset trashnet \
  --architecture resnet50

For more experiments (architecture comparisons, CLIP, SAM), see the scripts in the scripts/ folder.