https://github.com/elifirinci/cifar-10

This project explores the concept of transfer learning using the CIFAR-10 dataset. The work demonstrates how to reuse a convolutional neural network trained on a subset of image classes and then fine-tune it on a different set of classes. This approach is common in real-world deep learning applications where labeled data is limited.
https://github.com/elifirinci/cifar-10

cifar10 classification deep-learning fine-tuning googlenet inceptionv3

Last synced: 2 months ago
JSON representation

This project explores the concept of transfer learning using the CIFAR-10 dataset. The work demonstrates how to reuse a convolutional neural network trained on a subset of image classes and then fine-tune it on a different set of classes. This approach is common in real-world deep learning applications where labeled data is limited.

• Host: GitHub
• URL: https://github.com/elifirinci/cifar-10
• Owner: elifirinci
• License: mit
• Created: 2025-05-23T16:49:11.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2025-05-28T11:42:54.000Z (5 months ago)
• Last Synced: 2025-06-06T08:43:45.433Z (4 months ago)
• Topics: cifar10, classification, deep-learning, fine-tuning, googlenet, inceptionv3
• Language: Jupyter Notebook
• Homepage:
• Size: 512 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
 

## 📚 Dataset

The CIFAR-10 dataset contains 60,000 color images (32x32 pixels) divided into 10 classes:

- airplane

- automobile

- bird

- cat

- deer

- dog

- frog

- horse

- ship

- truck

In this project:

- The model is first trained on 5 classes: airplane, automobile, bird, cat, and deer.

- Then, the last layer(s) of the network are retrained to classify the other 5 classes: dog, frog, horse, ship, and truck.

## 🧠 Method

- A pre-trained InceptionV3 model is used for transfer learning.

- Two strategies were tested:

  1. Freeze all layers except the output layer.

  2. Freeze all layers except the fully connected and output layers.

### Results

- **Trainable Parameters**:

  - Strategy 1: ~262,917

  - Strategy 2: ~263,109

- **Accuracy**:

  - Strategy 1: ~93% (train and validation)

  - Strategy 2: ~94% (train), ~94.5% (validation)

- Strategy 2 performed better due to more layers being fine-tuned, allowing better adaptation to the new classes.

## Why Fine-Tuning is Faster

Fine-tuning is faster than full training because most layers are frozen and do not require gradient calculations. This reduces the computational load significantly.

## Libraries Used

- TensorFlow / Keras

- NumPy

- Matplotlib