https://github.com/manuelz/dlpt-food-classification

Project #2 for the OpenCV University course "Deep Learning with PyTorch".
https://github.com/manuelz/dlpt-food-classification

classification deep-learning python pytorch

Last synced: about 5 hours ago
JSON representation

Project #2 for the OpenCV University course "Deep Learning with PyTorch".

• Host: GitHub
• URL: https://github.com/manuelz/dlpt-food-classification
• Owner: ManuelZ
• Created: 2024-08-16T21:18:43.000Z (about 1 month ago)
• Default Branch: main
• Last Pushed: 2024-08-21T03:46:36.000Z (about 1 month ago)
• Last Synced: 2024-09-26T20:04:12.898Z (about 5 hours ago)
• Topics: classification, deep-learning, python, pytorch
• Language: Jupyter Notebook
• Homepage:
• Size: 1.69 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Classification: 13 Kenyan food types

This is the second project of the Opencv University course ["Deep Learning with PyTorch"](https://opencv.org/university/deep-learning-with-pytorch/).

It focuses on classifying images of 13 different Kenyan food classes.

## Introduction

Classification in computer vision is a task where the objective is to determine the class that the image belongs to. 

This project focuses on classifying images from 13 classes.

## Data

The dataset, containing 8174 images of various sizes, was split into 6536 training images and 1638 validation images.

These images are categorized into the following classes: 

    Bhaji, Chapati, Githeri, Kachumbari, Kukuchoma, Mandazi, Masalachips, Matoke, Mukimo, Nyamachoma, Pilau, Sukumawiki, Ugali

## The method used

Fine-tuning of a ResNet-50 backbone with a linear classifier using PyTorch Lightning to gain experience with high-level 

deep learning training.

- Various augmentations techniques were used to try to improve generalization:

  - Color jitter

  - Conversion to gray

  - Horizontal and vertical flips

  - Random shifting and rotation

  - Elastic transformations

  - Grid distortion

- The loss function used was Cross-Entropy.

- An SGD optimizer with weight decay.

- A learning rate scheduler that implements the 1-cycle policy. It adjusts the learning rate from an initial rate to a 

maximum, then decreases it to a much lower minimum.

## Discussion

Training this model for ~200 epochs resulted in an accuracy of 75.2% on the test set.

See the [notebook](project-2-deep-learning-with-pytorch-2024.ipynb).