https://github.com/reyhaneh-saffar/autoencoder-reconstruction-of-mixed-mnist-and-cifar-10-images

Exploring the use of an autoencoder neural network for data compression and reconstruction
https://github.com/reyhaneh-saffar/autoencoder-reconstruction-of-mixed-mnist-and-cifar-10-images

Last synced: 16 days ago
JSON representation

Exploring the use of an autoencoder neural network for data compression and reconstruction

• Host: GitHub
• URL: https://github.com/reyhaneh-saffar/autoencoder-reconstruction-of-mixed-mnist-and-cifar-10-images
• Owner: reyhaneh-saffar
• Created: 2025-01-11T00:14:12.000Z (20 days ago)
• Default Branch: main
• Last Pushed: 2025-01-11T00:18:19.000Z (20 days ago)
• Last Synced: 2025-01-11T01:21:46.573Z (20 days ago)
• Language: Jupyter Notebook
• Size: 88.9 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
This project explores the use of an autoencoder neural network for data compression and reconstruction. The datasets CIFAR-10 and MNIST were preprocessed, combined, and used to train the model, with promising results in reducing reconstruction loss.

### Data Processing

- **Dataset Preparation**:

  - CIFAR-10 and MNIST datasets were transformed and normalized to ensure compatibility.

  - MNIST images were converted to RGB and resized to match CIFAR-10 dimensions.

  - Both datasets were standardized in terms of image size and pixel value distribution.

- **Mean Image Computation**:

  - Corresponding images from CIFAR-10 and MNIST were averaged to create new "mean images."

  - This process enriched the dataset by combining features from both datasets, producing unique training and testing data.

- **Dataset Structuring**:

  - The mean images were reshaped to fit the input format required for model training and evaluation.

  - Structured training and testing datasets were created to ensure consistency and diversity for the model.

### Training the Autoencoder

- **Model Construction**:

  - The autoencoder was composed of:

    - **Encoder**: Compressed input data into a latent space representation using convolutional layers with non-linear activation functions.

    - **Decoder**: Reconstructed input data from the latent space using transposed convolutional layers, restoring spatial dimensions.

  - A loss function measured reconstruction accuracy, optimized using backpropagation.

- **Training Process**:

  - **Data Handling**: Training and testing datasets were loaded into data loaders for efficient batching and shuffling.

  - **Training Loop**:

    - Over 50 epochs, the autoencoder iteratively learned to compress and reconstruct input data.

    - The optimizer adjusted model parameters based on reconstruction loss.

  - **Evaluation Loop**:

    - After each epoch, the model was tested on unseen data to monitor its generalization ability.

    - Reconstruction loss was recorded for both training and testing datasets.

### Results and Insights

- **Performance Metrics**:

  - Training Loss: Decreased from **0.9747** in the first epoch to **0.2641** in the final epoch.

  - Testing Loss: Reduced from **0.9674** to **0.2609**.

- **Visualization**:

  - A plot of training and testing losses over 50 epochs showed steady decreases, indicating effective learning and generalization.