CIFAR-10 Image Classification using ResNet-50

This repository contains a deep learning project that implements a ResNet-50-based model for classifying images in the CIFAR-10 dataset. The CIFAR-10 dataset consists of 60,000 images distributed across 10 classes. This project achieves 94.66% accuracy on the test data using transfer learning and custom training.

---

Overview

The project includes:

• Dataset Preparation: Reduced dataset for training and separate unused data for testing.
• Model Architecture: ResNet-50 pre-trained on ImageNet with additional layers for CIFAR-10 classification.
• Training Strategy: Early stopping and model checkpointing for optimized training.
• Submission: Predictions saved for Kaggle submission in the required format.

---

Features

• Dataset Preprocessing:

  • Reduced CIFAR-10 dataset to 40,000 images for training, stratified across classes.
  • A separate folder, for_test_dir, contains unused images for model testing.

• Model Details:

  • ResNet-50 backbone pre-trained on ImageNet.
  • Upsampling and fully connected layers for CIFAR-10 image classification.
  • Regularization techniques like dropout to prevent overfitting.

• Evaluation Results:

  • Loss: 0.2436
  • Accuracy: 94.66%

• Kaggle Submission:

  • Creates a submission.csv file with image IDs and corresponding predictions.

---

Setup

Prerequisites

• Python >= 3.8
• TensorFlow >= 2.5
• Pandas, NumPy, Matplotlib, and scikit-learn